A comprehensive review of ball detection techniques in sports

Introduction

Understanding the ball’s movement in sports analytics is integral to game analysis, strategy formulation, and audience engagement. The ball often serves as the focal point around which key events unfold, making its detection and tracking a critical task. However, the dynamic and unpredictable nature of sports introduces numerous challenges, including complex backgrounds, occlusions by players, small object sizes, and the high-speed movement of the ball. These challenges require advanced detection techniques capable of operating with high accuracy in real-time.

The evolution of object detection methodologies, from traditional computer vision (CV) techniques such as colour segmentation and background extraction, to machine learning and feature-based algorithms like histogram of oriented gradients (HOG) (Dalal & Triggs, 2005) and scale-invariant feature transform (SIFT) (Lowe, 1999), to deep learning-based frameworks such as faster region-based convolutional neural networks (R-CNN) (Ren et al., 2017), you only look once (YOLO) (Redmon et al., 2016), single shot detector (SSD) (Liu et al., 2016), RetinaNet (Lin et al., 2017), and Mask R-CNN (He et al., 2017) has significantly improved detection capabilities in both performance and efficiency. Despite these advancements, the unique constraints of sports environments demand further innovation. Factors such as varying ball sizes and colours across sports, high object speeds, lighting conditions, different fields or courts, and the need for seamless integration with tracking systems underline the need for tailored solutions.

This article reviews the state-of-the-art techniques in ball detection across sports, providing a concentrated and comprehensive overview of existing challenges, tools, and applications. This review aims to bridge gaps in the literature and inspire future research in this domain by categorizing approaches based on their underlying methods and sports-specific adaptations. The subsequent sections explore traditional and deep learning techniques, benchmark datasets, and performance metrics, culminating in insights for developing robust ball detection systems suitable for diverse sporting scenarios and hardware. The present article also strives to give useful insights on the topic of object detection, especially applied to sports, earned from meticulously analysing and reviewing these various sources of information.

The proposed article aims to gather and concentrate information on object detection applied to sports from reliable sources, and give valuable insight on the topic, earned from meticulously analysing and studying said information.

This article is divided into six sections: Related Work, details articles that also review object detection applied to sports balls. Methodology presents research questions that this review pretends to answer and explains the strategy behind the research, while Results, introduces and explains the articles that resulted from the research. Discussion presents a summary of what was gathered from the articles researched and brings about a discussion on the most valuable information. Finally, the Conclusions brings this review to a close with conclusions on what was learned and contributed.

Related work

This section analyses some of the most recent and relevant reviews on this topic. Object detection has been, and continues to be, a highly debated and researched topic. The sheer number of research articles and conference articles that appear when searching for the term “object detection” is evident. The number of reviews focusing solely on spheric objects is significantly more limited. Most reviews that focus on sports ball detection also concentrate on a single sport or one or two specific detection methods, leaving little room for newer or less popular techniques to shine. This article compiles recently published reviews (2015–2024) on ball detection, regardless of method or application, as long as their contents may be useful to ball detection in the sports industry.

The number of articles found reviewing or surveying ball detection applied specifically to sports was quite low. Most of this low amount is specifically aimed at football applications. Because of this, this review includes a couple of additional articles that either analyse ball detection but do not aim to apply it to sports or analyse player detection in some type of sport.

Mishra, Chandra & Jaiswal (2015) reviews works on player tracking in generic sports videos, weighing challenges such as occlusion and fast movements, and highlighting the evolution of tracking methods from 2003 to 2014. This review focuses solely on traditional techniques, including background subtraction, particle filters, and feature-based methods, applied across diverse sports datasets.

Using handball-oriented datasets, the article by Burić, Pobar & Ivasic-Kos (2018a) compares Mask R-CNN and YOLOv2 on ball detection. Different training configurations and datasets were employed, and YOLO emerged as the preferred method for real-time use.

In another article by the same author, Burić, Pobar & Ivašić-Kos (2018b) evaluates object detection methods for sports videos, once again focusing on action recognition in handball scenes. It compares Mask R-CNN, YOLO, and a mixture of Gaussian (MOG) and attempts to address challenges such as fast movements and diverse backgrounds.

Badami et al. (2018) review analyses of football videos, detecting various players’ body parts and the ball to identify poses, actions, and events. This work proposes an automated refereeing system that utilizes computer vision techniques, such as HOG, to address the limitations of human referees and reduce controversies in football.

Taken together, these four studies mark the field’s transition from rule-based computer-vision pipelines (Mishra, Chandra & Jaiswal, 2015) toward the first CNN-powered, real-time detectors (Burić, Pobar & Ivasic-Kos, 2018a), highlighting a simultaneous gain in speed (YOLOv2) and in segmentation granularity (Mask R-CNN).

The survey, authored by Kamble, Keskar & Bhurchandi (2019), analyses techniques for detecting and tracking balls in generic sports. It differentiates and categorizes methods for ball detection and tracking separately. For ball detection, the methods are categorized into four main areas: colour, size, position, and motion. Kalman filter, particle filters, and trajectory are used to categorize and compare tracking efficiency. Although this work only surveys traditional computer vision techniques, the authors suggest that “the future lies in the right employment of deep learning techniques (…)”. Challenges like occlusion, rapid motion, and lighting variations are highlighted. Evaluations utilize metrics such as precision and recall, with datasets spanning various sports and camera setups.

Deepa et al. (2019) investigates real-time tennis ball tracking using YOLO, SSD, and Faster R-CNN, aiming to create a cost-effective alternative to systems like Hawk-Eye for analysing ball trajectories. The models were evaluated based on their accuracy and computational power required to detect and calculate metrics such as toss height, toss distance, serve speed, and hit speed, using diverse datasets collected from multiple cameras.

The article by Cuevas, Quilón & García (2020) surveys techniques and applications for football video analysis, with a focus on player and ball tracking for event detection. It categorizes various methods based on audiovisual information, such as support vector machines (SVM), neural networks (NN), hidden Markov models (HMM), and probabilistic-based approaches. Methods based on external sources, such as webcasting and social networks, are also mentioned. The authors give great importance to the real-time applicability and various challenges encountered in ball detection, such as its small size, occlusion due to players, misdetections when the ball stands on top of a marked line on the field, and changes in the ball’s size, shape, colour, and speed.

Jagadeesh, R & S (2023) surveyed deep learning techniques, such as CNN and transfer learning, to classify cricket shots, not in real-time. The research evaluates models using metrics such as Precision, Recall, and confusion matrices, with a dataset of cricket shot images that have been augmented and pre-processed for consistency. Results highlight the effectiveness of transfer learning models, particularly visual geometry groups (VGG), in accurately detecting shot types. Using a pre-trained 19-layer VGG-19 model, the authors claim that this deep architecture is capable of recognizing various objects and patterns.

The review by Şah & Direkoğlu (2023) evaluates and compares player detection methods and techniques in field sports. The authors discuss conventional computer vision techniques, including HOG, SVM, and heat diffusion (HD), as well as CNN architectures with various image representations, such as gray, red-green-blue (RGB), shape information image (SIM), and polar-transformed shape information image (PSIM). Transfer learning was also compared using deep networks, such as Faster R-CNN, SSD, and YOLO. The study utilizes two field hockey datasets to evaluate precision, recall, and F-measure under consistent settings, taking into account occlusion and overlap ratios. Their results state that conventional methods and CNNs with SIM and PSIM image representations are preferred over deep networks when treating low-resolution images, which is common in field sports videos.

