Research On Human Motion Prediction Method Based On Two-person Interaction

With the development of 5G applications,multimedia data,especially image and video data,is growing explosively.Human motion prediction,as an important component of human behavior analysis,is gradually becoming an increasingly important research topic,with wideranging potential for development in human-machine interaction,autonomous driving,sports coaching,and other areas.In natural scenes,there are various interaction relationships between people.When predicting the motion of two people in a scene,if the cooperative relationship between the body joints of the two individuals can be captured,the prediction error of the two-person motion can be further reduced.However,the existing graph convolutional network-based methods are not explicit enough in modeling interaction relationships.Therefore,there is an urgent need for a two-person motion prediction method with more explicit interaction features.This research topic focuses on the study of a motion prediction method based on two-person interaction,which is based on graph convolutional network and multi-scale modeling strategy to capture the interaction relationships between two individuals at different granularities.Firstly,the graph convolutional network structure is used to obtain multiscale motion features through multi-scale feature aggregation and separation in the feature extraction part,which enhances the model’s ability to capture the spatial relationships within the human body.Secondly,an interaction feature extraction module based on attention mechanism is designed to generate an interaction feature matrix based on the current twoperson motion pattern,representing the two-person interaction relationships at different granularities.By introducing an action recognition task in a multi-task framework,the model’s ability to judge the two-person motion patterns is further improved.This research topic is derived from Beijing Post-Telecom Visual Intelligence Joint Laboratory Project.The main research content and innovative results are as follows:To address the issue of unclear cooperative relationships between individuals in existing two-person motion prediction methods,this research proposes a multi-scale graph convolution-based method for twoperson motion prediction.Firstly,the multi-scale motion features are obtained through multi-scale feature aggregation and separation in the feature extraction part using the graph convolutional network structure,which enhances the model’s ability to capture the spatial relationships within the human body.Secondly,an interaction feature extraction module based on attention mechanism is designed to generate an interaction feature matrix based on the current two-person motion pattern,representing the two-person interaction relationships at different granularities.On the Extreme Pose Interaction two-person interaction action dataset,the average joint position error of the proposed method is reduced by 21.9%and 10.1%compared to the baseline method and the optimal method,respectively.To address the issue of large differences in two-person motion patterns between different action categories in two-person interaction scenes,this research proposes a multi-task-based two-person motion prediction method,which improves the model’s ability to judge two-person motion patterns by introducing an action recognition task.Firstly,multiorder motion information is modeled and fused,enhancing the representation ability of two-person motion features.Secondly,a multiscale feature fusion module is designed to fuse the two-person motion features at different scales.Finally,a loss function suitable for multi-task frameworks is designed.Compared to the baseline method under singletask conditions,the proposed method reduces the average joint position error on the Extreme Pose Interaction dataset by 4.1%.