Human Activity Prediction Algorithm Based On Spatial-Temporal And-or Graph

Due to the non-Markovian property of human activities and the increasingly complex background of human activities,in order to ensure the safety of people,people install cameras in many places,but the camera can only play a monitoring role,not completely avoid the occurrence of danger.In recent years,the hierarchical representation model of human activities proposed by scholars has provided new ideas for human activity prediction and danger warning.In this thesis,based on the accurate target tracking detection method,activity prediction is carried out by constructing human activity representation models in different scenarios.Moreover,activity prediction can prevent people from danger in advance,which is of great significance in people's life.In this thesis,the spatial-temporal And-Or graph(ST-AOG)model is used to capture the grammatical structure of events,The spatial-temporal And-Or graph is composed of the temporal And-Or graph(T-AOG)constructed by representing the temporal stochastic grammar of human sub-activities and the spatial And-Or graph(S-AOG)representing the interrelationship between objects in the video.Based on different scenarios of target tracking under video detection,and then through the analysis of the goal of relationship between the activity for extraction,and get the whole testing video event description,the last time of grammar and Earley parsing algorithm to predict the future activities,so as to realize human activity prediction.The specific work is as follows: Firstly,the interested targets are tracked and detected in different scenarios.As a result of the detection accuracy will affect the accuracy of tracking,so for the target detection technology,this article uses YOLO?v3 for target detection,can guarantee the accuracy of the target detection and real-time,and then use the multiple target tracking method for target tracking DeepSort,its continues the Sort of idea of Kalman filter and Hungarian algorithm,based on the appearance of evaluation index and the exterior information,increased the depth for a long time to keep out of target tracking,reduce the switching frequency of the target id.Improved spatial And-Or graph(S-AOG)by YOLO?v3+DeepSort for target tracking.Secondly,the spatial-temporal And-Or graph(ST-AOG)is constructed is constructed based on the improved spatial And-Or graph and the temporal And-Or graph based on the temporal stochastic grammar.The root node of the spatial And-Or graph(S-AOG)is taken as the leaf node of the temporal And-Or graph,Through the context random grammar model,the events in the scene are analyzed,and the temporal And-Or graph(T-AOG)of the hierarchical combination model of events is obtained.Through the analysis of relationship between different targets in video,define different sub-activities for each scene,according to the similar triangle method to calculate the actual location of the image coordinates,and the velocity of the target in the video is calculated,and then use the sub-activity extraction algorithm proposed in this thesis to obtain the sub-activity labels.The sub-activity of the target of interest in the detected video is taken as input,and the future sub-activity is predicted by spatial-temporal And-Or graph(ST-AOG)and Earley parsing algorithm,so as to obtain the event optimal parsing tree to predict human activities,and the danger warning is carried out under special scenes.Finally,using the algorithm of this thesis to extract sub-activities in different scene videos for human activity prediction,and constructing a confusion matrix for the extracted subactivity labels,experiments show that the accuracy of sub-activity extraction of the human activity prediction algorithm in this thesis can reach about 90%.The prediction results of human activities are also consistent with the actual human activities in the video.