Akan & Varlı (2023) explores deep learning in football video analysis, addressing both player and ball tracking for real-time event detection. They review classical computer vision techniques and deep learning-based methods, highlighting the superiority of the latter in capturing high-level semantic information compared to traditional blob-based and feature-based approaches. The authors conclude that deep learning models, such as CNN and recurrent neural networks (RNN), overcome the drawbacks of classical approaches, performing with better efficiency and adaptability to various scenarios and complexities.

Yang et al. (2024a) review advancements in football player detection and tracking, emphasizing the growing demand for automated football video analysis. The study examines traditional methods and modern approaches using multi-object tracking (MOT). It evaluates different models using performance metrics such as detection accuracy, multi-object tracking accuracy (MOTA), and high-order tracking accuracy (HOTA).

Markappa, O’Leary & Lynch (2024) evaluates various YOLO models and versions in detecting football-related entities like players, referees, and balls. Using the SoccerNet dataset (Giancola et al., 2018), the training and assessment of the models in various Precision metrics and across multiple epochs revealed superior performance by YOLOv8 and YOLOv9.

To better articulate the novelty and analytical depth of this review, the authors introduce a formal taxonomy that organizes prior work across three core dimensions: object of detection, application scope, and technical approach. These categories are reflected in Table 1 and are further clarified below:

• Object of detection: We distinguish between articles that study sports ball detection (DB: detects balls) and those that focus exclusively on ball detection (FB: focuses on balls). This distinction highlights the specificity of the work and helps identify gaps where ball-only detection is understudied;

• Application scope: This dimension captures the review’s generalization and domain focus. Articles are marked based on whether they apply to sports contexts (AS: applied to sports) and whether they span multiple sports rather than focusing on a single discipline (GS: generalized sports). This allows us to assess the transferability of techniques across different sports;

• Technical approach: We classify the detection methodologies as traditional computer vision and machine learning techniques (CV: computer vision and machine learning) or deep learning-based methods (DL: deep learning). This distinction reflects the field’s evolution and helps identify which methods are underrepresented in reviews.

This taxonomy allows for a structured comparison between existing reviews. It underscores where this article extends beyond the scope of previous works, particularly by being one of the few to offer a ball-specific, sport-agnostic, and DL-inclusive review.

Most review literature does not generalize its work findings to different sports. Doing so will increase its usefulness to a broader audience. While someone looking into football ball detection techniques might be better off researching reviews that specifically target football, since there are many, most other sports applications would benefit more from a generalized view of what is available for ball detection in sports. The proposed review intends to do just that: review all sorts of algorithms to search for one that can perform well in any ball-playing sport. Also, these do not specialize solely in ball detection. It leaves room for improvement in systems designed solely for ball identification and tracking, as training them for a specific task will significantly enhance their effectiveness. Some methods or techniques may be more suitable than others for specific objects, depending on factors such as object speed, size, and complexity. This means that a specific detector that excels in both player and ball detection in a football game, for example, might not be the best detector for football ball detection only. That again differentiates this review, as it strives to achieve generalized ball-only detection.

The proposed review aims to gather and explore information on object detection applied to sports from reliable sources, providing valuable insights into the topic through meticulous analysis and study of the information. The article focuses on detecting sports balls but does not specialize in a single specific sport, aiming to give generalized insight into sports as a whole and the challenges one model might face, regardless of the sport to which it is applied. This review also aims to address the limitations identified in other articles. The survey by Kamble, Keskar & Bhurchandi (2019) does most of the mentioned goals, but unlike this review, it does not research deep learning methods. These techniques have been evolving rapidly and are something that cannot be overlooked at the current time. The review by Şah & Direkoğlu (2023) seems to do what this review intends to, but it is applied to player detection instead of ball detection.

What further distinguishes this review is its broad scope, encompassing ball detection across 12 different sports, including less commonly studied ones such as boccia, padel, and rugby. Unlike prior surveys focusing on a single sport (often football), this work compiles methods and insights from various disciplines, making it more broadly applicable. This review addresses the fragmentation by intentionally selecting articles that extend beyond mainstream sports and balancing coverage across traditional computer vision and modern deep learning approaches. It delivers a more inclusive and sport-agnostic perspective.

The following section will outline the methodology of this review, examining its objectives and the approach taken during the investigation.

Methodology

This chapter will outline the objectives of the proposed review and the methodologies employed in the research. The initial and main research was conducted on October 17, 2024, and an additional search was performed on May 22, 2025, aiming to update and identify additional relevant and recent studies. The search was conducted using predetermined inclusion and exclusion criteria and well-defined terms to gather articles that seemed relevant from various databases. The following subsections will outline the research questions this comprehensive review aims to address and the search strategy employed to gather the reviewed manuscripts.

Search strategy

The initial search utilizes well-defined search terms to gather relevant articles from various databases. The removal of duplicates and unrelated or incomplete articles follows this initial step. A manual screening of the remaining articles removes any unwanted articles. From the results, a qualitative analysis is performed to identify relevant articles, culminating in a small group of high-quality articles that assess the most recent developments in object detection. Various searches were made in Google Scholar through generic but relevant keywords and phrases, such as “Sports balls object detection”, for example. Articles found within Google Scholar from these searches are indexed in other databases. The databases from which the articles were gathered for this review were obtained from IEEE Xplore, Springer, Scopus, and ArXiv, which are the most representative databases for computer vision and computer science. These researches followed the following criteria:

Inclusion criteria

• 1.

  Articles written in English;

• 2.

  Peer-reviewed manuscripts;

• 3.

  Full text access to manuscript;

• 4.

  Articles published from 2015 onward;

• 5.

  Articles that follow the keywords (Object AND Detection AND Sports) AND (Balls OR Circular OR Circle).

Exclusion criteria

• 1.

  Articles that are categorized as reviews or surveys;

• 2.

  Articles with an invalid DOI (Digital Object Identifier);

• 3.

  Articles that do not present experimental results.

A total of 927 studies came up from the initial search: 182 from IEEE Xplore, 458 from Springer, 236 from Scopus, and 51 from other sources from generic searches in Google Scholar. From this relatively large group of articles, a pre-screening was done, immediately excluding a significant amount of articles for a multitude of reasons: 39 were rapidly removed for duplication between databases, 139 articles were excluded for not being in the computer vision or engineering’s subject areas (disciplines in Springer), 186 for the publication date (articles published before 2015), four for their language (articles not written in English), and 43 articles were removed for being reviews or surveys.

After this initial filtering, 516 studies underwent a manual screening, eliminating 479 articles. This procedure, undergone by all authors of this review, aimed to assess the quality and reliability of a article. Starting with an inspection of the article’s title and abstract, the assessment considered the relevance of the study and its objectives to the topic at hand, adherence to writing standards, structural integrity, the display of results, and whether evaluations and assessments were conducted. After this screening process, 37 articles remained, of which six could not be retrieved due to the lack of access, incorporating 31 relevant studies in the review.

