Research On Efficient Human Keypoint Detection And Action Recognition Algorithm

Human keypoint detection is an important research direction in the field of computer vision.After obtaining the key points of the human body,it can be further used to identify human behavior.These technologies have a wide range of applications in real life,such as intelligent security monitoring in public places,human-machine interaction,and so on.Existing human keypoint detection researches mostly focus on improving the detection accuracy of the algorithm,which leads to the complexity of the network structure,the network parameters and calculations are too large,and the network prediction takes too long.Among the behavior recognition algorithms based on human key points,some algorithms encode human key point information into the form of 2D images,which reduces the efficiency of the algorithm's prediction;other algorithms do not make full use of human key point feature information.For these reasons,the main contents of this article are:1.Design an efficient human keypoint detection algorithm to greatly reduce the amount of network parameters and calculations and increase the speed of network prediction.First,this article uses a top-down approach.The keypoint detection network uses the basic network to extract features and combines the upsampling operation to predict the efficient structure of the keypoint heat map.Then,when extracting features from the basic network,in order to prevent positioning errors caused by downsampling,reduce the number of times of network downsampling,and use dilated convolution to improve the receptive field of the network.In order to balance the prediction effect and speed of the network,verify the impact of different network depths,channel numbers,upsampling methods,and attention mechanisms on the network.Finally,knowledge distillation is used to improve the detection accuracy of the network.In addition,the human detection network is compressed to improve the overall efficiency of the algorithm.2.After obtaining the keypoints of the human,perform single-frame human posture recognition for behaviors that focus on human posture;for general human behavior,use sequence information to identify behavior types.In the single-frame human pose recognition,the normalization of the key points of the human body is simplified,the human pose data is enhanced,and the accuracy rates of image recognition and key point coordinate recognition are verified.In sequence human behavior recognition,firstly,a variety of static feature coding is performed on the key points of the human body in the frame,and a variety of dynamic feature codes are performed on the key points of the human in the form of trajectories;then two networks are designed to learn static and dynamic,respectively.Information,and finally fuse the characteristic information of the two networks to predict the action type.