Design And Implementation Of Human Keypoint Detection Based On Video Features

With the development of computing power and deep neural network technology,"Machine Vision'"is being studied by more and more scholars and enterprises.People try to make machines"recognize"images by simulating human understanding mode of images.Among them,the research on object detection has attracted much attention within the industry.As a sub-classification of object detection,keypoint detection is also a research hotspot in the field of computer vision.Keypoint detection is expected to solve the problem of how to extract features from static images and speculate human body posture information according to the spatial relationship of human body.Limited by the diversity of human actions and the complexity of image acquisition environment,it is still facing great challenges to solve the keypoint detection problem in general scenarios.In order to achieve high-precision detection of human keypoints,the author designs and implements a system to detect human keypoints.Utilizing motion video as input,The system could get the detection results of 2D keypoints as well as predict the body length of inspector by means of detecting the pre-processed image frames and optimizes the detection results according to the context information.The structure of this thesis is as followed,This thesis describes a keypoint detection system,which consists of pre-processing module,keypoint detection module and final post-processing module.Firstly,the background information and research significance of keypoint detection are analyzed,and then introduce the related technologies in the second chapter.In the third chapter,the key algorithms used in this system are studied,designed and evaluated.Then in the fourth chapter,the detailed requirement analysis and the corresponding module outline design of the keypoint detection system are carried out.In the fifth chapter of detailed design,the functions proposed in the outline design are implemented.In the sixth chapter,the test results of the proposed system are shown.Finally,in the seventh chapter,the work done in this thesis is summarized and analyzed,and prospect the successor work.On the basis of that 2D keypoint detection results realizing common data set baselines,this thesis optimizes the testing method of data detected by video unit.By combining the optical flow chart with the front and back frames in the sliding window,the keypoint of errors in detection are corrected.In addition,additional 3D keypoint detection results are output and processed to obtain the prediction of the true body length of the inspector.The experimental results show that the keypoint detection system designed and implemented in this thesis performs well in both common data sets and actual testing.