An updated research was performed on May 22, 2025, aiming to gather more recent studies. This research focused on the indexing engine Google Scholar and the ArXiv database. The search in ArXiv included the terms “sports”, “object” and “detection” simultaneously, which resulted in a total of 81 articles. A pre-screening was performed, limiting the search to the subject of computer science, limiting the publication dates to 2020 onwards, and excluding reviews and surveys. From this pre-screening, 62 articles remained, 61 of which were from ArXiv, and one relevant article was found in IEEE Xplore through Google Scholar’s indexing. The articles underwent a manual screening, of which 55 were removed. The remaining seven articles, published between 2021 and 2024, were added to the review, bringing the total to 38 articles incorporated into the meta-analysis of this review. A flowchart illustrating this entire process is depicted in Fig. 1.

Summing up, all the articles gathered for reviewing consist of one (2.63%) article from 2015, four (10.53%) articles from 2018, seven (18.42%) articles from 2019, three (7.89%) articles from 2020, six (15.79%) articles from 2021, four (10.53%) articles from 2023 and 13 (34.21%) articles from 2024, amounting to a total of 38 articles reviewed.

Unlike prior reviews that often rely on a narrow or informal selection of studies, our approach involved a well-defined set of inclusion and exclusion criteria applied across multiple major databases. This process enabled us to compile a broad and diverse collection of 38 high-quality articles. By covering a wide range of sports, detection techniques, and publication years, this review offers a more comprehensive and up-to-date perspective.

In conclusion, after researching and screening the articles, an in-depth analysis of each individual article was conducted and compiled into the next section. The next chapter presents various statistics on the gathered articles through easy-to-read graphs and ratios, as well as some initial insights into what to expect from these results. The section provides a summary of each article and concludes with an organized table that lists the main points of each study.

Results

This chapter presents the research results and provides various statistics on these findings. As explained above, the search yielded 38 relevant studies for this review. As for the sport each study focuses on, 12 (31.58%) studies are centered around football/soccer, five (13.16%) in tennis, three (7.89%) articles are on basketball and table tennis, two (5.26%) articles focus on boccia, golf, handball, and volleyball each, and 1 (2.63%) article refers to padel, badminton and rugby each, the latter being the only sport with a ball that is not traditionally round. There are also three (7.89%) articles that do not pertain to any specific sport and instead research object detection applied to sports in general. Of all the sports-related articles in this review, seven (18.42%) focus solely on detecting humans. However, their contributions are still considered valuable to this analysis. At last, there is also one (2.63%) article that researches ball detection but is not sports-related, although it is a use-case that the authors mention.

These statistics can be visualized in Fig. 2. There is a clear bias towards football (and, to some extent, tennis), with the highest number of relevant studies found. This bias is due to research availability and is linked to the sport’s popularity. Football is the most popular sport in the world, and it shows. It is also possible that these sports can simply take advantage of computer vision more effectively than others, especially in detecting sports balls or players. This bias may skewer results, such as showing increased popularity and use of object detectors that are often used in football studies and systems, for example. Hence, it is recommended to take some of the following statistics with some caution.

Regarding the type of detectors, five (13.16%) articles solely research machine learning methods or traditional computer vision processing techniques, bearing no touch on deep learning-based methods, which in recent years have grown in popularity and shown capable of achieving greater accuracy in more complex tasks such as image classification and semantic segmentation. Of the deep learning-based methods used, a staggering 26 (68.42%) articles either employed some version of YOLO or extensively compared their developed algorithms to some version of YOLO as a benchmark. It demonstrates the increasing prevalence of this technique over the years since its debut in 2015. Also very popular, eight (21.05%) articles either extensively compared or used an established form of R-CNN, be this Faster R-CNN with five (13.16%) articles or Mask R-CNN with three (7.89%) articles. The SSD method also had a significant presence, being used in five (13.16%) of the studies. The authors of three (7.89%) articles self-proposed an enhanced CNN-based algorithm, and the RetinaNet method was used once (2.63%). The popularity of these base methods can also be analysed in Fig. 3. Note that a single article can adhere to multiple detection methods, whether by using, testing, or extensively reviewing, which is why the graph holds different values compared to the above statistics.

Deep learning methods can be further categorized according to the number of stages. As the name suggests, two-stage methods split object detection into two sequential stages, while one-stage methods bypass the need for region proposal. Table 2 aggregates every type of deep learning-based object detector researched throughout these works, divided by detector type.

This aggregation highlights the popularity of YOLO-based algorithms and the numerous sub-versions of YOLO, demonstrating their high versatility and adaptability. Many researchers also opt for a hybrid system, combining YOLO with other techniques to strengthen its weaknesses or build upon its already strong characteristics.

Taking an individual look at each base detector, Faster R-CNN has been a staple of object detection since its debut. While the original R-CNN model achieved desirable object detection results, it had some significant shortcomings, particularly in terms of speed. Faster methods were introduced, starting with Fast R-CNN, which was quickly followed by Faster R-CNN, one of the best versions of the R-CNN family. This method replaced a selective search algorithm to compute region proposals with a superior region proposal network (RPN) Zengin et al. (2020). Although this was a significant improvement over its predecessors, the delay in prediction is still its biggest limitation. However, it remains a good choice when object detection is done offline, and inference speed is not the highest priority. For more details on Faster R-CNN, refer to Ren et al. (2017).

As a later advancement to the Fast R-CNN, Mask R-CNN added a branch for predicting an object mask in parallel with the existing bounding box recognition. Utilizing pixel-to-pixel alignment, this state-of-the-art algorithm in image and instance segmentation achieves significant improvements in terms of accuracy and precision. It adds a small overhead to Faster R-CNN, inheriting its major weakness of high computation cost. For the more interested reader on the inner works of Mask R-CNN, refer to He et al. (2017).

The single-shot detector for multi-box predictions is one of the fastest methods for achieving real-time computation of object detection tasks. While the two-stage algorithms above can achieve high prediction accuracies, SSD removes the RPN and instead uses multiscale features and default boxes to improve its inference performance. The main limitations of this method are the decrease in image resolution to lower quality and the difficulty in detecting small-scale objects. For more details about SSD, refer to Liu et al. (2016).

The RetinaNet uses SSD detection capabilities to achieve superior results in accuracy-related metrics. This model is built upon the feature pyramid network (FPN) and utilizes focal loss to address some of the issues presented by the previous SSD algorithm. The result is a well-balanced algorithm that is not as fast as YOLO nor as accurate as Faster R-CNN, but it provides a safe choice for a wide array of applications. For more details on RetinaNet, the authors refer to Lin et al. (2017).

YOLO became one of the most popular models due to its neural network architecture, which enables high computation and processing speeds, especially in real-time, compared to most other training and detection methods, while providing an overall acceptable accuracy rate. This architecture allows the model to learn and develop an understanding of numerous objects more efficiently. However, it does not have all the advantages, as it fails to detect smaller objects accurately due to its lower Recall rate. Detecting multiple objects that are close to each other can also be challenging due to the limitations of bounding boxes. For more details on YOLO, refer to Redmon et al. (2016).

Table 3 summarizes how each of the assessed models performs in terms of inference speed and detection accuracy in an overall manner.

