A systematic review on hand gesture recognition techniques, challenges and applications

Gesture acquisition methods

Image hand gesture acquisition, as illustrated in Fig. 5 from Microsoft (2019) is to capture the human hand gesture image by the computer (Ananya, Anjan Kumar & Kandarpa Kumar, 2015). This could be done using vision-based recognition where no special gadgets are required and a web camera or a depth camera is used, furthermore special tools can be utilized such as wired or wireless gloves that detect the movements of the user hand, and motion sensing input devices (Kinect from Microsoft, Leap Motion, etc.) that capture the hand gestures and motions as showed in Fig. 5.

Several methods for hand gestures acquisition were discussed in the work studied using different acquisition tools, part of them used images that were captured in real time and others used previous images that were recorded in databases. Another classification to be considered is the nature of the acquisition which can be either a static hand gesture recognition where the gestures are represented by a constant image, or a dynamic hand gesture recognition were gestures are represented by the movement of the hand.

The first method is using vision-based hand gesture recognition to extract images which was proposed by Weiguo et al. (2017), Ananyaa et al. (2017), Alvi, Fatema & Mohammad (2016), Shome (2017). Where in Weiguo et al. (2017) it was built in a real-time system.

In Oinam et al. (2017) the authors implemented two different techniques of vision-based hand gesture recognition and one data glove-based technique. The vision-based techniques are static hand and real-time hand gesture recognition techniques. In data glove-based technique the glove had five flex sensors. Results showed that the vision-based technique was more stable and reliable compared to the data glove-based technique. Hand gestures are obtained by evaluating the contour captured from the image segmentation using a glove worn by the speaker in Rosalina et al. (2017). Also, in Danling, Yuanlong & Huaping (2016) they used a novel data glove called YoBu to collect data for gesture recognition.

The work in Gunawardane & Nimali (2017) compared using a data glove to track the motion of the human hand using flex sensors, gyroscopes and vision data with Leap Motion Controller. Results showed that the Leap Motion Controller had a high repeatability and high potential for soft finger type applications. Furthermore, in Eko, Surya & Rafiidha (2017), Shaun et al. (2017) and Deepali & Milind (2016) the gestures are also captured using Leap Motion Controller.

In Siji Rani, Dhrisya & Ahalyadas (2017) they used a new Hand Gesture Control in Augmented Reality System (HGCARS) where the gesture recognition is performed using a secondary camera and the reality is captured using an IP camera, the virtual object is added to the video feed obtained from an IP camera and controlled by using the position and depth of hand, measured using a webcam. Moreover, the authors of Salunke & Bharkad (2017), Rokhsana et al. (2017), Jessie et al. (2016) and Anshal, Heidy & Emmanuel (2017) used webcams for gathering input data.

In Aditya et al. (2017), a hand gesture recognition on top-view hand images observed by a Time of Flight (ToF) camera in a car for touchless interactions inside a car was proposed. The help of image capturing devices installed on computer was implemented in Nilima & Patil (2017). The authors of Hafız et al. (2017) presented a CMSWVHG (Control MS Windows via hand Gesture) to perform numerous windows actions using hand gestures using internal or external camera for taking input with the help of OpenCV. To control OS on the projected screen for a virtual mouse system without any hardware requirement one camera source was required (Rishabh et al., 2016). Also, data acquisition was done using Camera interfacing in Ashish & Aarti (2016a) and Ashish & Aarti (2016b).

To detect static hand gesture RGB cameras and depth data were used by Rytis & Arunas (2017) and Chenyang, Xin & Lianwen (2017), whereas in Chenyang, Xin & Lianwen (2017) they tested their approach on Sheffield Kinect Gesture (SKIG). A static hand gesture recognition using Kinect depth sensor to capture color, and infrared and depth frames was applied by Fuchang & Hijian (2016), Xiang, Lei & Qiang (2017), Jozef & Slavomír (2017), Sih-Huei et al. (2017) and Yiwen et al. (2017). Moreover, the authors proposed a Kinect based real time dynamic hand gesture recognition technique in Atharva & Apurv (2017), Devendrakumar & Kakade (2016) and Karthik et al. (2017).

