Research On Video Motion Analysis Based On Pose Estimation

With the improvement of living standards in recent years,people’s participation in sports has become higher and higher.However,physical exercise may not be well done without the guidance and the assessment of professional coaches.Nonstandard motions may cause the decrease of exercise effect or even physical damage.The emergence of motion analysis makes it convenient to evaluate the completion of actions and obtain professional guidance.Traditional motion analysis methods are mainly based on motion capturing devices,which suffer from inconvenient wearing,high cost,and low universality.Most related research in computer vision is limited within a certain part of video motion analysis,and there is a lack of an end-to-end solution that can overcome problems such as people’s motion blur,key point occlusion and similar action differentiation.Therefore,this thesis proposes a video motion analysis method based on pose estimation,which first removes the human motion blur and makes up the missing key points,then performs the action recognition and normative assessment.The main contributions of this thesis are summarized as follows:To solve the human motion blur caused by high-speed movement in the video,a method of human motion deblurring based on generative adversarial network is proposed,which uses Res Net and jumping connection to avoid the vanishing gradient problem and capture features in different scale.It also improves the composition of the loss function which helps improving the accuracy of motion analysis.The experiment results show that the peak signal-to-noise ratio and structural similarity of restored images generated by this method reach 35.27 d B and 0.956 on the Go Pro dataset,which are better than most related methods,and the details in those images are also well restored.To solve the problem that the existing action recognition methods are insufficient for similar actions,an action recognition method based on graph convolution network is proposed,which divides the human skeleton into several parts according to the human physical structure before extracting features.It makes it easier for the network to capture the differences between substructures,and helps improving the accuracy of action recognition.The experiment results show that the Top1 and Top5 accuracy on Kinetics dataset of this method reach 35.2% and 57.6% respectively,which are better than most related methods.To solve the problem that the existing motion normative assessment methods are not sufficiently universal,a method of motion normative assessment based on the spatialtemporal volume is proposed,which is able to assess the motion completion by the superposition of spatial-temporal volumes between human motion and standard motion,and can perform normative analysis on any motion without requiring specialized design.Besides,this method can be applied to a specific human body part for motion normative assessment,which solves the problem that the motion normative assessment is not precise enough.In summary,the experiment results show that the video motion analysis method based on pose estimation proposed in this thesis well solves the problems of human motion blur and key point detection missing,and can perform action recognition and normative assessment with higher accuracy and fineness compared with the existing methods.