Research On Crucial Issues Of Gesture Recognition Algorithm Based On Neural Network

Gesture interference is an essential component of human-computer interaction which has been widely used in recent years in areas such as Automotive Systems,Internet of Things,Industrial Presentations,Virtual Reality,and Physical Games.Research of gesture recognition has been a popular academic field,and scholars have proposed many methods,such as gesture recognition based on data gloves,etc.But these methods have some drawbacks,such as the users are required to carry the device,the operation process is cumbersome,which is contrary to the requirements of human-computer interaction for naturalness.By analyzing the gesture recognition-related technologies at both domestic and overseas,it is found that the precision rate of traditional methods is low,weak gesture segmentation effect,and some of them do not have robustness.Aiming to eliminate the above shortcomings,an approach to static gesture recognition based on Neural Network is developed in the paper which utilizes Neural Network to enhance the classical gesture recognition algorithm with translation and rotation invariance for image processing domain.The work accomplished here consists of the following components.1.Improving the gesture segmentation method.Firstly,conventional gesture recognition procedures are analyzed,the strengths and weaknesses of gesture segmentation methods are compared,and the skin color segmentation method based on YCr Cb Gaussian skin tone model is ultimately selected through comparisons.The face and hand skin tones are considered to be consistent and the face is masked using the face position detection algorithm.After processing the skin color,the largest connected area is selected as the gesture area using the connected area algorithm.Such an algorithm eliminates the background of the gesture image effectively and completely splits the gesture region.2.Applying Neural Network Algorithms to gesture recognition.For the traditional gesture recognition approach,template matching method and statistical learning method are presented with their advantages and disadvantages,and the two algorithms are limited and require heavy data set.The Neural Network based gesture recognition approach is finally adopted.The performance of various categories of Neural Networks is compared in this paper,while the effect of different optimization algorithms to the result are compared and the effect of varying migration learning training is examined.After analyzing the experimental results,the optimal algorithm was selected from them.Resnet-152,a deep residual network in Neural Networks,was trained on the gesture dataset,and the network generated from the training was used to design the gesture recognition system.3.Designed a real-time gesture recognition system.Using the 3D model-based gesture recognition method,the Mediapipe framework is utilized to achieve real-time recognition of Chinese numeric and alphabetic gestures.A simple real-time gesture recognition system is also designed based on Open CV software.These two subsystems complement the non-real-time gesture recognition system based on Neural Network in this paper.Eventually,three subsystems are engineered for the paper,which can recognize a total of 49 different static gestures.The experimental outcomes demonstrate that the gesture recognition accuracy of this paper reaches 93.58%.The proposed algorithm will contribute to a more natural and harmonious human-computer interaction in the promising future.