A survey on gait recognition against occlusion: taxonomy, dataset and methodology

Motivation

Gait recognition has yet to be substantiated as a more dependable biometric modality compared to well-established techniques like fingerprinting, facial recognition, and iris recognition (Bouchrika, 2018), its practical application remains limited by performance degradation due to occlusion. Real-world conditions inevitably introduce occlusion, whether from clothing, carrying items, or partial view. Therefore, it is vital important to find the challenges of gait recognition in real-world settings. As discussed in introduction part, one of the foremost challenges encountered in real-world gait authentication pertains to occlusion (Kumari & Bharti, 2023). Imagine the scenarios below: (i) Dynamic occlusion: A significant number of individuals are present within the frame, their visibility may be compromised, thereby impeding the accurate detection of their walking patterns, as shown in Fig. 2. (ii) Static occlusion: When individuals walk behind objects like pillars, trees, or vehicles, their legs and lower body may be hidden from view, making it difficult to capture their distinctive gait features, as shown in Fig. 2. (iii) Other occlusion: When subject cut out or too close to lens, the occlusion increases. When the media quality cannot be guarantee cause by bad transmission or signal interference, salt and pepper noise (Azzeh, Zahran & Alqadi, 2018) or frame lost (Lin et al., 2018) may hinder the algorithm as well, become a potential occlusion.

The presence of occlusions are hindering successful implementation of gait recognition in real-world. Overcoming this obstacle is crucial to realizing the goal of deploying gait recognition technology in practical scenarios. Some researches have been conducted to solve this thorny problem. Unfortunately, to the best of our knowledge, there are only few review articles (Kumari & Bharti, 2023) focusing on gait recognition under occluded conditions. This study aims to bridge the gap between academic research and real-world application by focusing on gait recognition under occluded conditions. It strives to provide insights, methodologies, and solutions to overcome the challenges posed by occlusions and advance gait recognition technology in practical scenarios.

Contribution

This article surveys recent advancements in gait recognition under occlusion conditions up to October 2024. The primary contributions are summarized as follows:

• We propose a structured taxonomy that systematically categorizes: (i) real-world occlusion types encountered in gait recognition, (ii) datasets that incorporate occlusions as sample variations, and (iii) methodologies designed to address the unique challenges posed by occluded gait sequences.

• We conduct a detailed review guided by this taxonomy, emphasizing datasets and state-of-the-art methods in occluded gait recognition. This comprehensive review offers critical insights for ongoing research and lays a foundation for developing novel algorithms in occlusion-resilient gait recognition.

• We outline potential directions for future research, grounded in current studies and established biometric frameworks, to guide researchers in addressing the complexities of occluded gait recognition in practical applications.

Target audience

This article will be of particular interest to researchers and developers working in the field of biometric identification, especially gait recognition. As occlusion presents a significant challenge to applying gait recognition in real-world scenarios, computer vision scientists seeking to advance occlusion-robust gait recognition algorithms will benefit from the proposed taxonomy and extensive literature review. Additionally, the suggested future research directions will aid investigators focused on improving gait recognition for practical applications. Beyond the research community, this survey offers value to government and private sector stakeholders interested in deploying biometric solutions for purposes such as surveillance and access control. By summarizing key datasets, methodologies, and gaps in current occluded gait recognition capabilities, this article provides guidance to support development of more robust systems capable of overcoming limitations imposed by occlusion. Ultimately, a wide audience encompassing technology researchers, developers, and end users stands to gain from progress toward accurate gait recognition under real-world conditions.

Organization

The structure of this article is as follows: In “Review Methodology”, we detail the workflow and methodologies applied in this survey, including data sources, search strategies, and screening techniques used to identify and assess literature focused on gait recognition under occlusion conditions. “Deep Gait Recognition” provides a comprehensive overview of the gait recognition pipeline, covering essential processes such as data acquisition, feature representation, dimension reduction, classification, and performance metrics, along with a discussion of potential application scenarios. This section establishes foundational knowledge for addressing challenges related to occlusions. In “Taxonomy”, we present a structured taxonomy specifically developed to classify different approaches for tackling occluded gait recognition. “Review of Datasets” surveys datasets pertinent to occluded gait recognition, examining the resources available for algorithm development and testing. “Review of Literature” provides a comparative evaluation of state-of-the-art methods in occluded gait recognition, with experimental results on widely-used datasets. Finally, in “Future Directions”, we synthesize our insights and outline prospective research directions for enhancing gait recognition systems under occlusion.