A real time wearable hand gesture recognition system, which decoded the information from surface electromyography (sEMG) was proposed in Jian et al. (2017), Jinxing, Jianhong & Jun (2017), David & Donghan (2018), Kuang-Yow et al. (2017), Jakub, Tomasz & Piotr (2017), Shunzhan et al. (2017) and Andrés & Marco (2017). Also, a real time system based on the EMG with the surface electromyography measured by the commercial sensor Myo which is an armband placed on the forearm was proposed by Marco et al. (2017a), Marco et al. (2017b), Karthik et al. (2017) and Stefano et al. (2018).

In Yuhui, Shuo & Peter (2018), Yuta et al. (2017), Donq-Liang & Wei-Shiuan (2018), Ananta & Piyanuch (2016) and Nabeel, Rosa & Chan (2017) a sensor-based wearable wristband was presented for static hand gestures. The authors in Hari et al. (2016) and Erhan, Hakan & Baran (2017) presented making use of the sensors in smartphones. The authors of Yifan et al. (2018) used wearable devices such as VR/AR helmet and glasses in a gesture recognition system. A Doppler radar sensor with dual receiving channels was used to acquire a big database of hand gestures signals in Jiajun & Zhiguo (2017). A gesture recognition system for human–computer interaction based on 24 GHz radars in Shengchang et al. (2017) and 2.4-GHz continuous radar in Takuya et al. (2017) were used.

Moreover, an interface based on a mechanomyography (MMG) signal to capture arm movement and hand gesture was applied in Yuntao et al. (2017) and Piotr, Tomasz & Jakub (2017).

The work of Ibrahim et al. (2018) proposed a hand gesture system using two monopole antennas to measure signatures at different locations, and the work of Zengshan et al. (2018) developed a wireless sensing WiCatch, where the motion locus of the gesture is developed by constructing virtual antenna array based on signal samples.

Feature extraction

Hand tracking process is the ability of the computer to trace the user hand and separate it from the background and from the surrounding objects (Ananya, Anjan Kumar & Kandarpa Kumar, 2015). This can be done using multi-scale color feature hierarchies that gives the users hand and the background different shades of colors to be able to identify and remove the background, or by using clustering algorithms that are capable of treating each finger as a cluster and removing the empty spaces in-between them. In this subsection, we will be discussing the different features used by the different applications implemented in the reviewed work. Furthermore, the feature extraction methods used will also be discussed.

Classification of hand gestures

The features extracted are sent to training and testing the classification algorithm (such as Artificial Neural Networks (ANN), K-nearest neighbor (KNN), Naive Bayes (NB), Support Vector Machine (SVM), etc.) to reach the output gesture, for example the output gesture in a simple case can contain two classes to detect only two gestures such as open and closed hand gestures, or in sign language as shown in Fig. 2.

The classification of the information extracted from the images of hand gestures is required to be able to recognize these gestures by the computer, there are several classification algorithms that can be used for this purpose.

The work of Haitham, Alaa & Sabah (2017), Weiguo et al. (2017), Rosalina et al. (2017), Jakub, Tomasz & Piotr (2017), Piotr, Tomasz & Jakub (2017), Shunzhan et al. (2017), Erhan, Hakan & Baran (2017), Vladislava, Predrag & Goran (2016), Deepali & Milind (2016) and Ananta & Piyanuch (2016) proposed using Artificial Neural Networks (ANN) for classification, whereas in Jakub, Tomasz & Piotr (2017) and Piotr, Tomasz & Jakub (2017) SoftMax output layer was used with feedforward ANN; in Shunzhan et al. (2017), Erhan, Hakan & Baran (2017), Alvi, Fatema & Mohammad (2016), Vladislava, Predrag & Goran (2016); Deepali & Milind (2016) and Ananta & Piyanuch (2016) backpropagation training methods were used, and the authors in Jessie et al. (2016) used Kohonen self-organizing maps as a type of ANN to classify data sets in unsupervised manner to convert hand gestures into Filipino words.

In Chun-Jen et al. (2017), a synthetically-trained neural network was used for 3D hand gesture identification. The training process of a deep-learning neural network required a large amount of training data. The authors of Chenyang, Xin & Lianwen (2017) proposed using the LPSNet, an end-to-end deep neural network for hand gesture recognition with novel log path signature features. In Yifan et al. (2018) and Biyao et al. (2017) they proposed using deep neural networks for classification.

