Gesture Detection And Recognition Based On Convolutional Neural Network

Gesture plays an important role in human life.It can not only help deaf people communicate,but also make human-computer interaction more intelligent.The requirements of gesture recognition algorithm are very high.The algorithm not only needs strong robustness,can adapt to complex environment,but also needs accuracy and real-time.Traditional gesture recognition methods based on computer vision generally divide gesture recognition into two parts: gesture feature extraction and gesture classification.Manual feature extraction often requires professionals to process gesture data and then extract gesture features.This feature extraction method is complicated and has no generalization.When the background of gesture is complex,the effect of gesture recognition is poor.With the continuous development of deep learning,the research of gesture recognition using convolutional neural network has become a new trend.In the modern social environment with a large amount of data,convolutional neural network has a good prospect.Therefore,this thesis uses convolutional neural network for gesture detection and recognition.First of all,this thesis studies and compares various convolutional neural network algorithms,and analyzes the Faster R-CNN two-stage target detection algorithm and efficient det,YOLO and other one-stage target detection algorithms.Considering the speed and accuracy of the algorithm,YOLO algorithm with good performance in both speed and accuracy is selected for gesture detection and recognition.In the second place,aiming at the problem of poor small target detection and recognition of YOLO algorithm,the layer number of the feature pyramid of YOLOv4 backbone feature extraction network is increased from three layers to four layers,which strengthens the detection of small target gesture.On this basis,the K-means++ algorithm is used to generate an anchor box suitable for gesture detection to speed up network detection and gesture recognition.Finally,the paper studies the YOLOv4-tiny algorithm which is suitable for mobile gesture detection and recognition,and proposes an improved YOLOv4-tiny tiny gesture detection and recognition algorithm.On the basis of YOLOv4-tiny network structure,the SPP spatial pyramid pooling module is added,which integrates the local and global features of the image,and enhances the accurate positioning ability of the network.At the same time,a 1 × 1 convolution is added after the three maximum pooling layers and the new SPP module of YOLOv4-tiny network,which reduces the network parameters and improves the network prediction speed.The improved algorithm is deployed in Android mobile terminal to realize real-time gesture detection and recognition.