Data acquisition: gait cycle capture

The gait cycle can be defined as the temporal interval between two consecutive instances of a repetitive characteristic associated with human locomotion by foot. Namely, the gait cycle demarcates the time from one particular phase or event of the walking motion to its subsequent repetition, such as from initial contact of the same foot to the next initial contact of that same foot (Kharb et al., 2011).

The human gait cycle can be delineated into seven distinct events that partition it into stance phase and swing phase, as shown in Fig. 5. Specifically, initial contact marking the beginning of stance phase followed by opposite toe off, heel rise, opposite initial contact and toe off which concludes stance and commences swing phase comprising of feet adjacent, tibia vertical. Stance, also termed the support or contact phase, encompasses four periods-loading response, mid-stance, terminal stance and pre-swing from initial contact to toe off. Meanwhile, swing spanning from toe off to the subsequent initial contact contains initial swing, mid-swing and terminal swing. Notably, a full gait cycle is quantified as the cycle time bifurcated into stance time spent on the ground and swing time airborne between footfalls. Kharb et al. (2011) Occlusion can impede accurate gait recognition when obstacles block view of the subject’s motion. Occlusion obscure gait cycles and limit the information gained of later gait recognition process (Kumari & Bharti, 2023).

Gait cycle capture is a fundamental component of the gait recognition pipeline. This process involves the use of data acquisition devices to bridge the gap between the physical and digital domains, facilitating the translation of human locomotion into quantifiable data. The devices employed can be categorized into four primary types: camera, radar, accelerometer, and sensor (Rani & Kumar, 2023). Camera systems, for instance, are commonly employed in vision-based gait recognition and offer the advantage of non-intrusively capturing gait patterns from a distance. Radar technology provides an alternative means of capturing gait, offering advantages in terms of penetration through obstacles and the ability to measure velocity. Accelerometers, often used in wearable devices, are effective in capturing dynamic movements and have been integral in the development of model-based gait recognition techniques. Sensors, on the other hand, provide a complementary perspective by capturing pressure distribution and have been useful in clinical gait analysis. In Table 1, we divided mainstream gait data acquisition devices into those four categories with their advantages and disadvantages.

Vision-based gait recognition has gained widespread attention as a practical biometric identification method due to its numerous advantages (Singh et al., 2018; Khan, Farid & Grzegorzek, 2021), as shown in Table 1. One key benefit is its non-intrusive nature, as it does not require physical contact with the subject (Derbel, Vivet & Emile, 2015). Unlike sensor-based approaches, such as accelerometers or gyroscopes, cameras can capture gait data from a distance, making them ideal for public and large-scale applications where attaching devices to individuals is impractical. This remote, real-time capability makes vision-based systems particularly well-suited for surveillance and security applications. The practicality of vision-based gait recognition is further enhanced by the widespread availability of camera technology. Cameras, particularly RGB cameras and depth sensors like Microsoft Kinect (Zhang, 2012), are already widely deployed in various environments such as public spaces and buildings. This existing infrastructure reduces the cost and complexity of implementing gait recognition systems. In contrast, wearable sensor systems are limited in scalability, as they require equipping individuals with specific devices, whereas vision-based systems can monitor large populations simultaneously with minimal additional resources. Another significant advantage is the versatility of vision-based systems, which can perform multiple computer vision tasks beyond gait recognition, such as face identification and behavior analysis, using the same visual data. This multifunctionality adds value, especially in environments like law enforcement and public safety, where different forms of surveillance are needed. Moreover, the rich spatial and temporal data captured by vision-based systems, particularly those with depth-sensing capabilities, allows for detailed analysis of gait dynamics, improving the accuracy and reliability of these systems in real-world conditions. Recent advancements in deep learning and computer vision have further improved the performance of vision-based gait recognition, enabling automatic feature extraction and robust handling of variations in lighting, clothing, and walking surfaces. However, some limitations remain, such as sensitivity to environmental factors like lighting conditions and potential privacy concerns due to the detailed visual data collected. Despite these challenges, vision-based systems offer scalable, non-intrusive, and cost-effective solutions for biometric identification in various fields, from security to healthcare (Parashar et al., 2024).

Feature representation

To the best of our knowledge, only few recent researches achieved end-to-end gait recognition (Song et al., 2019; Li et al., 2020b; Liang et al., 2022; Mehmood et al., 2020; Delgado-Escano et al., 2018; Li et al., 2021; Luo et al., 2019). Most of methods follow step-by-step manner, have not achieved true end-to-end gait recognition from video alone. Direct extraction of gait features from a walking video remains a challenge task without intermediate representation. An effective way is needed to encode both bodily structure and motion sequence from visual walking data. An intermediate body model serves as a routing to encapsulate essential body and movement information, facilitating subsequent analysis and recognition.