In general, and depending on algorithm enhancements and hardware available, YOLO and SSD are preferable solutions when real-time inference times are a necessity. In contrast, two-stage algorithms take precedence when higher precision and accuracy are of the utmost priority.

To evaluate all the different models and algorithms, various evaluation metrics (PaperspaceMetrics, 2019) are used, most of which are based upon the basic concepts of true positives (TP, correct detection—both the prediction and true value were positive), false positives (FP, incorrect detection—prediction was positive, but true value was negative), false negatives (FN, ground truth not detected—prediction was negative, but true value was positive) and true negatives (TN, correct misdetection—both the prediction and true value were negative). One of these metrics is precision, the ability of a model to identify only the relevant objects. It is the percentage of correct predictions, given by Eq. (1):

(1) $$Precision=\frac{TP}{TP+FP}.$$

Another metric is recall (or detection rate), the ability of a model to find all relevant cases, given by Eq. (2):

(2) $$Recall=\frac{TP}{TP+FN}.$$

F1-score is the weighted average of precision and recall, as given by Eq. (3):

(3) $$F1=\frac{2\ast TP}{2\ast TP+FP+FN}.$$

Average precision (AP) and mean average precision (mAP) are also widely used to assess a detection model. In essence, AP is a weighted sum of the precision values, where the weights are the increase in recall between consecutive points on the precision-recall curve. Considering n is the index of a point on the precision-recall curve, a graph that visualizes the trade-off between precision and recall of a model for different confidence thresholds, Pn is the precision at point n, Rn is the recall at point n, and R0 is assumed to be 0, the formula for AP is given by Eq. (4):

(4) $$AP=\sum _n(R_n-R_{n-1})\ast P_n.$$

The mAP is calculated by averaging AP values across all classes. Put simply, considering n the number of classes, mAP is given by Eq. (5):

(5) $$mAP=\frac{\sum (AP)}{n}.$$

Accuracy measures the overall correctness of a model’s predictions, encompassing all four basic concepts of TP, TN, FP, and FN, as given by Eq. (6):

(6) $$Accuracy=\frac{TP+TN}{TP+TN+FP+FN}.$$

Other metrics include the inference time (or inference speed), which corresponds to the calculation time for each frame in milliseconds and is critical to evaluate a model’s computational cost and real-time application, and Intersection over Union (IoU), which measures the overlap between the predicted and ground truth bounding boxes to determine if a predicted object is a TP or an FP. In sum, IoU is given by the area of overlap by the area of union between bounding boxes, mathematically written as in Eq. (7):

(7) $$IoU=\frac{A\cup B}{A\cap B}.$$

From the 38 articles evaluated, as depicted in Fig. 4, the most used metric was Inference Time, with 23 appearances (60.53%), revealing that real-time application (or, at least, the system’s computational cost) is a major concern when applying object detection to sports. The popularity of the remaining metrics is as follows: Accuracy with 20 (52.63%), precision and recall with 18 (47.37%), F1-score with 12 (31.58%), mAP with 10 (26.32%), AP with 6 (15.79%), and finally, IoU with 5 (13.16%) appearances.

The following subsections present brief reviews of the articles gathered, organized by sport, and chronologically. The studies focused on football are presented first, as they represent the majority of sports, followed by tennis. After that, the reviews on the remaining studies are presented, given the minimal number of studies on each of these remaining sports.

Summary of researched studies

In this subsection, the research articles reviewed are aggregated in Table 4, along with their year of publication, purpose, target objects, detection methods used, evaluation metrics used, and dominant challenges felt.

Table 4:

Listing and summarization of studies.

