Research On Human Action Recognition Based On Skeleton Feature

In recent years,with the rapid development of information technology and the internet,videos have gradually become a primary carrier of information in many fields.Human actions are important information in videos.As a key technology for video understanding,human action recognition has received increasing attention and has great demand and broad application prospects in public safety monitoring,elderly care,sports and other fields.Although some progress has been made in action recognition research,there are still some challenges: the contextual information in video data affects the recognition effect of models;Existing research mainly trains and tests on pre-defined datasets without considering real-world application scenarios.This thesis proposes a research on human action recognition based on skeleton features,and combines attention mechanisms to make the model focus more on useful features for the tasks.At the same time,in order to solve the problem of open set,transforming the open set problem into an uncertainty estimation problem based on evidence neural networks is proposed.The specific research content and innovation points of this thesis are as follows:(1)Unlike other methods of action recognition that use RGB video or optical flow,this thesis uses skeleton features for human action recognition.Skeleton features only extract key points of the human body and do not involve contextual information.It is robust to environmental changes such as background and lighting,reducing the impact of subjective factors on model performance.Additionally,skeleton features have the advantage of being lightweight.(2)Action recognition based on skeleton features mainly uses graph convolutional networks,but the computational complexity of graph convolutional networks is linearly related to the number of people and body joint nodes because each human joint is treated as a node.This thesis uses a three-dimensional convolutional neural network,combined with 3D human heat map volume to recognize human actions,and combines attention mechanisms based on previous work to make the network model more focused on task-relevant features.(3)In the research of human action recognition,the model is trained on a given dataset.It classifies actions into one of the predefined categories during testing.However,the categories that can be defined by the dataset are limited and in reality there will always be actions that fall outside of these defined categories.Therefore,action recognition is essentially an open set problem which requires models to not only recognize known actions but also reject unknown ones.This thesis proposes research on open set action recognition based on evidence neural networks which transforms open set action recognition into a problem of uncertainty estimation.The evidence function outputs evidence for each category,allowing for calculation of prediction uncertainty and probability for each predicted category.Known actions have low uncertainty,and known classes can be classified based on predicted class probabilities,while unknown actions have high uncertainty,and are classified as a class of "unknown classes.".This thesis conducted experiments on two mainstream action recognition datasets,UCF101 and HMDB51,and the experimental results showed the effectiveness of the proposed method.