Both silhouette (Collins, Gross & Shi, 2002; Lam, Cheung & Liu, 2011; Wang et al., 2003) and skeleton (Rashmi & Guddeti, 2022; Teepe et al., 2022; Choi et al., 2019) can be use as representation. Actually, most existing gait recognition based on silhouette or skeleton. Silhouette-based extraction offers computational efficiency but remains somewhat vulnerable to variabilities in environmental and subject-related factors such as occlusion. By contrast, skeleton-based modelling demonstrates greater robustness when faced variants. But this approach relies heavily on the accuracy of the extracted skeletal joint configuration. To represent temporal information, both silhouette and skeleton animations require constructing a sequence of frames rather than just a single frame. A sequence of silhouette or skeleton can effectively represent one’s gait based on gait cycle theory illustrated in section Gait Cycle5.

Cai et al. (2021) proposed an integrated human body representation method combining both skeleton and silhouette information. By leveraging state-of-the-art pose estimation techniques, robust human skeletons can be extracted from monocular RGB images. A fusion strategy is then applied to compactly encode both skeleton and silhouette cues into a single silhouette-skeleton gait descriptor. Preliminary results demonstrate the efficacy of this hybrid representation for accurate and efficient human body modeling from visual data alone.

Dimension reduction

Dimension reduction (Ma & Zhu, 2013; Carreira-Perpinán, 1997) plays a pivotal role in the gait recognition process, as it addresses the challenge of dealing with high-dimensional data while preserving the most relevant information for identification (Santos et al., 2022). The extracted features from the gait data, such as silhouettes, temporal sequences, or kinematic parameters, often result in a feature space that is too large to be processed efficiently. Moreover, not all features contribute equally to the recognition task; some may even introduce noise or redundancy. Dimension reduction techniques aim to mitigate these issues by transforming the original high-dimensional feature space into a lower-dimensional one, enhancing the performance and computational efficiency of the gait recognition system.

This subsection provides an overview of the significance and types of dimension reduction methods commonly employed in gait recognition research. We categorize these methods into two main types: feature reduction and outliers removal. Feature reduction techniques focus on selecting or transforming the most relevant features, while outliers removal aims to eliminate noisy or abnormal data points that may negatively impact the recognition process (Marsico & Mecca, 2019).

Feature reduction techniques encompass various algorithms, such as principal component analysis (PCA) (Kherif & Latypova, 2020), linear discriminant analysis (LDA) (Balakrishnama & Ganapathiraju, 1998), tensor-based methods, and wavelet transforms. These methods help in capturing the essential characteristics of the gait data, while reducing the dimensionality to a more manageable level. For instance, PCA (Kherif & Latypova, 2020) and LDA (Balakrishnama & Ganapathiraju, 1998) are widely used to identify the principal components or discriminative features that best represent the variance or class separation in the dataset. Tensor-based methods like general tensor discriminant analysis (GTDA) (Khan, Farid & Grzegorzek, 2023) and sparse bilinear discriminant analysis (SBDA) (Lai et al., 2014) are employed to exploit the multi-linear structure of gait data, preserving the most informative features. Additionally, wavelet transforms, such as discrete wavelet transform (DWT) (Zhang, 2019), are utilized for decomposing the gait signals into different frequency bands, filtering out noise, and reducing dimensionality.

By effectively reducing the dimensionality of the gait data, these methods contribute to improving the accuracy, speed, and robustness of gait recognition systems. Furthermore, they facilitate the visualization of data and enable more efficient storage and transmission of gait templates.

Deep learning-based classification approaches

In the gait recognition pipeline, classification constitutes a pivotal process. Traditional machine learning (Mahesh, 2020; Alpaydin, 2021; Jordan & Mitchell, 2015; El Naqa & Murphy, 2015; Zheng et al., 2024b; Li et al., 2023a) classifiers, including support vector machines (SVM) (Xue et al., 2010; Kamruzzaman & Begg, 2006), hidden Markov models (HMM) (Panahandeh et al., 2013; Nickel et al., 2011), K-nearest neighbors (K-NN) (Choi et al., 2014; Bajwa, Garg & Saurabh, 2016), Bayesian (Chen, Liang & Zhu, 2011; Gupta et al., 2015), and random forest (Kececi et al., 2020; Shi et al., 2020), have played a fundamental role in this domain, have been widely discussed in Rani & Kumar (2023). Although these traditional machine learning classifiers have achieved some success in gait recognition, deep learning classifiers have garnered increasing attention in recent years, owing to their superior capabilities in handling large-scale datasets and complex feature extraction (Parashar et al., 2024). We mainly focus on deep learning-based classifier in this subsection.

