| With the rapid development of computer technology and artificial intelligence technology,human-computer interaction technology has been widely concerned.Hand gesture interaction is simple and efficient,and can be applied in many fields,such as intelligent driving,smart home and virtual games.At present,in the field of vision based dynamic gesture recognition,there are some problems,such as the part of the gesture can not be recognized,the complex background interference of the gesture,and the gesture can not be detected when the movement speed is fast.Therefore,we need to find an efficient dynamic gesture recognition method.Aiming at the above problems,this paper designs a dynamic gesture recognition model combined with gesture tracking trajectory by using the knowledge of convolutional neural network.This paper compares and analyzes the key frame extraction algorithms such as clustering method,frame sampling method,background weakening method and inter-frame difference method.Finally,three-frame difference method is used to extract key frames of video,and some dynamic gestures with large application degree and strong proportion of gestures in real life are screened out.Then,the data set is annotated by Lableme open source tagging software,and the data set is made by image clipping and image expansion.Yolov4 algorithm is improved in this paper for the gesture detection stage,which has the advantages of multi-objective detection,and the redundancy and complexity of the single-objective detection network structure make the training difficult and the detection efficiency low.1×1 convolution kernel is added after each residual module to further reduce the output dimension;a linear activation function is used in the first convolution layer to avoid the loss of feature images in the low-dimensional convolution layer;the number of layers of residual network in each module is adjusted in the residual module.Combined with the Deep-sort target tracking algorithm,the trajectory of the hand is finally obtained.DenseNet-BC network is selected in this paper to complete the trajectory classification.To save training time without loss of recognition rate,this paper compares the number of DenseNet-BC-169 、 DenseNet-BC-121 and DenseNetBC-201 networks,and finally selects the DenseNet-BC-169 network layer.Finally,the model is compared with the champion FLIXT of dynamic gesture recognition challenge.The results show that the accuracy is improved by 6.13%under the same data set.Can process the number of video frames up to 30 fps,to meet the continuity of video detection requirements.The feasibility of this algorithm is proved. |