The authors in Sungho & Wonyong (2016) came up with two dynamic hand gesture recognition techniques using low complexity recurrent neural network (RNN) algorithms for wearable devices, the first was based on video signal and uses convolutional neural network (CNN) with RNN for classification, and the other used accelerometer data and applied RNN for classification. In contrast, Xinghao et al. (2017) used a bidirectional recurrent neural network (RNN) with the skeleton sequence to augment the motion features for RNN were used.

Deep convolutional neural network was used for classification in Ibrahim et al. (2018), David & Donghan (2018), Javier, Paula & Robinson (2017), Jose et al. (2017), Aditya et al. (2017), Yuntao et al. (2017), Peijun et al. (2017), Jozef & Slavomír (2017), Jiajun & Zhiguo (2017), Takuya et al. (2017) and Fabian et al. (2017), where in Aditya et al. (2017) the authors aimed to improve the detection of hand-gestures by correcting the probability estimate of a Long-Short-Term Memory (LSTM) network by pose prediction performed by a Convolutional Neural Network (CNN). They used Principal Component Analysis (PCA) as a training procedure to reduce the labelled data of hand pose classification to perfect the initialization of weights for the CNN.

The authors in Fuchang & Hijian (2016), Salunke & Bharkad (2017) and Marlon, Alistair & Mohamed (2016) used K-nearest neighbor (KNN) classifier to recognize hand gestures. The work presented in Kuang-Yow et al. (2017) classified the recognition using k-Nearest Neighbor (KNN) and Decision Tree algorithms combination.

Support vector machine (SVM) was used for classification in Yuhui, Shuo & Peter (2018), Chenyang, Yingli & Matt (2016), Yuta et al. (2017) Xi, Chen & Lihua (2017), Rasel et al. (2017), Reshna & Jayaraju (2017), Zengshan et al. (2018), Ashish & Aarti (2016a), Ashish & Aarti (2016b), Zhiwen et al. (2017) and Stefano et al. (2018). Decision tree was used to classify the gestures in Shengchang et al. (2017). To recognize the patterns in Jian et al. (2017), they used a modified deep forest algorithm. In Shaun et al. (2017) and Riya et al. (2017), they used a random regression forest algorithm for classification.

In Jinxing, Jianhong & Jun (2017) the hand gesture was modeled and decomposed by the use of Gaussian Mixture Model-Hidden Markov Models (GMMHMM), GMMs are used as sub-states of HMMs to decode sEMG feature of gesture. Moreover, a hand gestures recognition system was implemented using the incorporation of Bhattacharyya divergence into Bayesian sensing hidden Markov models (BS-HMM) in Sih-Huei et al. (2017). Also, in Devendrakumar & Kakade (2016) Hidden Markov Model was used for classification.

The work of Hari et al. (2016), Atharva & Apurv (2017), Yiwen et al. (2017) and Washef, Kunal & Soma (2016) presented using the dynamic time warping algorithm for classification. Whereas Marco et al. (2017a) and Marco et al. (2017b) used the k-nearest neighbor rule together with the dynamic time warping algorithm for classification.

Naive Bayes was applied as the training method for classification of Eko, Surya & Rafiidha (2017). Multiple linear discriminant analysis (LDA) classifier was adopted to classify different hand gestures in Isack, Zhongfu & Jamal (2017) and Yuefeng et al. (2016).

Classification of digits gesture from 11 to 20 was done using Principal component analysis by Rajeshree & Jayashree (2016). The work (Danling, Yuanlong & Huaping, 2016) used extreme learning machine (ELM) for gesture recognition.

The work presented in Himadri et al. (2018) used Support Vector Machine (SVM), Artificial Neural Network, Naive Bayes and K-Nearest Neighbor (K-NN) classifiers as the training methods to recognize the features extracted. In contrast, in Ananyaa et al. (2017) the best classification accuracy was achieved using Euclidean distance and Eigen vector, but this result is for a very small dataset, and the best result was a dataset containing nearly 720 images that used Support vector machine for classification of images; also, using Artificial Neural Network provided an accuracy of 89.48%. Furthermore, Multi-class Support Vector Machine (SVM) and k-Nearest Neighbors (KNN) classifiers were used to classify the hand gestures in Rania, Mohammed & Mahmoud (2017), experiments showed that the accuracy of KNN classifier were better than SVM classifier. In Nabeel, Rosa & Chan (2017) Support Vector Machine (SVM), Decision Tree (DT), K-Nearest Neighbors (KNN), and Linear Discriminant Analysis (LDA) were compared in classification performance, and results showed that LDA got the highest accuracy.