Performance evaluation metric

In evaluating gait recognition methods, a critical initial step is partitioning the dataset into a gallery set and a probe set. The gallery set, containing pre-enrolled gait sequences for each individual, serves as a reference for identifying individuals, while the probe set introduces new sequences to test recognition accuracy. This setup enables systematic performance assessment across varied occlusion scenarios, simulating real-world identification challenges. Evaluation protocols such as closed-set, open-set, and cross-dataset testing further enhance assessment rigor. In the closed-set protocol, all individuals in the probe set are present in the gallery, allowing for controlled testing of identification accuracy within known subjects. Conversely, the open-set protocol challenges the model with individuals in the probe set who are not in the gallery, simulating real-world applications where subjects may vary. Cross-dataset evaluation adds an additional layer of complexity, testing models on entirely different datasets to assess generalization across new environments, camera setups, or occlusion types. Complementing these partitioning strategies, several key metrics serve as indicators of performance under occlusion, notably Rank-N accuracy, false acceptance rate (FAR), and equal error rate (EER). Rank-N accuracy gauges the model’s ability to identify the correct individual within the top-N results, an essential feature in occluded scenarios where visual data may be partially obstructed. FAR and EER provide insights into robustness under occlusion, measuring incorrect acceptances and the balance between acceptance and rejection errors, respectively. A low FAR, in particular, highlights the system’s capability to avoid mistaking occluded patterns for unauthorized individuals. Together, these protocols and metrics offer a comprehensive framework for assessing a model’s robustness, accuracy, and adaptability to occlusion in gait recognition, making them indispensable for understanding a method’s suitability for real-world applications.

Application scenarios

Gait recognition, with its non-intrusive identification approach (Devarapalli, Dhanikonda & Gunturi, 2020), offers promise in various applications. Unlike traditional biometrics requiring active cooperation, gait recognition identifies individuals through their walking patterns, often without their knowledge. As conventional biometrics face increased forgery risks, gait recognition’s resistance to imitation makes it a more secure option. In this section, we outline potential gait recognition applications.

Forensic gait recognition technology, as demonstrated in recent research, finds crucial applications in crime investigation (Marra, Di Silvestro & Somma, 2023) and forensics (Xu et al., 2023a). When video footage records a suspect’s walking pattern, algorithms analyze it to extract unique biometric data, facilitating identity verification through database matching. Gait’s inherent resistance to disguise or forgery enhances police tracking of suspect movements over time.

Access security control gait recognition technology enhances security access control systems (Ali, Sabir & Ullah, 2019) by verifying individual identities based on unique walking patterns. Unlike traditional methods susceptible to forgery, such as ID cards, passwords, or recent biometrics like face and fingerprint recognition, gait authentication is exceptionally hard to fake or mimic. This robust approach bolsters security and deters unauthorized access to sensitive areas and buildings.

Border security control. The gait recognition authentication system, when used at border checkpoints (Ngo & Hung, 2019), enhances security by accurately identifying travelers. It collects and analyzes gait data, extracting distinctive features and cross-referencing them with an identity database for swift and precise identification. This non-intrusive approach eliminates the need for additional documents or passwords, reducing the risk of errors.

High traffic scenarios. In highly populated areas such as transportation hubs, shopping districts, and streets, surveillance systems play a critical role in urban infrastructure. Within surveillance videos, “pedestrians” are a common focal point. Manual visual identification is labor-intensive, time-consuming, and prone to both visual fatigue (Souchet et al., 2022) and misidentification (Jung et al., 2020) over extended periods. Therefore, the adoption of gait recognition authentication technology is essential for efficient and accurate surveillance in such scenarios.

In such situational contexts, the phenomenon of occlusion is quite a regular occurrence. Various real-world scenarios involve complex spatial arrangements where one aspect of the visual scene is partially or fully blocked from sight by intervening obstructions. Figure 2 demonstrates two most common types of real-world occlusion.

