Research On Scenario Oriented Traffic Vehicle Motion Prediction Methods

Autonomous vehicles must engage and cooperate effectively with surrounding traffic vehicles to achieve safe and efficient driving.Despite significant advancements in autonomous driving technology,challenges still exist in complex traffic scenarios such as lane changing,merging,roundabouts,and unsignalized intersections.To navigate these scenarios safely,autonomous vehicles need to anticipate the motion states of other vehicles in advance.Different traffic scenarios exhibit distinct road structure characteristics and environmental interaction patterns,necessitating the use of appropriate motion prediction methods tailored to each scenario.Accurately and reliably predicting vehicle motion states using suitable prediction methods enables the realization of safe and dependable autonomous driving.Based on a comprehensive analysis of the road structure features and environmental interaction characteristics of common driving scenarios,this dissertation conducts research on scenario oriented traffic vehicle motion prediction methods for typical driving scenes.A vehicle behavior prediction framework based on game theory and hidden markov model is proposed to solve the unity problem of vehicle behavior recognition and intention prediction on highways.A data-driven multimodal vehicle behavior and trajectory prediction model is constructed to address the integrity issue of vehicle behavior prediction and trajectory prediction on highways.By utilizing road structure information in different types of main roads in urban areas and designing different types of attention mechanisms,accurate and reasonable vehicle trajectory prediction on urban main roads is achieved.An interactive object selection method suitable for different types of dense and heterogeneous traffic scenes in urban areas is designed,and a trajectory prediction model that meets vehicle dynamic constraints is established.A dense and heterogeneous traffic dataset without traffic signals in urban areas is constructed to verify the the vehicle trajectory prediction model’ performance.The main research content of this dissertation is as follows:Firstly,this dissertation proposes a vehicle behavior prediction framework based on game theory and hidden markov model(HMM)to address the problem of vehicle behavior prediction on highways.Considering the inertia of driver behavior,i.e.,the tendency of vehicles to maintain their current behavior,a vehicle behavior recognition model based on Gaussian HMM is designed.The forward-backward algorithm is used to solve the vehicle behavior recognition model.The parameters of the vehicle behavior recognition model are estimated using the NGSIM dataset and the expectation-maximization algorithm.A vehicle intent prediction model based on game theory is constructed,and a vehicle trajectory reward function is designed and calibrated.The proposed algorithm is validated using a testbed based on datasets and simulation environments to predict vehicle behavior on highways.Next,in purpose to address the completeness issue of vehicle behavior prediction and trajectory prediction in highway scenarios,a multi-modal vehicle behavior and trajectory prediction algorithm for highway scenarios is proposed,which can predict the multi-modal behaviors of vehicles in highway scenarios and the corresponding trajectories under each behavior.By analyzing the road structure features of highways and the behavioral characteristics of vehicles in highway scenarios,a method for selecting interactive objects in highway scenarios is designed,with the predicted vehicle as the center.To more fully extract the motion features of each vehicle in the scene,a vehicle trajectory encoder based on Bi GRU is designed.To extract richer interactive features in the scene,a dilated convolutional social layer is constructed.Comparative experiments and ablation experiments based on the NGSIM dataset and the High D dataset are conducted to verify and analyze the behavioral prediction performance and trajectory prediction performance of the model.Furthermore,a road structure feature extraction method based on Vector Net is established for main roads in different types of urban areas.Three different vehicle and road feature input methods are designed to investigate the impact on the performance of the trajectory prediction model.To model the influence of different road structure features on vehicle motion in the scene,a map information aggregation method based on multi-head attention mechanism is designed.A graph-structured vehicle interaction feature extraction model is constructed,which models the impact of other dynamic vehicles in the scene on the target vehicle’s motion trajectory using various attention mechanisms.Based on the interaction features between the target vehicle and other vehicles,the target vehicle’s map-weighted features,and the target vehicle’s motion features,a trajectory prediction module based on Conditional Variational Auto Encoder(CVAE)is constructed.Comparative experiments and ablation experiments are conducted on the Argoverse Motion Prediction dataset and the INTERACTION dataset to verify and analyze the effectiveness of the trajectory prediction model and some key modules in the model.Finally,a vehicle trajectory prediction model is established for dense and heterogeneous traffic in urban areas.To make the prediction model applicable to different types of dense and heterogeneous traffic scenarios,a graph-based interaction feature selection method is designed.A category layer is constructed to extract the motion patterns of different types of traffic participants.To extract spatial-temporal interaction features between interest traffic participants and the target vehicle,a graph convolution model is established based on graph operation layers and temporal convolution layers.The dynamic characteristics of the target vehicle are constructed and different types of trajectory prediction models are explored and designed.The HID dataset,consisting of dense and heterogeneous traffic without traffic signals in urban areas,is constructed,and comparative experiments and ablation experiments are conducted on multiple different datasets to validate and analyze the effectiveness of the model.