Research On Skeleton-based Group Activity Recognition With Interactive Relation

Group activity recognition refers to the systematic modeling and recognition of the human actions and overall activity,which has important scientific research and application value in the fields of social security,sports,and video analysis.At present,there are still several key problems to be solved in the field of group activity recognition:Firstly,the parameter quantity optimization problem under the group scale,the number of individuals in the group scene is huge,resulting in high parameter amount and computational complexity,and the application is more difficult.Secondly,the modeling problem of individual multi-dependency interaction relationship,which has complex oneto-many and many-to-many interactions,making it difficult to extract interactions feature.Thirdly,the problem of interactive action capture in a specific scene.Individual actions potentially obey the rules of the scene,and actions behave differently in different scenes,making it difficult to capture specific interactions.Based on the above problems,this thesis integrates the pose estimation algorithm and graph convolution theory into the modeling process of group activity recognition,and takes parameter quantity optimization,interaction relationship extraction and scene information fusion as the key goals for research:(1)This thesis proposes a parameter quantity optimization method based on human skeleton for shift graph convolution.Via adjusting the skeleton connection structure,combined with the point convolution method to reduce the amount of parameters and calculation steps,and use the shift operation to reduce the loss of information;(2)This thesis proposes a multi-dependency interaction modeling method based on global context information.It uses global context information to model related actions between individuals,infers multi-dependent interactions,and extracts more discriminative interaction features;(3)This thesis proposes a scene information fusion method based on temporal attention mask matrix.It generates an attention mask matrix based on scene information,judges the relevance of actions and scenes through time sequence weighting,and captures interactive action features that conform to the rules of the scene.The experimental results on two benchmark datasets,Collective Activity Dataset and Volleyball Dataset,demonstrate that the method proposed in this thesis effectively improves the accuracy and efficiency of group activity recognition,and provides a research reference for this field.