Public datasets with native occlusion

The study of gait recognition under occlusion conditions has been greatly facilitated by the release of various datasets. Table 2 provides a detailed overview of 10 popular datasets that incorporate occlusion scenarios, each offering a distinct set of challenges and environments. For instance, OccGait (Huang et al., 2025) and SUSTech1K (Shen et al., 2023) capture data under indoor and outdoor conditions, respectively, with different types of occlusions such as static and crowd occlusions for OccGait and more diverse challenges including clothing and illumination for SUSTech1K (Shen et al., 2023). Datasets such as OG RGB+D (Li & Zhao, 2023) further leverage multi-modal data to enhance recognition under occluded conditions. The figure setups in Fig. 7 highlight the experimental environments for these datasets. For example, SUSTech1K (Shen et al., 2023) uses a LiDAR sensor and an RGB camera setup, while the OG RGB+D (Li & Zhao, 2023) dataset relies on a multi-Kinect SDAS system. Similarly, Depth Gait captures depth and skeleton data using Kinect sensors. VersatileGait (Zhang et al., 2023), another significant dataset, simulates occlusions in a game-engine environment, creating realistic walking animations. These datasets collectively provide a comprehensive platform for testing gait recognition algorithms under varying occlusion conditions.

OccGait. The OccGait benchmark is a pioneering occluded gait recognition dataset designed to address the challenges posed by real-world occlusions in gait recognition. It includes 101 subjects, with over 80,000 sequences captured from eight camera views. OccGait introduces four types of occlusion scenarios—None (normal walking), Carrying (with items like umbrellas or luggage), Crowd (interactions with other individuals), and Static (obstructions like plants or chairs). The dataset provides detailed annotations of occlusion types, offering a robust platform for the analysis and evaluation of gait recognition algorithms under varying occluded conditions. Data collection took place in an indoor lab, and silhouettes were extracted using MaskFormer for preprocessing. By simulating realistic occlusions, OccGait facilitates both qualitative and quantitative assessments, serving as a foundation for developing more resilient gait recognition systems. Extensive experiments on this benchmark demonstrate significant improvements in handling occlusion compared to previous methods (Huang et al., 2025).

SUSTech1K. The SUSTech1K dataset comprises a total of 1,050 subjects, encompassing 25,239 gait sequences. Each subject is captured from 12 different views in outdoor environment, collected by a LiDAR sensor and an RGB camera. Occlusion is considered in process of data collection. SUSTech1K encompasses the diverse range of variations observed within established datasets, incorporating factors such as Normal, Bag, Clothes Changing, Views, and Object Carrying. Furthermore, it acknowledges the presence of other prevalent and demanding variations encountered in outdoor scenarios, including Occlusion, Illumination, Uniform, and Umbrella. By systematically categorizing these walking sequences into distinct subsets based on their specific variations, we can gain a comprehensive understanding of how different variations impact the performance of gait recognition for both modalities (Shen et al., 2023).

CCPG. The CCPG dataset consists of 200 subjects and over 16,000 sequences captured both indoors and outdoors. Each subject appears in seven distinct outfits, making it one of the few datasets specifically designed to handle variations in clothing, including changes to tops, pants, entire outfits, and the presence of bags. Unlike previous datasets that focus on controlled conditions, CCPG introduces real-world challenges such as varying illumination, occlusion, and viewing angles. Both RGB and silhouette data are available, enabling a comprehensive evaluation of person re-identification and gait recognition methods under changing clothing conditions. This dataset provides a valuable resource for researchers aiming to improve the robustness of gait recognition systems in practical applications (Li et al., 2023b).

OG RGB+D. The OG RGB+D dataset consists of a total 96 subjects, encompassing 58,752 gait sequences. Each subject is captured from 12 different views in indoor environment. By utilizing a multiple synchronous Azure Kinect DK sensors data acquisition system (multi-Kinect SDAS), this database overcomes the limitations of existing gait databases by including multimodal gait data with various types of occlusions, such as self-occlusion, active occlusion, and passive occlusion. The Azure Kinect DK sensors allow for the simultaneous collection of multimodal data, supporting different gait recognition algorithms and enabling the effective capture of camera-centric multi-person 3D poses. The use of multiple views provided by the multi-Kinect setup is advantageous in handling occlusion scenarios compared to single-view systems. Moreover, the OG RGB+D database offers accurate silhouettes and optimized human 3D joints data (OJ) by combining data collected from the multi-Kinects, which ensures improved accuracy in representing human poses under occlusion (Li & Zhao, 2023).

VersatileGait. The VersatileGait dataset consist of a total 10,000 subjects, encompassing more than one million gait sequences. Each subject is captured from 44 different views in wild simulation environment by toolkit Unity3D. It offers a synthetic yet realistic representation of gait variations encountered in real-world scenarios. By incorporating diverse pedestrian attributes, walking styles, and a complex camera layout, the dat aset aims to bridge the domain gap between synthetic and real data (Zhang et al., 2023).