Sign language

Sign languages, as illustrated in Fig. 2, use manual communication to explain a certain meaning. This can include using hand gestures, movements, fingers orientation to deliver a certain word or meaning.

The work of Weiguo et al. (2017), Chenyang, Yingli & Matt (2016), Rasel et al. (2017), Shaun et al. (2017), Deepali & Milind (2016), Nabeel, Rosa & Chan (2017); and Anshal, Heidy & Emmanuel (2017) proposed using hand gesture recognition for American Sign Language (ASL) for people who are hearing impaired, whereas in Shaun et al. (2017) they did not evaluate the letters that are dynamic (like: j and z). In Nabeel, Rosa & Chan (2017) 36 gestures were studied including 26 ASL alphabets and 10 ASL numbers. In Anshal, Heidy & Emmanuel (2017) two different translation paradigms; English characters (alphabet) and complete words or phrases were proposed.

An alternative to talking for deaf & dumb people who have hearing or speech problems using sign language recognition and hand gestures was proposed in Himadri et al. (2018), Oinam et al. (2017), Reshna & Jayaraju (2017), Vivek, Ramesh & Kagalkar (2017), Oyndrila et al. (2016), Ashish & Aarti (2016a), Ashish & Aarti (2016b) and Rishabh et al. (2016). Whereas in Rosalina et al. (2017) a sign language that consists of the alphabetic from “A” to “Z”, the numbers from “0” to “9”, and some additional punctuation marks like “Period”, “Question Mark”, and “Space” using static hand gesture recognition was produced.

The work of Jose et al. (2017) used the alphabet of sign language of Peru (LSP). Also, Vimonhak, Aranya & Somsak (2017) proposed a technique for the recognition of Lao alphabet sign language for the deaf people. In Eko, Surya & Rafiidha (2017), Indonesian Sign Language System (SIBI) was applied for normal people to learn sign language and communicate to people with hearing disability. Moreover, in Donq-Liang & Wei-Shiuan (2018), the authors used 29 Turkish fingerspelling signs. A framework for recognizing Sinhala sign language gestures and translating them in to natural language for hearing & speaking impaired people was proposed in Vladislava, Predrag & Goran (2016). The authors of Jessie et al. (2016) converted hand gestures into Filipino words. The work presented in Marlon, Alistair & Mohamed (2016) used the alphabet of Irish Sign Language. Finally, Indian Sign Language (ISL) hand gesture recognition system was introduced in Riya et al. (2017).

Robotics

Hand gestures can also be employed in the process of building robotics, as illustrated in Fig. 6 (Elecbits, 2019) which shows the use of a glove to control a robotic vehicle. A recognition frame of continuous hand gestures of upper-limb exoskeleton robot was proposed in Jinxing, Jianhong & Jun (2017). In Yuhui, Shuo & Peter (2018) they had three groups of hand gestures: six wrist gestures, five single finger flexions, and ten Chinese number gestures that were built to control a robotic arm. The authors of Sungho & Wonyong (2016) came up with two dynamic hand gesture recognition techniques using low complexity recurrent neural network (RNN) algorithms for a wearable device.

The experiments of Xiang, Lei & Qiang (2017) produced a control system for robotic wheelchair for the aged and the disabled, which contained two parts: gesture interaction and intelligent wheelchair. Also, the approach of Javier, Paula & Robinson (2017) allowed the consequent execution of a robotic agent for the delivery of objects. In Yuntao et al. (2017) the authors introduced a wearable sensor suite fusing arm motion and hand gesture recognition for operator control of UAVs.

The authors of Yuta et al. (2017) proposed a wearable device with photo-reflective sensors arranged in an array to recognize hand gestures by measuring the skin deformation of the back of the hand. Also, in Danling, Yuanlong & Huaping (2016) hand gestures recognition was used in human–robot interaction (HRI). In experiments of Devendrakumar & Kakade (2016), continuous trajectory path made by hand over a period of time was considered for real time dynamic gesture recognition purpose using Kinect sensor for controlling robotic arm. A hand gesture-based control design was proposed for mobile robots in Hang et al. (2017), where mobile robots can move according to the signals encoded by hand gestures. Work of Vedha Viyas et al. (2017) dealt with a robotic arm using hand gestures for many applications such as automation, medical applications and gaming.

