Multi-object Behavior Recognition And Trajectory Prediction Around Intelligent Vehicle

As an effective solution to improve traffic safety,traffic congestion,energy shortage and other problems,intelligent vehicles are one of the important strategic directions for the upgrading and development of the global automobile industry.Among them,environment perception is an important basis for realizing automobile intelligence.Behavior recognition and trajectory prediction,as high-level perceptual output,are important preconditions for scientific decisionmaking and safety control of intelligent vehicles.At present,the difficulties in the research of behavior recognition and trajectory prediction lie in the inadequate extraction of vehicle interaction in complex urban traffic scenes,the neglect of static environment constraints,and the difficulty in modeling heterogeneous target coupling relationship.To solve this problem,this paper first studied the extraction of interaction relations in the isomorphic multi-vehicle scene,then extended the research object to more complex heterogeneous multi-targets,and established the hierarchical interaction behavior recognition and trajectory prediction model of heterogeneous targets.The main research work is as follows:(1)A vehicle behavior recognition and trajectory prediction model based on graph convolutional interactive network is proposed.In view of the insufficient extraction of multivehicle interaction in existing studies,a vehicle trajectory prediction model based on lane topology constraints was constructed by using graph convolutional interactive network.On the basis of realizing individual vehicle interaction,the lane topology structure was innovatively obtained from the high-precision map,and the implicit modeling of the interaction between road structure and driving trajectory was realized.Based on the construction of the relationship diagram of vehicle interaction,the expression method of edge weight based on vehicle space distance is improved,so as to capture the influence weight of vehicle close interaction.The model was verified on the public dataset Argoverse.The test showed that the average displacement accuracy of the proposed model was increased by 39.9% compared with the baseline method,and the final displacement accuracy of the multi-action intention output trajectory was increased by 40.4%,which improved the rationality of vehicle trajectory planning under the complex road structure.(2)A heterogeneous multi-object behavior recognition and trajectory prediction model based on hierarchical graph attention is proposed.For complex urban roads containing heterogeneous objects,firstly,a three-level diagram of category-target-lane was constructed innovatively,and the gated cycle unit and graph convolutional network were used to independently encode different types of targets to capture the characteristics of different categories.Secondly,in order to strengthen the edge representation of heterogeneous object interaction graph,the attention mechanism of hierarchy diagram is constructed to obtain the interaction between categories and between objects and the interaction between objects and lanes respectively,so as to realize the efficient interaction and sharing of heterogeneous multi-objects.Finally,based on target track information and regional lane information,a prediction network is constructed to predict multitarget track.In order to evaluate the performance of the model,experiments were conducted on the INTERACTION and Nuscenes data sets.Experiments show that the average error loss and final displacement loss of single-object track output of the proposed model on the Nuscenes dataset are reduced by more than 20%,and the average displacement loss effect of multi-object track output on the INTERACTION dataset is reduced by 2m compared with the baseline method,showing high prediction accuracy in different traffic environments.(3)The results of the proposed prediction models in the isomorphic multi-vehicle scenario and heterogeneous multi-objective scenario are verified respectively in the real vehicle environment.Firstly,the hardware environment required by the model is introduced and deployed.Secondly,the corresponding nodes and logic of trajectory prediction algorithm are designed based on ROS,an intelligent computing platform,and used for data collection and algorithm loading.The test results show that the proposed model still has good reliability and effectiveness in real urban traffic roads.