Gait3D. The Gait3D dataset is an extensive collection of gait sequences, comprising a total of 4,000 subjects and 25,309 gait sequences. This dataset is particularly noteworthy due to its large sample size, which allows for a more comprehensive analysis of gait characteristics and variations among individuals. Furthermore, the Gait3D dataset encompasses a diverse range of viewing angles, with a total view number of 39. This multifaceted perspective facilitates the examination of gait patterns from different angles, enabling a more thorough understanding of gait dynamics and kinematics. The Gait3D dataset is an invaluable resource for researchers in the field of gait analysis, providing a rich source of data for developing and validating gait recognition algorithms, as well as for studying the underlying mechanisms of gait disorders (Zheng et al., 2022).

GREW. The GREW dataset consists of a total of 26,345 subjects, encompassing 128,671 gait sequences. Each subject is captured from 882 different perspectives. Unlike other outdoor datasets, GREW is uncontrolled, which exhibits greater irregularity and complexity compared to controlled outdoor settings, representing what is commonly referred to as a wild environment. This research endeavors to shift the testing environment to real-life conditions. Existing benchmarks for gait recognition inadequately address the challenges encountered in fully unconstrained real-world scenarios. This underscores the necessity for more comprehensive and practical evaluation standards. GREW represents the pioneering effort in focusing on the wild environment, with the potential to serve as a significant milestone in the advancement of gait recognition (Zhu et al., 2021).

Multi Gait. The Multi Gait dataset is specifically designed to address the challenges associated with gait recognition in diverse conditions. With a total of 20 subjects, each captured from three different views, this dataset provides a comprehensive collection of 240 gait sequences. These sequences include variations in walking directions, such as back and forth, as well as the presence of dynamic occlusions, which are crucial for evaluating the robustness of gait recognition algorithms. Furthermore, the dataset also includes sequences without any occlusion, serving as a baseline for comparison. By incorporating such a wide range of conditions, the Multi Gait dataset offers valuable resources for researchers to develop and test gait recognition algorithms that can perform effectively in real-world scenarios. This dataset is an essential contribution to the field of computer vision and biometrics, and it will undoubtedly facilitate advancements in gait analysis and recognition technologies (Singh et al., 2019).

Depth Gait. The Depth Gait dataset comprises a total of 29 subjects, encompassing 464 gait sequences. Each subject is captured from both the back and the front views using two Kinect cameras mounted at the entry and exit points of a surveillance zone. Kinect, introduced by Microsoft in 2010, encompasses a series of motion sensing input devices. These devices are equipped with RGB cameras, infrared projectors, and detectors that facilitate depth mapping using structured light or time of flight calculations. The utilization of the dataset requires the involvement of Microsoft SDK due to the use of Kinect for data acquisition. The Depth Gait dataset specifically focuses on occlusion scenarios that occur in congested places such as airports, railway stations, and shopping malls, where multiple people enter the surveillance zone one after another, leading to the occlusion of the target (Chattopadhyay, Sural & Mukherjee, 2015).

TUM-IITKGP. The TUM-IITKGP dataset comprises a total of 35 subjects, encompassing 840 gait sequences. Each subject is captured from only one single view in indoor environment. This datset was created to simulate real-world surveillance scenarios. Specifically, this dataset addresses the problem of dynamic and static inter object occlusion. The physical setup involves a camera positioned at a height of 1.85 m in a narrow hallway, perpendicular to the direction of people walking from right to left and left to right in the image, reflecting a potential real-world surveillance application (Hofmann, Sural & Rigoll, 2011).

While current datasets have made significant strides in advancing gait recognition under occluded conditions, there are several areas for improvement that could further elevate their utility. Future datasets should prioritize more realistic and diverse occlusion scenarios, integrate additional data modalities such as thermal or radar imaging, and expand environmental coverage to include a wider range of real-world conditions. Increasing subject diversity and including longitudinal data will also enhance the generalizability of models. Furthermore, better annotation and metadata, along with synthetic data augmentation, will allow for the creation of richer datasets that can comprehensively address the challenges posed by occlusion in gait recognition. These improvements will ultimately drive the development of more robust and reliable gait recognition systems.

Occlusion synthetic methods based on public gait dataset

Introducing occlusion into normal gait datasets is a relatively straightforward process. It is feasible to generate a occluded version dataset derived from a publicly available gait dataset that lacks occlusion. In the literature, many different kind of methods to generate occlusion pattern exist. In this section, we present a collection of synthetic methods employed by researchers to contaminate normal gait video or it’s representation, thereby facilitating their investigations into occluded gait recognition. These methods serve as valuable tools for simulating occlusion scenarios and evaluating recognition algorithms under controlled conditions, with a summary provided in Table 3.

