Research On Multimodal Trajectory Prediction Method Of Intelligent Vehicle Based On Deep Inverse Reinforcement Learning

In the context of the strategy of "strong transportation and prosperous industry" advocated by the state,the rise of artificial intelligence has boosted the development of the field of autonomous driving,and the development opportunities and challenges of autonomous driving coexist,with safety issues being a major challenge all the time.As an important component of autonomous driving perception,trajectory prediction plays a crucial role in subsequent safety decision control,and applying deep reinforcement learning to the vehicle trajectory problem can improve its prediction accuracy.To further improve the safety of assisted driving by making full use of complex traffic scene information,a study on multimodal trajectory prediction of intelligent vehicles based on deep inverse reinforcement learning is carried out,which mainly includes the following:1.A network based on deep inverse reinforcement learning GIRL-ATF is proposed.The network architecture consists of two parts,including a strategy generation subnet Goal-inferred Maximum Entropy Inverse Reinforcement Learning(Ginf-MaxEnt IRL)and a trajectory fusion subnet Attention-based Coding and decoding Trajectory Fusion(Att-Codec TF).The Ginf-MaxEnt IRL subnet maps the traffic scene information into the reward function of the two-dimensional grid by inverse reinforcement learning based on the maximum entropy policy,and optimally iterates the optimal policy;the Att-Codec TF subnet generates multiple paths to different targets on the two-dimensional grid based on policy sampling,and predicts the future trajectory by coding and decoding.2.In the Ginf-MaxEnt IRL sub-network,to solve the problem that the original network cannot fully extract the traffic scene information features,an improved Fused Dilated Convolution(FDC)module is introduced,which combines the dilation convolution with the point-by-point convolution,and then fuses it with the original residual module,which can lighten the existing feature extraction It can reduce the loss of edge information in traffic scenes by expanding the perceptual field,thus making a better trade-off between network lightweighting and prediction accuracy,while taking into account the comparable or even higher generalizability of the original network.In the process of inverse reinforcement learning,the policy generation is closely related to the predefined target state,which is often unknown and needs to be inferred artificially.Therefore,introducing the Goal-inferred Maximum Entropy(Ginf-MaxEnt)inverse reinforcement learning strategy with target inference into the existing trajectory prediction network can help reduce the need for predefined goal states.It alleviates the defect of strong subjectivity in policy generation due to manual setting in the original process,and helps to generate a more objective and accurate policy.3.In the Att-Codec TF subnetwork,in order to solve the problem that the feature information of the original network is still not sufficiently utilized,an attention mechanism is introduced to encode the fusion of motion features and scene features,which helps to further improve the accuracy of the trajectory prediction results in the next stage by considering the scene context information comprehensively.A reasonable and effective loss function is crucial,which not only affects the optimization process of the whole network,but also determines whether the feature information is effectively utilized.Therefore,an improved correction factor is added to the original loss function,which can encourage the generation of diverse trajectories by penalizing pairwise distances with small differences,which is more in line with the multimodal nature of future trajectories of smart vehicles.The ablation experiments and simulations were validated on the public datasets Stanford Drone and nuScenes,showing that the introduced component modules facilitate the overall network performance and have high prediction accuracy compared with other existing methods.On the Stanford Drone dataset,the average displacement error and final displacement error at K=20 are reduced by 0.18% and 7.42%,respectively,compared with the advanced HBA-flow method.On the nuScenes dataset,the average displacement errors at K=5 and K=10 are reduced by 38.93% and 28.02%,respectively,compared with the classical method MTP;and by 17.67%,8.87% and 8.58% and 12.40%,respectively,compared with the two advanced methods MHA-JAM and Cxx.In summary,the proposed improved network GIRL-ATF takes into account the multimodal characteristics of the future behavior and trajectories of smart vehicles and is able to generate diverse trajectory prediction results that are more consistent with the scene structure.