Others

A static hand gesture recognition in a real-time human–computer interaction application was proposed in Salunke & Bharkad (2017) to control the Power Point presentations from a distance. A human–computer interaction based on hand gesture in a projector-camera system was presented in Qingrui et al. (2017). Also, real-time hand gesture recognition for flight control of multiple quadrotors was applied in David & Donghan (2018). The authors of Javier, Paula & Robinson (2017) presented the design of a convolutional neural network architecture using the MatConvNet library in order to achieve the recognition of 2 classes: “open” and “closed” and “unknown”. The approach studied in Marco et al. (2017a) and Marco et al. (2017b) could be used in multiple applications in medical and engineering fields.

The study of Aditya et al. (2017) showed hand gesture recognition on top-view hand images observed by a Time of Flight (ToF) camera in a car for touchless interactions inside a car. In Peijun et al. (2017) they used seven kinds of hand gestures that can command a consumer electronics device, such as mobile phones and TVs. The authors in Siji Rani, Dhrisya & Ahalyadas (2017) proposed an Augmented Reality (AR) to merge the virtual world into the real world and to enhance the perception of the real world by adding 2-D virtual objects.

The experiments of Yu et al. (2017) described a method of hand gesture recognition using Principle Component Analysis (PCA) implemented in Android phone. A real-time hand gesture recognition system was proposed using electromyography (EMG) in the field of medicine and engineering, with a higher number of gestures to recognize in Marco et al. (2017a) and Marco et al. (2017b). The work of Rokhsana et al. (2017) presented a hand gesture-based computer mouse control system. The authors of Hafız et al. (2017) presented a CMSWVHG (Control MS Windows via hand Gesture) to perform numerous windows actions using hand gestures.

The work discussed in Anna, Sung & Jae-Gon (2018) presented a method to detect and to recognize hand gestures for generating hand gesture-based commands to control the media consumption in smart glasses. Hand gesture recognition or finger angle prediction for Ultrasound imaging was proposed in Yuefeng et al. (2016). A hand gesture recognition technique on a smartphone using Google Cardboard (GC) and Wearality2 in phones was applied in Srinidhi et al. (2016).

A real-time hand gesture recognition technique for presentation was proposed in Rishabh et al. (2016), to control OS on the projected screen for a virtual mouse system without any hardware requirement. Hand-gesture-based commands were proposed in Lalit & Pritee (2017) to replace touch and electromechanical input panels using vision-based mid-air character input.

Gesture recognition system was proposed in Soumya & Muzameel (2017), Mudra is an expressive form of gesture that is mainly used in Indian classical dance form where the gesture is in visual form to connect with the audience. The authors in Zhiwen et al. (2017) presented a real-time hand gesture recognition by using Kinect sensor, to control mouse by user hands for operations such as ‘clicking’, ‘dragging’ and ‘dropping’, and engaged/disengaged gestures. Gesture recognition in Karthik et al. (2017) was used for Bio Robotics, the paper focused on presenting a sensor based human gesture recognition for the Hand Cricket game.

Environmental surroundings

In Fuchang & Hijian (2016) the environmental background, light and rotation, translation and scale change proposed a difficulty; the authors used depth data to help in hand separation and to accomplish a synchronized color and depth images in Kinect. Different light brightness (dull, medium, and bright) conditions were tested in Salunke & Bharkad (2017) to overcome their problem, the results of the bright light level gave the best accuracy as expected. Furthermore, the results of Oinam et al. (2017) showed that the vision-based techniques gave an accuracy of 100% in bright lighting condition with a white background.

The work of Qingrui et al. (2017) showed that it was difficult to achieve proper accuracy when having complex backgrounds, variable external illumination, and shadows of hand gesture. The authors of Hari et al. (2016) presented a continuous hand gestures recognition technique using three-axis accelerometer and gyroscope sensors in smartphones, a gesture coding algorithm was also used to reduce the influence of unstableness of the user hand.

Digital image processing techniques were used in Jose et al. (2017) to eliminate noise, to improve the contrast under different illumination, to separate the hand from the background and to cut the region containing the hand, where results of images with complex background got low accuracy.

In Rytis & Arunas (2017) the authors suggested that RGB cameras can be used to detect hand, but it has limited applications because hand detection was hard in some lighting conditions or in different skin colors, using depth camera data was better in distinguishing hands in front of camera.