Study	Purpose	Target	Detection methods	Evaluation metrics	Challenges
Li et al. (2015)	To describe a novel approach for generating object proposals for ball detection	Sports balls	R-CNN, contouring, CHT,	mAP, Recall, Inference time	Small or distant objects, computational cost
Ali Shah et al. (2018)	To achieve an embedded approach for ball detection and tracking using image processing	Sports balls	Colour detection, Gaussian blur, contouring, centroid detection	Accuracy	Complex backgrounds, motion blur
Reno et al. (2018)	To present an innovative deep learning approach to tennis balls identification	Sports balls	CNN	Accuracy, Precision, Recall	Motion blur, high number of false positives
Burić, Pobar & Ivasic-Kos (2018a)	To compare YOLO and Mark R-CNN for handball detection in real-world conditions	Sports balls and humans	YOLOv2, Mask R-CNN	Precision, Recall, F1-score	Occlusion, computational cost, high number of false positives
Burić, Pobar & Ivašić-Kos (2018b)	To provide an overview of CNN detection methods for handball analysis	Sports balls and humans	YOLO, Mask R-CNN, MOG	Precision, Recall, F1-score	Occlusion, lighting, computational cost
Teimouri, Delavaran & Rezaei (2019)	To propose a low-cost ball detection method for football robots	Sports balls	CNN	Accuracy, Precision, Recall, Inference time	Lighting, motion blur, underfitting
Renolfi de Oliveira et al. (2019)	To investigate the performance of a vision system for object detection under constrained hardware, trained on a football dataset	Sports balls	SSD-MobileNet	Accuracy, Precision, mAP, Recall, F1-score, Inference time	Accuracy trade-off, computational cost
Barry et al. (2019)	To develop a lightweight real-time model (xYOLO)	Sports balls and goalposts	xYOLO	mAP, F1-score, Inference time	Accuracy trade-off
Deepa et al. (2019)	To describe a systemic approach for trajectory estimation of tennis balls	Sports balls	YOLO, SSD, Faster R-CNN	Accuracy, Inference time	Occlusion, motion blur
Calado et al. (2019a)	To propose a ball-detection system for boccia to motivate elderly physical activity	Sports balls	Contouring, centroid detection, colour detection	Accuracy	Lighting, underfitting
Calado et al. (2019b)	To have a versatile algorithm for boccia balls detection	Sports balls	Tiny-YOLO, HOG-SVM	Precision, AP, Recall, Inference time	Computational cost, high number of false negatives
Wang et al. (2019)	To present a high-speed stereo vision golf ball tracking system	Sports balls	ROI, P-tile, noise filtering	Accuracy, Recall, Inference time	Occlusion
Tian, Zhang & Zhang (2020)	To propose an anchor-free tennis balls object detector	Sports balls	YOLOv3	Accuracy, Precision, Recall, F1-score	Motion blur, small or distant objects
Zhang et al. (2020)	To propose an efficient solution for real-time golf ball detection and tracking	Sports balls	YOLOv3, YOLOv3-tiny, Faster R-CNN, Kalman filter	Precision, mAP, Inference time	Small or distant objects, underfitting
Sheng et al. (2020)	To propose a real-time one-stage algorithm based on feature fusion for table tennis balls detection	Sports balls	YOLOv3-tiny	mAP, Inference time	Underfitting
Fatekha, Dewantara & Oktavianto (2021)	To enhance detection algorithms for sports robotics in football using colour-based segmentation	Sports balls	Colour-based segmentation, morphological operations, contouring	Inference time	High number of false positives
Meneghetti et al. (2021)	To evaluate the performance of detection systems in constrained hardware	Sports balls	YOLO (v3, v4, tiny-v3 and tiny-v4), SSD (v2 and v3)	AP, Inference time	Computational cost
Hiemann et al. (2021)	To address the detection of small and fast-moving balls in sports in real-time	Sports balls	YOLOv3	Precision, AP, Recall, F1-score, IoU, Inference time	Underfitting, motion blur
Balaji, Karthikeyan & Manikandan (2021)	To research volleyball player detection using innovative Metaheuristic algorithms	Humans	Firefly, TLBO, Cuckoo Search	Accuracy, Precision, Recall	Occlusion
Pawar et al. (2021)	To create a robust tracking algorithm based on a custom rugby dataset to replicate industrial object detection	Sports balls	SSD-MobileNet	Accuracy, Inference Time	Underfitting, computational cost
Liu et al. (2021)	To propose a method that detect and matches players to their equipment with a single bounding box	Humans	YOLO, FPN	Accuracy, AP, IoU, Inference time	Occlusion, Overlap of bounding boxes
Hassan, Karungaru & Terada (2023)	To propose a method to detect handball events on football games	Sports balls and humans	YOLOv7, instance segmentation, Kalman filter, separating axis theorem	Accuracy	Occlusion, High number of false positives
Keča et al. (2023)	To explore the combination of HT and deep learning methods for ball detection on a low-powered device	Sports balls	Faster R-CNN	Precision, Recall, F1-score, IoU, Inference time	Computational cost
Li et al. (2023)	To improve immersiveness in live sports broadcasting via real-time 3D detection	Sports balls and humans	YOLOv3	Accuracy, Precision, Recall, IoU, Inference time	Complex backgrounds
Kulkarni et al. (2023)	To detect and identify table tennis strokes using ball trajectory data, eliminating the need for player-focused video and wearable sensors	Sports balls	YOLOv4, TrackNetv2	Accuracy, Precision, F1-score, Inference time	Accuracy
Zhao (2024)	To propose an improved version of YOLOv5 for football and player recognition	Sports balls and humans	YOLOv5	Precision, AP, mAP, Recall	Occlusion
Modi et al. (2024)	To propose a hybrid enhanced system for object tracking in videos	Sports balls	YOLO (v3, v5, v8), NMS. Gaussian blur	Precision, recall, F1-score, Inference time	Accuracy, occlusion, computational cost
Esfandiarpour, Mirshabani & Miandoab (2024)	To enhance detection algorithms for sports robotics in football	Sports balls	YOLOv8, colour detection, CHT, frame differencing	Accuracy, Inference time	High number of false positives, lighting
Li & Zhao (2024)	To propose an improved version of YOLOv5 to solve problems in tennis balls recognition	Sports balls	YOLO (v3, v4, v5), SSD, Faster R-CNN, FPN	mAP, Inference time	Computational cost
Decorte et al. (2024)	To introduce a predictive model for detecting padel hits based on audio signals, pose tracking, distance calculating framework, and a padel open dataset	Humans	YOLO, TrackNet, SED	Accuracy, F1-score	Occlusion
Fujimoto et al. (2024)	To enable table tennis ball detection with higher Accuracy using fine-tuning	Sports balls	Mask R-CNN, Faster R-CNN, RetinaNet	AP, Inference time	Computational cost
Luo, Quan & Liu (2024)	To improve the Accuracy of small fast-moving ball detection in sports	Sports balls	YOLOv8	Precision, mAP, Recall, F1-score	Small or distant objects
Li, Luo & Islam (2024)	To propose an hybrid YOLO-T2FLSTM detection system for basketball players and action recognition	Humans	YOLO-T2LSTM	Accuracy, IoU	Occlusion, Underfitting
Yang et al. (2024b)	To propose a novel model, YOLO-HGNet, to enhance feature learning, applied to badminton action recognition	Humans	YOLO-HGNet	Precision, mAP, Recall, F1-score	High number of false positives, computational cost
Hu et al. (2024)	To address the challenges of complex multi-object occlusion in basketball	Humans	YOLOv8	Accuracy	Occlusion
Fu, Chen & Song (2024)	To propose a YOLOv8n-based model capable of solving the challenges of large deformation and small size in football object detection	Sports balls and humans	HPDR-YOLOv8	Precision, mAP, Recall, Inference time	Computational cost
Solberg et al. (2024)	To automate the creation of player-specific football highlight videos	Sports balls and humans	YOLO	Accuracy, Inference time	Accuracy, Inference time
Yin et al. (2024)	To address the challenges posed by frequent occlusions and visual similarity of players in basketball	Humans	YOLOv8	Accuracy	Occlusion

DOI: 10.7717/peerj-cs.3079/table-4

Analysis of Table 4 shows clear recurring challenges across sports and models. Occlusion and small-object recall are the most frequent limitations, especially in sports such as table tennis and volleyball, where the ball is fast, small, or partially hidden (e.g., Hiemann et al. (2021)). Many studies also cite trade-offs between model size and detection accuracy when optimizing for real-time inference on constrained hardware (e.g., Hassan, Karungaru & Terada, 2023, Sheng et al. (2020)). To address these, some solutions appear repeatedly. Attention modules (SE, CBAM, and Transformer blocks) are often used to improve focus on small or occluded targets, while GhostNet-based pruning and feature fusion are commonly employed to maintain speed without sacrificing accuracy (e.g., Fu, Chen & Song, 2024). These patterns suggest that combining lightweight design with spatial feature enhancement yields the most consistent gains. The following discussion builds on these observations to address RQ2 and RQ3, particularly regarding cross-sport model design and performance trade-offs.

From a different perspective, the studies reviewed in this article utilize a variety of datasets tailored to the unique challenges of ball detection in sports. Key datasets include ImageNet, a large-scale collection with millions of labelled images that serves as a foundation for model pre-training. However, it may not be directly applicable to specific sports ball detection. The COCO dataset, comprising over 200,000 images across 80 categories, is widely used for object detection tasks, including sports ball identification, due to its diverse environmental conditions. Additionally, datasets like SoccerDB are designed explicitly for football-related ball detection, incorporating a range of lighting conditions and ball types. The RoboCup Dataset, used in robotics competitions, features diverse ball patterns and lighting conditions, making it valuable for testing the generalizability of ball detection systems across different settings. Table 5 summarizes the datasets in the reviewed studies.

Table 5:

Summary of Datasets presented in the reviewed studies.

Study	Dataset name	Size (Images)	Key features	Annotations
Teimouri, Delavaran & Rezaei (2019)	RoboCup	N/A	Football, various ball patterns, lighting	Ball position
Modi et al. (2024)	D-FL Soccer Ball Detection Dataset	>1,000 images	Soccer ball, diverse environments (outdoor, indoor)	Ball position, size
Tian, Zhang & Zhang (2020)	MS-COCO + Custom Tennis Dataset	>200,000 images	Sports balls, various environments	Object detection (balls, players)
Burić, Pobar & Ivasic-Kos (2018a)	Custom Handball Dataset	>500 images	Handball, fast motion, occlusion	Ball position
Deepa et al. (2019)	Custom Tennis Dataset	N/A	Tennis ball, high-speed motion, occlusion	Ball trajectory

DOI: 10.7717/peerj-cs.3079/table-5

Synthesizing the findings across football, tennis, and other sports reveals several recurring challenges and solution strategies. Football, with its larger field of view and frequent occlusions, has driven the adoption of one-stage detectors such as YOLOv5 and its lightweight variants (e.g., YOLOv5-Lite), which balance speed with robustness to partial visibility. In tennis, the key issue is high-velocity motion coupled with a smaller frame region of interest, leading researchers to favor detectors like SSD with temporal smoothing or motion priors to handle fast, linear trajectories. Meanwhile, in smaller-object sports such as table tennis and badminton, the focus has shifted to enhancing detection of tiny and blurry balls, often through pruning strategies (e.g., GhostNet modules) and attention mechanisms (e.g., SE or CBAM blocks).

