Human Pose Estimation And Tracking In Crowded Scene

Pose estimation enables computers to recognize the keypoints of human in images.As the basic for human motion recognition and behavior analysis,pose estimation is widely used in intelligent monitoring,human-computer interaction and other fields.It is a hotspot in the field of computer vision.With the introduction of deep learning,the performance of multihuman body pose estimation based on static images has been improved,and then the multihuman body pose tracking based on video has been developed.However,due to factors such as interactive occlusion of the human body,diversified clothing,self-occlusion,changes in lighting,and interference with scene objects,the estimation and tracking of human poses for multiple people is still a challenging topic.In this thesis,a new method for human pose estimation and tracking is proposed to improve the accuracy in crowded scenes.The main research contents are as follows:(1)The cascaded high-resolution representation network(CHRN)is constructed for human keypoint detection,which can accurately locate human keypoints through the combination of Global Net and Refine Net.In the cascaded high-resolution representation network,the Global Net extracts the deep feature to get more comprehensive and effective information in images;Refine Net cascades the multi-stage deep features extracted by Global Net,and improves the overall accuracy of keypoints with online hard keypoint mining.(2)This thesis proposes a two-stage training optimization method which combines bottom-up and top-down.In the training process,CHRN is first trained with bottom-up method to get the basic model.According to the transfer learning,and then train the basic model with top-down method once again to get the pose estimation model.This train method can reduce the influence of human detection results,complex background and occlusion,and increase the attention of human target.In the process of testing,the top-down method is used to output the human pose estimation results with human keypoints clustering,avoiding the influence of human target detection on the final results,and ensuring the accuracy of human pose estimation.(3)A human pose tracking method based on Kalman filter optimization is proposed.In this method,the CHRN optimized by two-stage training is used to estimate the human pose,and then the Kalman filter is used to optimize the human pose estimation results by introducing timing information.And the detection results of the model are modified during the tracking process.Further this method improves the accuracy of human pose tracking in the video.Experiments on MSCOCO dataset,MPII dataset and Pose Track dataset show that the pose estimation and tracking method proposed in this thesis are superior to the current stateof-the-art algorithms.