Study of Peijun et al. (2017) suggested that spatial localization of the hands when it contains background, human face, and human body could be a challenging task, where the results achieved an accuracy of 97.1% in the dataset with simple backgrounds and 85.3% in the dataset with complex backgrounds.

The authors of Yu et al. (2017) were capable of solving the problems that faced them such as different size of gesture image captured, different angle of gesture’s rotation and flipped gesture. After experiments of Atharva & Apurv (2017) they found that if there was a larger object than the hand that the Kinect captures, wrong results will be achieved. The results of Hafız et al. (2017) were highly affected by the background noise where they achieved an accuracy of 82.52%.

In experiments of Yifan et al. (2018) and to strengthen the recognition accuracy, and as the dataset used contained more than 24,000 gesture and 3,000,000 frames for both color and depth modalities from 50 subjects, they tested 83 different static and dynamic gestures with six diverse indoor and outdoor scenes respectively with variation in background and illumination, they also tested when people perform gestures while they are walking and they achieved an acceptable accuracy.

In Reshna & Jayaraju (2017) The authors used complex backgrounds with respect to the skin color, and results showed a disadvantage in this system which was the lighting conditions. The results of Rania, Mohammed & Mahmoud (2017) showed that the proposed algorithm achieved recognition rate of 97.69% under different hand poses and complex background with changes in lightning, moreover their system was robust to rotation, scale, translation, and lighting.

Also, the results of Yiwen et al. (2017) showed that their method was adjusting to translation, rotation scaling and articulated deformation. Time of Flight (ToF) data was used in Fabian et al. (2017) which tends to be overly noisy depending on various factors such as illumination, reflection coefficient and distance that were studied in their research process. The system of Washef, Kunal & Soma (2016) overcame the limitations of a glove-based approach and the vision-based approach concerning different illumination conditions, background complexity and distance from camera. The system in Lalit & Pritee (2017) was applicable for rotation, scale, and translation variations, directions, and the dataset contained digits, alphabets, and symbols.

Training and testing Data

The training and testing data problems can be classified as either having a small set of gestures that are insufficient for recognition training or having a large set of gestures that are imbalanced leading to a low recognition accuracy. The training data of Haitham, Alaa & Sabah (2017) had only 25 images and the results reached an accuracy of 82.2%. In Weiguo et al. (2017) they trained their system using 24 static American Sign Language (ASL) and their accuracy reached 93.3%, whereas in Fuchang & Hijian (2016) they used detailed features to avoid the problem of having imbalanced data and images.

To overcome the problem of having more than one meaning and move to represent a gesture the authors of Oinam et al. (2017) used vision-based hand gesture recognition systems and data glove-based hand gesture recognition systems. In Jian et al. (2017) 16 hand gestures include 10 dynamic and six static gestures were used, and results showed that gestures were correctly recognized and the accuracy was 96%.

The work in Ibrahim et al. (2018) tested the system using ten hand gestures achieving their goal which is to maximize the recognition accuracy using Pre-trained networks. The data of Javier, Paula & Robinson (2017) contained six architectures with variance in the hyperparameters and depth, their results achieved good accuracy. Classification of 24 hand gestures to improve the rate of recognition accuracy was proposed in Jose et al. (2017) reaching 96.20% for simple backgrounds.

In Aditya et al. (2017) five hand poses with nine class hand gestures were used reaching an accuracy of 89.50% as a result. The work of Vimonhak, Aranya & Somsak (2017) used 54 Lao alphabets in the experiments, testing data had 540 gestures from four individuals, it was difficult to maintain constant speed of hand gestures, thus the recognition rate was 79%.

The main difficulties of the hand gesture recognition (Andrés & Marco, 2017) with EMG using machine learning are: the noisy behavior of EMG signal, and the small number of gestures per person relative to the number of generated data by each gesture (overfitting). Whereas, in Chun-Jen et al. (2017) the training process of a deep-learning neural network required a large amount of training data, they combined a large set of computer-generated 3D hand images with few real camera images to form the training data set, the testing and training sets had 24 classes of hand gestures, and the results accuracy was 77.08%.

The authors of Deepali & Milind (2016) considered 26 different alphabets of American Sign Language, with a total of 520 samples (consisting of 20 samples of each alphabet), and results gave an accuracy of 96.15%. In study of Stefano et al. (2018) to avoid correlation between training and testing datasets the Leave One Subject Out (LOSO) cross-validation technique was used, and the accuracy was 92.87%.