Some techniques developed for one sport show promise in others. For instance, feature-fusion techniques used to amplify small object cues in table tennis (e.g., Sheng et al. (2020)) have since been applied in football-specific detectors, such as HPDR-YOLOv8 (Fu, Chen & Song, 2024), indicating the cross-sport transferability of model enhancements. Similarly, multi-camera 3D triangulation, as explored in football (Li et al., 2023), is largely absent in studies of tennis and badminton, representing an untapped opportunity for these domains, especially in handling occlusion or improving bounce detection.

This cross-sport synthesis highlights how confident architectural and training choices (e.g., one-stage vs. two-stage, lightweight pruning, temporal fusion) reflect the underlying constraints of each sport, while also suggesting a roadmap for future research through the adaptive reuse of successful strategies across domains.

To enhance the comparative analysis of the reviewed studies, a structured quality assessment framework was considered. Each of the 38 studies was evaluated across five dimensions: (1) Dataset availability, assessing whether the dataset is publicly accessible, conditionally available, or not accessible; (2) Dataset size and diversity, reflecting the volume of data and the extent of variability in conditions such as lighting, motion, and occlusion; (3) Annotation quality, indicating the level of detail in the dataset labeling (e.g., bounding boxes, trajectories, semantic annotations); (4) Method reproducibility, considering the clarity of methodological description and the availability of source code or implementation details; and (5) Evaluation rigor, measuring the comprehensiveness of performance evaluation, including the use of established metrics and comparative or ablation analyses. Each dimension was scored on a scale from 0 to 2, resulting in a maximum possible score of 10, as depicted in Fig. 5. This scoring system enables a relative ranking of the studies, providing a more objective basis for assessing their methodological quality and applicability.

Finally, to address the need for a clear, comparative overview of algorithm performance across various sports, Table 6 is presented. This table aggregates the best-performing detection methods identified for each sport type reviewed, alongside their accuracy, inference speed, strengths, and limitations. This comparative view is particularly valuable for identifying algorithms that generalize well across multiple sports or are better suited for specific contexts, such as embedded systems or offline processing.

Table 6:

Algorithm performance by sport.

Sport	Best algorithm	Accuracy/mAP	Inference time	Strengths	Challenges
Football	YOLOv8 (Hybrid Enhancements)	High	Real-time	Handles occlusion, real-time tracking	Issues with distant ball tracking
Tennis	SSD-MobileNet	Medium-high	Fast	Lightweight, CPU-friendly	Sensitive to lighting
Table tennis	Optimized YOLOv3-Tiny	Very high	Very Fast	Great for small object detection	Needs high-res input
Basketball	YOLO-T2LSTM	Very high	Fast	Accurate action recognition	Complex model fusion
Golf	Faster R-CNN	Very high	Slow	High precision tracking	Not real-time capable
Handball	Mask R-CNN	Very high	Slow	High segmentation accuracy	High computational cost
Volleyball	YOLOv3 (Enhanced)	High	Fast	Improved detection of fast-moving objects	Training data limits generalizability
Boccia	Tiny-YOLO	High	Fast	Good color-based accuracy	Lighting sensitivity, narrow FOV
Rugby	YOLOv3	High	Fast	Compact architecture for embedded use	Lower resolution support
Padel	YOLO + Audio SED	High	Fast	Context-aware detection	Audio interference risks
Badminton	YOLO-HGNet	Very high	Fast	Enhanced classification and detection accuracy	Requires tuning and large datasets

DOI: 10.7717/peerj-cs.3079/table-6

In addition to the detailed qualitative summaries provided, a quantitative benchmarking summary and comparison of key detection models based on their reported performance across the reviewed studies. Table 7 presents aggregated values for the most frequently used object detection methods, highlighting trade-offs in accuracy, inference speed, and detection robustness. For instance, YOLO-based detectors achieved the lowest inference times (20 ms) with competitive accuracy (nearly 85%) and F1-scores (approximately 83%), making them ideal for real-time sports applications. In contrast, Mask R-CNN achieved the highest accuracy (around 92%) and F1-score (about 91%), though at the expense of significantly greater computational cost (about 120 ms). This comparative analysis underscores the importance of selecting detection models based on application-specific needs such as speed, precision, and computational constraints.

Table 7:

Comparative performance metrics of object detection models used in sports applications.

Model	Avg accuracy (%)	Avg inference time (ms)	Precision (%)	Recall (%)	F1-score (%)
YOLO (various versions)	85	20	84	82	83
SSD	78	25	76	75	75
Faster R-CNN	90	100	89	91	90
Mask R-CNN	92	120	91	92	91
RetinaNet	87	60	85	83	84
Traditional CV/ML	70	15	65	60	62

DOI: 10.7717/peerj-cs.3079/table-7

Discussion

This review consolidates knowledge on ball detection in sports, covering a broad range of applications, methodologies, challenges, and results, and offers valuable insights that inform this discussion. This chapter summarizes, analyzes, and discusses the main findings, emphasizing object detection methods, evaluation metrics, strengths, and challenges while also mentioning some options to address various challenges.

The research highlighted a broad spectrum of methodologies. A clear trend emerged in the evolution of object detection techniques from traditional computer vision methods to advanced deep learning models. Earlier approaches, such as CHT and color-based segmentation, may have been computationally efficient but lacked robustness and generalizability in complex environments. These traditional methods are still very relevant for simple tasks or constrained systems where the need for simplicity and speed outweighs Accuracy, such as in robotic competitions or low-end devices.

Although closely related, machine learning and deep learning have been growing more distant due to the rapid development of the latter. While the core of deep learning is feature learning, which is automatically evolved through multi-layer neural networks, machine learning methods require the manual selection and design of features. Machine learning remains highly relevant and useful, particularly for specific yet complex tasks where traditional computer vision may not be sufficient to achieve the desired results on its own.

Deep learning methods dominated recent studies, particularly one-stage detectors like YOLO (in its various versions) and SSD, as well as two-stage models such as Faster R-CNN and Mask R-CNN. YOLO was especially prevalent, cited in more than 60% of the studies, due to its detection speed, making it ideal for real-time applications, especially in constrained hardware settings. Innovations like YOLO adaptations and hybrid approaches combining YOLO with traditional methods or advanced modules further extended its applicability. Two-stage methods excelled in scenarios requiring precision, such as small object detection or occlusion handling. However, these methods are considerably slower, falling short in scenarios where real-time application is required.

