Table 1.

Comparison of the most recent selected gesture recognition studies.

Author	Findings	Challenges
[56]	In this study, image processing techniques such as wavelets and empirical mode decomposition were suggested to extract picture functionalities in order to identify 2D or 3D manual motions. Classification of artificial neural networks (ANN), which was utilized for the training and classification of data in addition to the CNN (CNN).	Three-dimensional gesture disparities were measured utilizing the left and right 3D gesture videos.
[57]	Deaf–mute elderly folk use five distinct hand signals to seek a particular item, such as drink, food, toilet, assistance, and medication. Since older individuals cannot do anything independently, their requests were delivered to their smartphone.	Microsoft Kinect v2 sensor’s capability to extract hand movements in real time keeps this study in a restricted area.
[58]	The physical closeness of gestures and voices may be loosened slightly and utilized by individuals with unique abilities. It was always important to explore efficient human computer interaction (HCI) in developing new approaches and methodologies.	Many of the methods encounter difficulties like occlusions, changes in lighting, low resolution and a high frame rate.
[59]	A working prototype is created to perform gestures based on real-time interactions, comprising a wearable gesture detecting device with four pressure sensors and the appropriate computational framework.	The hardware design of the system has to be further simplified to make it more feasible. More research on the balance between system resilience and sensitivity is required.
[60]	This article offers a lightweight model based on the YOLO (You Look Only Once) v3 and the DarkNet-53 neural networks for gesture detection without further preprocessing, filtration of pictures and image improvement. Even in a complicated context the suggested model was very accurate, and even in low resolution image mode motions were effectively identified. Rate of high frame.	The primary challenge of this application for identification of gestures in real time is the classification and recognition of gestures. Hand recognition is a method used by several algorithms and ideas of diverse approaches for understanding the movement of a hand, such as picture and neural networks.
[61]	This work formulates the recognition of gestures as an irregular issue of sequence identification and aims to capture long-run spatial correlations in points of the cloud. In order to spread information from past to future while maintaining its spatial structure, a new and effective PointLSTM is suggested.	The underlying geometric structure and distance information for the object surfaces are accurately described in dot clouds as compared with RGB data, which offer additional indicators of gesture identification.
[62]	A new system is presented for a dynamic recognition of hand gestures utilizing various architectures to learn how to partition hands, local and global features and globalization and recognition features of the sequence.	To create an efficient system for recognition, hand segmentation, local representation of hand forms, global corporate configuration, and gesture sequence modeling need to be addressed.
[63]	This article detects and recognizes the gestures of the human hand using the method to classification for neural networks (CNN). This process flow includes hand area segmentation using mask image, finger segmentation, segmented finger image normalization and CNN classification finger identification.	SVM and the naive Bayes classification were used to recognize the conventional gesture technique and needed a large number of data for the identification of gesture patterns.
[64]	They provided a study of existing deep learning methodologies for action and gesture detection in picture sequences, as well as a taxonomy that outlines key components of deep learning for both tasks.	They looked through the suggested architectures, fusion methodologies, primary datasets, and competitions in depth. They described and analyzed the key works presented so far, focusing on how they deal with the temporal component of data and suggesting potential and challenges for future study.
[65]	They solve the problems by employing an end-to-end learning recurrent 3D convolutional neural network. They created a spatiotemporal transformer module with recurrent connections between surrounding time slices that can dynamically change a 3D feature map into a canonical view in both space and time.	The main challenge in egocentric vision gesture detection is the global camera motion created by the device wearer’s spontaneous head movement.
[66]	To categorize video sequences of hand motions, a long-term recurrent convolution network is utilized. Long-term recurrent convolution is the most common kind of long-term recurrent convolution. Multiple frames captured from a video sequence are fed into a network to conduct categorization in a network-based action classifier.	Apart from lowering the accuracy of the classifier, the inclusion of several frames increases the computing complexity of the system.
[67]	The MEMP network’s major characteristic is that it extracts and predicts the temporal and spatial feature information of gesture video numerous times, allowing for great accuracy. MEMP stands for multiple extraction and multiple prediction.	They present a neural network with an alternative fusion of 3D CNN and ConvLSTM since each kind of neural network structure has its own constraints. MEMP was developed by them.
[68]	This research introduces a new machine learning architecture that is especially built for gesture identification based on radio frequency. They are particularly interested in high-frequency (60 GHz) short-range radar sensing, such as Google’s Soli sensor.	The signal has certain unique characteristics, such as the ability to resolve motion at a very fine level and the ability to segment in range and velocity space rather than picture space. This allows for the identification of new sorts of inputs, but it makes the design of input recognition algorithms much more challenging.
[69]	They propose learning spatio-temporal properties from successive video frames using a 3D convolutional neural network (CNN). They test their method using recordings of robot-assisted suturing on a bench-top model from the JIGSAWS dataset, which is freely accessible.	Recognizing surgical gestures automatically is an important step in gaining a complete grasp of surgical expertise. Automatic skill evaluation, intra-operative monitoring of essential surgical processes, and semi-automation of surgical activities are all possible applications.
[70,71]	They blur the image frames from videos to remove the background noise. The photos are then converted to HSV color mode. They transform the picture to black-and-white format through dilation, erosion, filtering, and thresholding. Finally, hand movements are identified using SVM.	Gesture-based technology may assist the handicapped, as well as the general public, to maintain their safety and requirements. Due to the significant changeability of the properties of each motion with regard to various persons, gesture detection from video streams is a complicated matter.
[72,73]	The purpose of this study is to offer a method for Hajj applications that is based on a convolutional neural network model. They also created a technique for counting and then assessing crowd density. The model employs an architecture that recognizes each individual in the crowd, marks their head position with a bounding box, and counts them in their own unique dataset (HAJJ-Crowd).	There has been a growth in interest in the improvement of video analytics and visual monitoring to better the safety and security of pilgrims while in Makkah. It is mostly due to the fact that Hajj is a one-of-a-kind event with hundreds of thousands of people crowded into a small area.
[74]	This study presents crowd density analysis using machine learning. The primary goal of this model is to find the best machine learning method for crowd density categorization with the greatest performance.	Crowd control is essential for ensuring crowd safety. Crowd monitoring is an efficient method of observing, controlling, and comprehending crowd behavior.