Reconstruction free

Reconstruction-free approaches for gait recognition are designed to bypass the conventional step of reconstructing occluded or corrupted input data, opting instead to directly extract features from the available, unoccluded portions of the subject’s gait. By skipping the reconstruction process, these methods reduce computational overhead, making them advantageous for real-time applications where processing speed and low latency are critical. This capability is particularly valuable in scenarios with partial occlusions, where the visible gait patterns offer sufficient information to achieve accurate identification. These approaches are built on specialized algorithms capable of emphasizing unaffected portions of the data, making them highly suitable for dynamic environments, such as crowded surveillance areas or live-streamed footage.

One standout example of a reconstruction-free strategy is GaitMoE (Huang et al., 2025), which uses a mixture of experts (MoE) model to adaptively handle different occlusion scenarios. In this framework, distinct expert models specialize in recognizing specific occlusion patterns, and the system dynamically selects the most suitable expert for each input. This adaptive approach enables GaitMoE to perform robustly across a range of environments with varying occlusion levels. However, this method involves a trade-off: while it increases flexibility and robustness, the MoE architecture adds complexity and computational demands due to the dynamic expert selection mechanism. This challenge becomes especially apparent in applications requiring continuous, real-time processing, where the multi-expert setup may affect overall efficiency. GMSN (Wei et al., 2024b) adopts a different approach by focusing on static occlusions through multi-scale feature extraction and part-based horizontal mapping. In this method, the silhouette of the subject is divided into segments, with emphasis placed on unoccluded body regions to optimize feature extraction. By concentrating on visible parts of the body, GMSN effectively mitigates the impact of occlusions, enhancing its resilience in scenarios with consistent occlusion patterns. Nevertheless, the reliance on multi-scale features introduces additional computational costs, particularly in environments where diverse occlusion patterns necessitate repeated adjustments. The segmentation and multi-scale strategy allow GMSN to perform well in moderately occluded environments, though handling fine-tuned adjustments across large datasets may require substantial computational resources. Further expanding on reconstruction-free methods, GaitASMS (Sun et al., 2024) incorporates adaptive spatial representation with multi-scale temporal aggregation to improve occluded gait recognition. This approach enables the model to emphasize visible portions of the gait while maintaining adaptability to changing occlusion patterns. GaitASMS’s temporal aggregation methods are particularly effective in dealing with partial occlusions across extended gait sequences; however, the complexity of aggregating temporal features at multiple scales can lead to increased processing time, especially when managing long gait sequences or high-resolution data. The GLGait method (Peng et al., 2024) offers an innovative take on reconstruction-free approaches by combining global and local temporal receptive fields. This fusion allows GLGait to capture both macro-level gait trends and micro-level body movements, thus providing robustness across varying degrees of occlusion. By integrating global and local cues, GLGait excels in recognizing gait patterns in real-world scenarios where occlusion levels fluctuate. Nonetheless, the complex temporal attention mechanisms utilized in this approach can result in higher computational requirements, making it less ideal for applications with constrained resources or where large-scale data must be processed rapidly.

While these reconstruction-free approaches demonstrate effectiveness in moderate occlusion scenarios, their performance tends to decline as occlusion severity increases. In cases where substantial portions of the body are obscured, the ability of these methods to extract discriminative features diminishes due to the lack of visible gait information. This limitation highlights a core difference from reconstruction-based methods, which are designed to infer missing data and perform better in extreme occlusion settings. The primary advantage of reconstruction-free approaches lies in their efficiency, as they directly process the available data without the need for intermediate reconstruction. This streamlined operation is particularly suitable for environments where processing speed is essential, such as in surveillance systems deployed in crowded public spaces, transportation hubs, or live video feeds, where rapid response is prioritized over absolute precision. These methods, by focusing on extracting features from visible, unoccluded regions, allow for a significant reduction in computational load compared to reconstruction-based techniques. The use of models like GaitMoE (Huang et al., 2025), GMSN (Wei et al., 2024b), and GLGait (Peng et al., 2024) demonstrates how specialized, feature-focused extraction can provide robust performance with partial occlusions. However, the reliance on visible data limits the use of reconstruction-free methods in high-occlusion scenarios, where critical gait features may be obscured. The multi-expert setups in models like GaitMoE and the complex temporal mechanisms in GLGait, while enhancing flexibility, also contribute to computational overhead that may constrain their real-time capabilities in highly dynamic environments. As such, reconstruction-free methods are best suited for applications with low-to-moderate occlusions and environments prioritizing speed over exhaustive detail. For more severe occlusions, reconstruction-based approaches may still offer superior recognition accuracy by filling in missing gait information, thereby preserving discriminative features necessary for reliable identification.

