Design And Implementation Of Procession Scoring System Based On Human Pose Estimation

The procession training has been evaluated by human eye observation for all the time,and since there is no unified quantitative standard,this traditional method lacks of objectivity and has low efficiency.This thesis will combine human pose estimation technology with pedestrian detection,and use top-down architecture to implement multi-person pose estimation,then extract each pedestrian's skeleton structure from video frames.Based on the above,an automatic scoring system will be developed to evaluate the uniformity of procession quantitatively,which can not only assess the effect of procession training objectively and accurately,but also indicate the direction in which the procession training can be improved.Human pose estimation is the process of detecting the position of the human body's each part and calculating the direction and scale information from the image.Human pose estimation has been a research focus in the field of computer vision for all the time.Deep learning,based on convolutional neural network,has the ability of learning image features automatically,and it can obtain the optimal representation of implicit features by learning massive images.Deep leaning has powerful ability of feature extraction and representation,and then gets the nonlinear mapping from image to human pose by regression or classification operation.Deep learning has become the primary method in the field of human pose estimation.This thesis is based on the classical multi-stage regression convolutional neural network CPM(Convolutional Pose Machines)to achieve single-person pose estimation.The CPM uses the form of full convolution,and uses the multi-stage regression thermogram to extract the joint coordinates of human body.The network search space of CPM is large,so this thesis uses the Hough line detection technology to extract the straight lines of human body,then constructs the HOG feature vectors,which are injected into the convolutional pose machines as external prior knowledge to guide the network learning and narrow the search space.Compared with the original network,the improved network has achieved a precision improvement to some extent.Pedestrian detection is the process of determining whether there are pedestrian targets or not from static images or video sequences by acquiring region proposals,extracting feature descriptors,and applying the human classifier,and then getting thebounding boxes of detected pedestrians.HOG+SVM is a classic method of pedestrian detection,while the HOG feature descriptor has a large feature dimension.In this thesis,the PCA algorithm is used to reduce the dimension of extracted HOG features,eliminate redundant information,reduce noise interference,and retain the effective features that contribute to the SVM classification most,thus improves the accuracy of pedestrian detector.Based on the improved convolutional pose machines model and pedestrian detector,this thesis constructs a multi-person pose estimation model,which realizes multi-person pose estimation for images and video.Finally,this thesis designs and implements a procession scoring system.The system provides a GUI interaction window and builds a Web visualization platform,which has good usability and user experience.