Research On Single Person Pose Estimation And Tracking Algorithm Based On Object Detection

In recent years,developments in the fields of intelligent monitoring,human-computer interaction,smart home,and autonomous driving have put forward demands for research on the detection and tracking of human and their behaviors.Recognizing and tracking the human skeleton structure has become an important basis for human motion analysis.Object detection and tracking also have new development requirements,which is human pose estimation and tracking.Pose estimation is to accurately identify and locate the positions of human keypoints in the image;pose tracking is to detect and track the human keypoints in the video by using the relationship between the previous and subsequent frames in the video.This paper mainly studies the single person pose estimation and tracking based on autonomously identifying the target person,that is,the target person's pose estimation and tracking tasks are performed after the person's position is automatically detected using the object detection algorithm in the first frame.This research is of great significance to promote human behavior analysis and prediction and scene understanding.The current object detection algorithms usually list potential target positions in the form of bounding boxes,and then classify the bounding boxes for foreground and background and regress to their precise positions.This detection method is relatively inefficient and requires expensive post-processing method;Meanwhile,there are few studies on single-person pose tracking algorithms,and most multi-person pose tracking algorithms based on optical flow focus more on the accuracy of pose estimation than the speed of the algorithm,failing to meet the standards of online tracking.In view of the above problems,the main contents are as follows:(1)The basic theories of existing object detection methods based on anchor box mechanism and human pose tracking algorithm based on optical flow are investigated.The shortcomings of the anchor box mechanism and optical flow method are analyzed,and the overall schemes of anchor-free object detection and pose estimation and tracking based on Siamese networks are designed.(2)An anchor-free object detection algorithm based on Center Net is proposed to detect the target person in the first frame.The detection of the person is equivalent to the detection of the target center point and the regression of the target size.It is verified through experiments that the human detection algorithm in this paper has improved the target detection speed and accuracy compared with the classic anchor-based object detection algorithm.(3)An end-to-end pose estimation and tracking algorithm based on the Siamese network is designed.The proposed algorithm combined with region proposal network achieves multi-scale detection.The image preprocessing process and network framework are introduced in detail.In order to track the pose of the target person,the pose estimation branch is added to the Siamese network,and it shares the feature extraction network with the Siamese network.The network structure of pose estimation branch is designed.Based on the keypoints coordinates predicted by the heatmap,the discrete error caused by the convolution step is compensated,and the accuracy of keypoints detection is improved.The performance of the proposed algorithm is tested on a public dataset,and experimental results show that the proposed algorithm achieves a better trade-off between keypoints detection accuracy and tracking speed.(4)The trained model is transplanted to a portable embedded AI platform for testing.The difference of target tracking speed,accuracy and success rate on different hardware platforms is compared.Experiments on pose estimation and tracking in real-world scenarios were performed on the embedded AI platform.Experimental results show that the algorithm can achieve real-time pose tracking while achieving good pose estimation results.Using Siamese networks for pose tracking tasks was first proposed in this paper and it greatly improves the speed of pose tracking.Combined with the regional proposal network,the multi-scale object detection in the tracking process is realized.The proposed pose estimation and tracking algorithm has broad application prospects in the fields of intelligent monitoring and autonomous driving.