Research On Gesture Recognition Algorithm Based On Deep Learning

People often use some gestures to communicate in daily life,the technology of realizing human-computer interaction through gesture recognition can be naturally applied to sign language recognition,vehicle equipment control,smart home and other aspects,which is one of the key research directions of human-computer natural interaction in the future.But in terms of interaction,the real-time and accuracy of gesture recognition is the key.Traditional gesture recognition research has entered a bottleneck period,and it is difficult to go further in real-time and accuracy.The rapid development of deep learning provides a new research direction for gesture recognition.This article first analyzes the research background and significance of gesture recognition,introduces the research status and development trends,and then explains the main research content and organizational structure of this article.Then,a brief introduction is made to the traditional gesture recognition method and the deep learning-based gesture recognition method,and the advantages and disadvantages of these two methods are compared.The deep learning-based gesture recognition method is selected as the research focus of this article.Then it analyzes the superiority of YOLOv3 algorithm in the same type of algorithm,and chooses to realize gesture recognition based on YOLOv3 algorithm.This article briefly introduces the current mainstream learning framework,chooses to use the Py Torch framework to develop gesture recognition algorithms,and explains the algorithm implementation environment of this article.Then the network structure and loss function of the YOLOv3 algorithm are studied,related evaluation indicators of the algorithm are introduced,and three improvements are proposed for the YOLOv3 algorithm for gesture recognition,which lays the foundation for the improvement of the algorithm.Next,the importance of the data set to the deep learning algorithm is explained.By comparing the existing public gesture data sets,the NUS Hand Posture dataset-II gesture data set suitable for the study of gesture recognition in this article is fused with the homemade gesture data set.And through the data enhancement method,a fusion gesture data set containing 12600 complex background pictures was created,and the superiority of the fusion gesture data set was verified through experiments.Then the Darknet-53 network structure is optimized,and the optimization ideas and optimization methods are explained,and the effect of the optimized network structure is verified.Then,the prediction output layer of YOLOv3 is improved,the improved ideas and methods are explained,and the effect of the improved prediction output layer is verified.Then the application test of the F-YOLOv3 algorithm proposed in this paper is conducted,the average gesture recognition speed is 31 ms,and the accuracy rate is 97.1%.The test results show that the improved F-YOLOv3 gesture recognition algorithm based on YOLOv3 has faster recognition speed and higher accuracy.