One important note to keep in mind is that modern deep learning does not completely replace machine learning or conventional computer vision. On the contrary, some of the most insightful and impressive studies focused on a hybrid system, which combined the versatility of deep learning for the most complex tasks with the efficiency of traditional techniques or machine learning for simpler tasks or enhancements. For example, YOLO is widely regarded for being lightweight and quick on prediction and detection, but there is a significant trade-off in terms of Accuracy. Other algorithms and techniques, such as the enhancements mentioned above, can be expertly applied to the system to strengthen that weakness without compromising its speed. Tweaks can also be made to improve the system faster, all coming down to priorities. A good example of a use case for this enhancement is object detection applied to tennis. The tennis ball moves at high speeds, requiring significant computational power for real-time application, making lightweight detectors desirable. However, the object is also relatively small, making it challenging for a detector like YOLO to achieve desirable performance results. A hybrid model would likely be favoured in this situation, offering a balanced solution.

The following confusion matrix (Fig. 6) illustrates the most commonly employed detection methods in each of the researched sports. While it was already stated that YOLO was the most popular object detector and football was the most popular sport, it is interesting to observe the contrast between some sports and their approaches to more precise and accurate, or faster, algorithms. It is also worth reiterating that a big part of the articles gathered were studies on football, and although YOLO was still, by far, the most popular object detector when ignoring football studies, it has got a significant boost in the various statistics related to the popularity of each object detector.

Upon closer examination of each sport and its respective methods, the differences in challenges are readily apparent. In football, the ball can be kicked at speeds exceeding 100 km/h, making its detection particularly challenging, especially from a distance. This mainly utilizes fast algorithms, such as one-stage methods or traditional computer vision techniques. Football detection also suffers significantly from occlusion due to the large number of players on the field simultaneously. Since it is a sport played outdoors, varied lighting and weather pose hardships.

While YOLO is still popular within tennis detection systems, given its relatively weak Accuracy and the small size of a tennis ball, more systems rather lean towards more robust algorithms. Given the small size of a tennis ball, YOLO has a harder time accurately detecting it. The most common issue recorded in tennis studies is motion blur, given the speed at which the balls are served. Occlusion may also pose a problem for continuous tracking if detection cannot be done from a top-down view, although it is much less problematic than in football. Table tennis differs from regular tennis due to smaller object sizes, which necessitate the use of conventional detection or R-CNN-based systems. However, these are rarely applied in real-time, according to reviewed studies. provides a good example (Fujimoto et al., 2024), who fine-tuned Mask R-CNN, Faster R-CNN, and RetinaNet models (generally slower models) for improved accuracy. The difficulties faced in table tennis are identical to those in regular tennis.

Due to the lack of relevant studies on the other sports, it is challenging to accurately determine which detection systems are most suitable for each occasion. It is safe to say that one-stage algorithms are preferred over two-stage algorithms due to the need for quick detection and decision-making, especially when real-time application is involved. An example of this is football, where YOLO and SSD, the fastest models reviewed, were essentially the only deep learning-based object detectors used due to the need to detect and referee the game in real-time, similar to the VAR system used in today’s professional football. In contrast to refereeing-type applications, systems that aim to provide visual training and clues to players do not need to be applied in real-time and favor the higher precision and accuracy offered by slower algorithms.

Many challenges were repeatedly reported throughout the studies, showcasing the difficulties that object detection in sports faces, independent of the sport itself. The most common were occlusion, lighting and background variations, underfitting, small or distant object size, the detection of fast-moving objects, and hardware constraints.

Occlusion is, generally speaking, always a problem in the detection and tracking of objects. There are many examples of occlusion in sports, such as players in team sports, trees in large golf courses, light posts, fences, and goalposts, as found in football, among others. It is easy to understand how occlusion would be a major problem for sports with multiple players on the field or court, such as volleyball, handball, basketball, and rugby, as well as sports with multiple objects to track, like boccia. Almost all articles mentioned occlusion, with many proposing different methods for combating it. The most common approaches were image processing and data augmentation. Zhao (2024) provides an example that improves occlusion handling with a multiscale feature extraction module. The angle at which the system is trying to detect the object is also essential. In most cases, a top-down view is extremely helpful in avoiding object occlusion in sports, although this is rarely possible, especially in outdoor activities. Another way to minimize the effects of occlusion is to use more cameras. Using more sensors not only increases Precision in tasks like distance measurement and identifying distant objects, but it may also help counteract occlusion if a system can access different angles and points of view. Depth information can also help significantly in detecting partially occluded objects.

Lighting and background variation problems can also be, for the most part, attributed to underfitting as a cause of heavy training under particular conditions. This is fine when the system is to be fixed and used indoors. Still, the performance is heavily compromised if generalizability is essential, as in outdoor environments where the time of day and weather conditions may vary, such as in football. This can be aided through robust pre-processing or data augmentation, such as enhancing a training dataset with diverse images that simulate different lighting conditions, shadows, brightness, contrast, and saturation levels. A good example can be found in the article authored by Keča et al. (2023), where a handmade dataset of images under various lighting conditions was used for training and testing. Some examples of image processing techniques that can also be applied to accommodate varied lighting conditions include shadow removal, brightness adjustment, and colour normalization. The choice of sensor is also significant, as some have a wider dynamic range or can compensate for changes in illumination. Infrared cameras are also an option, as they are immune to lighting variations.

Problems with small or distant object detection were primarily found in high inference speed systems, mostly using a YOLO-based method, due to their accuracy trade-off for computational efficiency. This occurs when the image resolution is too low, making it difficult for the detector to identify objects with insufficient pixel resolution. Luo, Quan & Liu (2024) authored an interesting research on enhancing YOLOv8 for better small object detection.

The movement of objects, which causes motion blur, is also a very common challenge in sports. This is especially problematic when fast movement is combined with small-sized objects, such as in golf, tennis, badminton, and padel. To avoid motion blur, the shutter speed must be fast enough to freeze the scene, allowing the object to move as little as possible during the exposure time. For this, reducing the camera’s exposure time is crucial. However, reducing the exposure time will also make the image darker, as there is less time to capture light, so the choice of sensor and lens is important. As mentioned above, regarding varied lighting conditions, some image processing techniques can be applied to increase or normalize brightness and contrast, which helps with these darker images. Very high-framerate cameras can be used, as was the case with the 810 FPS binocular camera by Wang et al. (2019), to combat quick movement. However, this usually comes at a cost and may bring other drawbacks, such as reduced detection distance. The anchor-free detector by Tian, Zhang & Zhang (2020) was another good approach to this issue. Generally, there must be a compromise between a high frame rate, sufficient image resolution, and specificity in training.

One of the most common challenges was hardware constraints, which are prevalent in works that integrate robotics or embedded devices. These systems often lack a GPU, forcing the CPU to handle all processing. Generally, GPUs outperform CPUs due to their parallel processing capabilities, making them faster and more efficient for training and inference of deep learning models. This is especially the case when working with large datasets or real-time applications. Most scholars train their detection model on a medium to high-power machine and only then install it on the constrained hardware system. Achieving real-time inference speeds while maintaining favorable levels of accuracy remains one of the biggest challenges in object detection. When faced with hardware constraints, traditional computer vision techniques, hybrid models, and YOLO-tiny variants were the most popular detection methods, especially in embedded devices with no GPU available. Barry et al. (2019) managed to create a detector based on YOLOv3-tiny and XNOR layers, running on a Raspberry Pi 3B with improved performance over other base YOLO variants, albeit still not reaching consistent double-digit inference speeds, proving how challenging it is to run smooth real-time object detection in constrained hardware. On the other hand, base (non-tiny) YOLO variants, such as YOLOv8 and SSD-based algorithms, are ideal for systems that aim to achieve real-time performance but can rely on more powerful computers, generally equipped with a dedicated GPU.

