Research On Multi-Task Learning Algorithm Of Human Pose Estimation In Video

Human pose estimation in video has become an important method for analyzing human behavior,and it has been widely used in intelligent monitoring and human-computer interaction.The single-task learning method uses multi-layer network to extract joint features for human pose estimation.The multi-task learning method can obtain a large amount of task-related information in limited inputs with multiple related tasks,which provides a new idea of human pose estimation.Therefore,based on multi-task learning,this thesis study the neural network structure design and post-processing method of human pose estimation based on multi-task learning,and analyze the impact of multi-task learning on the detection effect of human pose estimation.The following work is completed in this thesis.After analyzing the joint jitter and undetected causes in multi-task networks,DP-Pose as a post-processing method for human pose estimation is proposed based on dynamic programming.The points in the heatmap are used as candidate joint points to form candidate sequences.The algorithm solves the stable joint sequence from the candidate sequences,and the experiment shows that the method can effectively recover the undetected joint points and correct the jittery joint points.A multi-task pose estimation network FRCNN-PoseNet is proposed to solve the problem of ambiguity in the matching of joint points in the image.FRCNN-PoseNet combines the advantages of the top-down and bottom-up methods of human pose estimation,and uses the bounding box of the human body obtained by object detection to limit the joint matching area in the bottom-up method.The method can improve the accuracy of human pose estimation effectively.For human pose estimation in video sequence.MTS-PoseNet is proposed in this thesis.Combining with the FRCNN-PoseNet network structure,it fuses the spatiotemporal features of the input data and enhances the input data to improve the model generalization performance.Experiments show that the spatiotemporal features have a good expression of human pose,and further improve the performance of human pose estimation.This paper studies the possibility of multi-task learning and different post-processing methods in human pose estimation,and provides some ideas for subsequent research on the application of multi-task learning in human pose estimation.