Motion Assistant System Based On The Key Of Anti-Confusing Human Pose Estimation

In the process of sports or fitness training,people's irregular movements will affect the training effect and even lead to sports injury.However,the standard movements of various sports activities require professional guidance,so it is difficult to find out whether the movements are standard.In recent years,human pose estimation has become a hot spot in computer vision technology research.Deep learning models can effectively recognize joint points of human bodies and their trajectories in pictures or videos,and can help evaluate the actions of target human bodies.The use of deep learning methods to automatically determine whether the movements during the exercise are standard or not can enable athletes to evaluate movements without the guidance of professionals to correct incorrect movements,thereby achieving healthy and efficient exercise and preventing sports injuries.However,it is common for other people to block or move in the vicinity of the movement process,which leads to deviations in human body movement recognition and affects the evaluation of movements.As a result,incorrect movements cannot be corrected effectively,but trainers are misled instead.This paper proposes a human pose estimation network model based on regional perception network to strengthen anti-aliasing ability in the case of severe occlusion,close human overlap and occlusion due to symmetrical appearance.This method involves three key technologies:data augmentation,feature learning and prediction fusion.First,a set of data enhancement methods is proposed to specifically solve the problem of human body pose estimation texture obfuscation.A large amount of data is generated using analytic-based data enhancement methods to synthesize obfuscated textures,and the obfuscated limb component pixels are overlaid onto the current human body as a whole.The ability of the model to distinguish between confused limbs.Secondly,a feature pyramid root module is designed to fuse high-and low-level network features.Compared with the previous work,which effectively reduced the pixel loss caused by downsampling the input image,this module can extract richer image features during the network input stage,which is more sufficient.The effective pixels of the input image are used to solve the problem of texture confusion,and the end-to-end image region processing structure extracted from the effective region can be used to mine better target features.Third,a cascading fusion strategy that is effective for the multi-stage output results of the network is designed.The heat map confidence information of each key point is considered to effectively exclude outliers,so that bad predictions are explicitly excluded,and more accurate prediction results are obtained.Experimental results on two common human pose estimation datasets,MPII and LSP,prove the effectiveness of the method,especially for confusing joints,and the recognition accuracy is significantly improved.Based on the target human joint point recognition technology,a motion assistance system is implemented,which can determine whether the trainer's movement is standard through key frame pictures.The system recognizes the joint points of the key frames of the movement,compares the standard movement with the practice movement,calculates the similarity,and obtains the evaluation result of the practice movement.