Evaluation metrics varied across studies, reflecting the diversity of applications and priorities. Based on how different object detection algorithms function, some evaluation metrics may be more or less valuable in assessing overall accuracy and efficiency. For example, detectors that do not use bounding boxes will have no use for IoU. In this comprehensive review, all deep learning-based methods analysed use bounding boxes. Hence, the metrics used are very similar, varying more based on the application than on the object detector used. Systems not meant for real-time applications might not prioritize computational costs as highly. Hence, inference time may be a more useful metric for faster detectors, like one-stage algorithms. Many articles emphasized real-time capabilities, underscoring the need for high FPS without compromising accuracy, hence why the use or benchmark of some version of YOLO was a common sight in these articles. On the other hand, many researchers didn’t prioritize speed; instead, they focused on the best mAP values rather than fast inference speeds. One such example is given by Tian, Zhang & Zhang (2020), who makes use of varied performance-assessing metrics such as precision, recall, F1-score, and accuracy, but pays no heed to inference time, which may be in line with the aims of the authors if real-time application is not applied. This disparity underscores the dual priorities in sports analytics: delivering actionable insights or real-time decision-making systems and achieving high post-game analysis precision.

To check how well a system performs, it mostly comes down to training and testing. For that, an adequate dataset is crucial. Various components make or break a dataset for object detection, the most relevant being the number of images or videos, the diversity or specificity of the data, the quality and precision of the frame annotations, and the ratio of data allocated to the training, validation, and testing tasks. With the use of annotated images to have a reference of what the ground truth is, the training portion of a dataset is used to train the model, the validation portion is used to evaluate the model’s performance during training and helps prevent overfitting, and the testing portion is used to evaluate the final performance of the trained model on data unused in the steps before. When training and testing a detection model, precisely annotating images is necessary to achieve results that closely approximate the ground truth. The annotation quality reflects how closely the bounding boxes align with the actual targets, thereby affecting performance. Annotation consistency and quality can vary greatly between datasets. Inaccurate and inconsistent annotations, such as loosely drawn bounding boxes or inconsistent labelling of partially occluded objects, can negatively impact model training and evaluation. Conducting manual correction may be needed to ensure reliable benchmarking.

If a dataset is too small or not diverse enough, it may lead to overfitting (Montesinos López, Montesinos López & Crossa, 2022). A model suffers overfitting when it learns the training data too well, including random noise and specific features, leading to poor performance on new, unseen data. On the other hand, a dataset with a very large or overly diversified collection of data may cause underfitting. A model suffers from underfitting when it is too simple to capture the underlying patterns and relationships in the data and will perform poorly on both the training data and the new, unseen data. Many reviewed studies utilize large, diverse, publicly available datasets like ImageNet or COCO. These offer the convenience of having a ready-to-use large dataset with thousands of images and many classes but may overfit a model for specific tasks with irrelevant information. Public datasets specifically tailored for sports, specially SoccerDB, also saw significant usage in the works reviewed. Starting from a pre-trained model is often effective, but it is worth creating or using a tailored dataset for the task at hand, with caution to avoid excessive simplification that leads to underfitting. Hiemann et al. (2021) does a commendable job of testing and comparing different models trained on different datasets.

Another critical consideration is class imbalance, where some object classes are significantly underrepresented. This is particularly relevant in sports where lighting, occlusion, and motion blur may disproportionately affect balls of different types. Models trained on imbalanced datasets often become biased, resulting in poor performance in underrepresented classes. Techniques such as data augmentation, class weighting, or resampling are commonly employed to mitigate this issue. When handling a generalized model intended for application to various sports, this is especially problematic. A model trained in one sport (such as football) may not generalize well to another (such as golf) due to differences in ball material, size, color, speed, or environment. In such cases, to combat the challenge proposed by domain transfer, fine-tuning domain-specific data or using adaptation methods is recommended to maintain performance across different applications.

Lastly, comparing developed systems and algorithms to deployed real-world examples, such as video-assisted refereeing (VAR) technologies in football, provides exceptional value in terms of quality assurance and benchmarking. A great example is provided by Wang et al. (2019), who directly tested and compared their system against GC2 (ForesightSports, 2025), a professionally used and certified device with industry-leading performance in tracking the trajectory of golf balls. Additionally, Deepa et al. (2019) briefly mention Hawk-Eye (HawkEye, 2025), a professionally-used computer vision system that tracks the trajectory of balls (and much more), displaying a profile of its statistically most likely path as a moving image. It is implemented in several major sports, including tennis, football, and badminton.

Table 8 presents all the reviewed articles, focusing on the strengths and limitations of the methods/algorithms each article provided:

Overall, YOLO is expected to continue being one of the most predominantly used methods when detection speed and Inference Times are of the utmost priority, as well as when constrained hardware is a challenge. The appearance of an open, faster, and more reliable deep learning method would shake the world of object detection. The difficulties in detecting tiny or distant objects are ever-present, especially in faster yet less reliable methods such as YOLO. Hybrid systems with quicker and more robust techniques, enhancements, and augmentations are expected to produce better results. Another major problem is occlusion, arguably the most prominent challenge in object detection, which has little to no way of fully countering without spending a larger budget on multiple cameras or systems, such as radars, that can work through obstacles.

Conclusions

This review analyzed state-of-the-art techniques in ball detection across various sports, covering traditional computer vision, machine learning, and deep learning methods. Based on 38 key studies, it categorized detection approaches, benchmarked model performance, and identified major challenges, including high-speed motion, occlusion, and the demand for lightweight models. The review highlighted the strengths and limitations of various methods, as well as the importance of optimizations and enhancements.

Deep learning dominates complex detection tasks. One-stage detectors, such as YOLO and SSD, are preferred for real-time use due to their speed, although they struggle with small or distant objects. Two-stage detectors such as Faster R-CNN and Mask R-CNN offer higher accuracy but are often too resource-intensive for real-time applications. Hybrid models that combine traditional and modern methods have shown promise in addressing issues like occlusion, lighting changes, motion blur, and hardware limitations. Common evaluation metrics, such as inference time and accuracy, reflect the need to strike a balance between speed and precision.

The growing use of AI in sports analytics demands scalable, accurate ball detection systems. Applications range from autonomous refereeing to enhanced broadcasting. Despite advances in detection accuracy and speed, balancing model complexity with processing constraints remains a key challenge. Sports environments differ significantly in terms of lighting, pace, and ball characteristics, making universal solutions challenging. The lack of general-purpose models that work across various sports without significant tuning is another significant gap. Additionally, few studies focus on deployment on low-power or embedded systems—an essential step for real-world use. Continued innovation is crucial for enhancing sports technologies, facilitating better decision-making, in-depth analysis, and increased audience engagement.