Multi-Class Target Trajectory Prediction Based On Scene And Interaction Influence

With the increasing of intelligent systems,unmanned driving,intelligent robots and so on have gradually become a reality.Especially with the emergence of deep neural network,artificial intelligence science and technology have been developed unprecedentedly.With the prosperity and development of intelligent systems,people put forward more intelligent and safer requirements for artificial intelligence.Among them,the reasonable trajectory prediction of moving objects such as pedestrians is one of the mainstream research directions at present.Through reasonable trajectory prediction,the intelligent system can fully understand the current scene,provide reliable data sources for subsequent trajectory planning and active decision-making,and further improve the intelligent level of the intelligent system.Therefore,this paper proposes the research work of trajectory prediction for multi category targets based on scene and interaction,which is of great significance to the research of trajectory prediction theory and methods.By analyzing the influence of obstacles in the scene,the interaction between moving objects,the speed of the moving object itself and other factors on the motion trajectory,the Semantics-STGCNN trajectory prediction algorithm architecture is used to creatively propose the trajectory prediction research of multi category targets based on the scene and interaction.The main contents are as follows:(1)In order to accurately express the difference degree of interaction between different types of agents,the interaction is divided into three levels: weak,normal and strong according to prior knowledge.Through one-hot coding of three types of features,the adjacency matrix model of interaction features is constructed to extract the interaction features(2)In order to solve the influence of static obstacles on the trajectory of agents in the scene,a grid map of obstacles in the scene is constructed,and the scene features are extracted by combining the receptive field of agents.(3)The interaction characteristics,scene characteristics and the speed attribute space-time map in Semantics-STGCNN algorithm are fused to complete the construction of the space-time map containing the scene and interaction information.Then the space-time map is introduced into the Spatial-Temporal Graph convolution neural network to learn the space-time trajectory characteristics of the agents.Finally,the time extrapolator convolution neural network is used to achieve trajectory prediction.In order to verify the effectiveness of the proposed model algorithm,experiments were carried out on Stanford UAV dataset with multi-class trajectories,and m ADE,m FDE,a ADE,a FDE were used as the prediction error evaluation indicators.The experimental results show that the four indicators are 19.05,32.77,32.57 and 61.25 respectively.Compared with the Semantics-STGCNN algorithm,the prediction performance is improved.The accuracy of trajectory prediction is further improved,and the real-time requirement is guaranteed.