Research On Action Recognition Algorithm Based On Human Skeleton

Human action recognition is a very challenging problem in the field of computer vision,which has received more and more attention in recent years because of its broad application scenarios.In general,human behavior can be recognized in many ways,such as appearance,depth,optical flow,and human skeletons.In these modes,the information usually conveyed by the dynamic human skeletons is more focused on the movement characteristics of the human body than other types of patterns.Based on the human skeleton information in the video data,this work divides the whole thesis into two research questions: the extraction of the skeleton information in the video and the construction of the action recognition model.Aiming at the task of effectively extracting human skeleton data from video,this thesis designs the multi-target tracking algorithm for human skeleton extraction by combing the problems existing in the actual human pose estimation tool in actual scenes.Firstly,in order to solve the problem that the skeleton sequence is easy to cause the wrong matching of skeleton object in the process of frame-by-frame processing,a new skeleton matching algorithm is proposed to retain the correct characteristic information of the skeleton sequence data in the time dimension.Then,this thesis proposes to smooth the motion path of the joints according to the spatial position of the inner joints of adjacent frames,so as to eliminate the errors in the results of attitude estimation tools and improve the accuracy of the spatial position of the data.For the smooth implementation of the above two modules and the effective training of the subsequent recognition model,a normalized unit module of skeleton data is designed,which enables the model to simultaneously accept data of different resolutions,different shooting angles and different proportions of characters(the proportion of characters in the video picture).Aiming at the construction of the human skeleton-based action recognition model,this thesis first selects a scheme based on the Spatio Temporal Graph Convolution Network.The human action recognition model of this scheme makes good use of human expert prior knowledge,which helps to improve the interpretability of the model.But relatively,because the setting of the topological correlation between joint points in this scheme is relatively fixed,it reduces or limits the ability of deep learning networks based on data-driven learning of potentially unknown information characteristics.Therefore,the Adapt Spatial Temporal Graph Convo-lutional Networks is proposed to release the driving ability of big data as much as possible and give full play to the learning ability of recognition model based on dynamic human skeleton.On the other hand,in order to improve the possibility of the algorithm model landing on the end-side entity,this thesis studies the reduction of model parameters by using the knowledge distillation,and reduces the number of model parameters on the premise of retaining the learning ability of the original model as much as possible,so as to provide a basis for the future scheme landing on the real-time scene.Finally,this thesis designs corresponding simulation experiments for the above research modules,and rigorously demonstrates the performance of the scheme.The results show that the schemes involved in this thesis show better stability and convergence speed than other models when they complete the basic functional tasks.