Reconstruction based

Reconstruction-based gait recognition methods offer a robust solution for handling occlusion by first reconstructing missing or occluded portions of gait sequences before proceeding to feature extraction. These methods commonly leverage deep learning architectures, such as autoencoders (Tschannen, Bachem & Lucic, 2018; Yang et al., 2022; Li, Pei & Li, 2023) and generative adversarial networks (GANs) (Goodfellow et al., 2020; Saxena & Cao, 2021), to restore incomplete representations of gait data. By providing a more holistic view of the occluded data, reconstruction-based approaches enable the extraction of rich and informative features, significantly enhancing identification accuracy, even in challenging scenarios with substantial data loss. These methods are particularly valuable in high-security environments and critical applications where maximizing recognition accuracy is essential, albeit at the cost of higher computational demands.

One notable example is GaitAE (Li et al., 2024a), which utilizes an autoencoder model to map corrupted input gait sequences into a latent space and subsequently reconstructs them into a more complete form. The reconstruction process in GaitAE allows for the recovery of essential gait features, making it highly effective in scenarios with extensive occlusions. However, GaitAE’s dependence on synthetic occlusion data during training introduces potential challenges in generalizability, as real-world occlusions may differ from synthetic patterns, potentially limiting the model’s performance in authentic, unpredictable occlusion conditions. Adaptations, such as transfer learning, could enhance GaitAE’s applicability in real-world environments by enabling it to generalize better to diverse occlusion patterns encountered outside the training data. BGaitR-Net (Kumar et al., 2024), another influential reconstruction-based model, targets static occlusions by employing a deep learning framework to restore occluded gait segments before feature extraction. This model is particularly effective under high levels of occlusion, as it reconstructs missing portions of the gait data, allowing for robust recognition despite significant visual data loss. Nevertheless, BGaitR-Net’s computationally intensive reconstruction process can be a limiting factor in real-time applications, especially in high-stakes settings where rapid decision-making is required. SSGait (Xi et al., 2024) introduces a more sophisticated approach by combining semi-supervised and self-supervised learning techniques to enhance recognition performance in dynamic occlusion scenarios. Utilizing methods such as tube-masked reconstruction and clip order prediction, SSGait effectively handles occlusions without the need for extensive labeled data, improving data efficiency. However, the complexity of these self-supervised techniques increases computational demands, making SSGait less feasible for real-time deployment in environments that require quick processing and response times. OccSilGait (Peng, Cao & He, 2023) provides an advanced reconstruction-based framework specifically designed to address both static and dynamic occlusions. By using multiple modules to generate realistic occluded gait sequences, OccSilGait achieves high recognition accuracy even in severe occlusion scenarios. The multi-module architecture, however, introduces substantial computational costs during both training and inference, limiting its practicality in resource-constrained or real-time applications. Additional models, such as GaitSTR (Zheng et al., 2024a) and QAGait (Wang et al., 2024), also adopt reconstruction-based approaches to tackle occlusion. GaitSTR is notable for its adaptability across various occlusion types, using complex reconstruction techniques to address a wide range of missing data scenarios, while QAGait incorporates a quality assessment stage to ensure that only high-quality reconstructed features are used in the identification process. This added quality assessment step enhances QAGait’s performance in real-world environments where data fidelity is often compromised.

Although reconstruction-based methods offer an effective approach to managing severe occlusions, they are computationally intensive, which limits their applicability in real-time scenarios. The reconstruction process itself requires significant computational resources, and the added feature extraction step further contributes to the overall processing load. As such, these methods are generally more suited for high-security or high-stakes environments, such as forensic analysis or post-event security assessments, where recognition accuracy is prioritized over speed. With advancements in hardware acceleration and optimized neural network architectures, however, there is potential to bring reconstruction-based methods closer to real-time applicability. Improved computational efficiency could expand the utility of these methods to broader domains, enabling more responsive gait recognition systems capable of managing high levels of occlusion in dynamic, real-world environments. For applications requiring immediate responses, though, the trade-off between computational complexity and accuracy makes reconstruction-free methods a more practical option, albeit at the potential cost of reduced effectiveness under heavy occlusions.