Research On Autonomous Vehicle Trajectory Prediction Method Incorporating Surrounding Vehicle And Road Environment

With the deepening of vehicle intelligence,the research of various key technologies for autonomous driving systems is receiving attention from governments and many researchers around the world.Vehicle trajectory prediction,as an important foundation for autonomous driving decision control,aims to enable intelligent vehicles to anticipate the future motion of surrounding vehicles for safer,more efficient,comfortable and green transportation.The current vehicle trajectory prediction faces the bottleneck and challenge of complex traffic scenarios that are difficult to model effectively.In order to enable intelligent vehicles to find general patterns from complex traffic scenes,this thesis proposes a vehicle trajectory prediction method that incorporates the surrounding vehicle and road environments for the vehicle trajectory prediction problem in autonomous driving scenarios.The method uses Long-Short Term Memory(LSTM)and encoder-decoder architecture as the main body,and models the driving scenario by using vehicle-vehicle relationship spatio-temporal graph and vehicle-road relationship mapping.The advancedness of the model is also verified on the Next Generation Simulation(NGSIM)public dataset,which improves the long-term accuracy of trajectory prediction and can better support the decision planning of autonomous driving.The main research work of this thesis is as follows:(1)A spatio-temporal graph structure-based inter-vehicle interaction model is proposed.The model extracts the social interaction information between vehicles in the spatial dimension and the physical motion information of the target vehicle in the temporal dimension,which solves the problem that the instantaneous motion of the target vehicle cannot be accessed instantaneously.Thanks to the enhancement of the spatio-temporal graph structure to the original trajectory data,both local and overall features of vehicle motion in dynamic traffic scenarios can be effectively extracted by the model.(2)A serializable vehicle-road relationship mapping method is proposed for the autonomous driving scenario of highways.The method fuses discrete arbitrary vehicle trajectory points and highway shapes from geographic information system(GIS),and establishes a sequence of vehicle-road relationships through pre-processed spatial location calculation.The serialized vehicle-road relationship not only preserves the spatio-temporal dependence between vehicles and roads,but also expresses explicitly the vehicle maneuvering interaction behaviors along the road and perpendicular road directions,making the impact of roads on vehicles and the traffic maneuvering behaviors of target vehicles more closely related.(3)A new two-layer LSTM environment encoder module is proposed and applied to VRR-Net,a trajectory prediction model based on the spatio-temporal graph of vehicle-road relationship.The environment encoder module extracts environmental features from both social interactions and road constraints.VRR-Net accepts the spatio-temporal graph of vehicle-vehicle relationship and vehicle-road relationship as input,and after extracting features using the environment encoder,it generates by maneuver type future trajectory of the target vehicle.All the above works are implemented on the US-101 and I-80 road datasets of NGSIM.The comparison experiments and ablation experiments demonstrate that the vehicle-vehicle relationship spatio-temporal graph can fully exploit the contribution of historical trajectories,the vehicle-road relationship mapping can effectively fuse the influence of road constraints,and VRR-Net achieves comparable or better performance than the state-of-the-art methods on the NGSIM dataset.