3D Multiple Object Tracking And Trajectory Prediction For Unmanned Vehicles

The safety of unmanned driving has always been the focus of public attention.3D multi-target tracking technology and trajectory prediction technology are prerequisites for unmanned vehicles to detect potential collisions in advance and avoid planning jumps.And they are important for safe operation.Due to the diverse types of traffic participants in the actual driving environment and the complex and changeable road structure,accurate and stable 3D multi-target tracking and real-time and accurate trajectory prediction have become research hotspots.This thesis has launched related researches.First,the problem of 3D multi-target tracking based on Lidar is studied.Aiming at the insufficiency of the single detection and filtering threshold in the multi-target tracking benchmark algorithm,the limitation of the target matching metric,and the shortcomings of the original survival management strategy,a 3D multi-target tracking algorithm combining automatic detection and filtering and dynamic survival management is proposed.The automatic detection filter network is designed to make up for the shortcomings of the single threshold filtering in the benchmark method and reduce the false detection.Using 3D GIOU,which is devoted to non-overlapping areas,the problem of not matching when 3D IOU is 0 is solved.The trajectory score update method is analyzed,and a dynamic survival management mechanism is designed accordingly,which reduces the trajectory misdetection and target loss.Through ablation experiments on the KITTI multi-target tracking dataset,the effectiveness of the above modules and the overall effectiveness of the algorithm are verified.Second,the issue of trajectory prediction is investigated.Following most algorithmic construction idea,we select a prediction architecture based on recurrent neural networks,and choose the classic method Social-GAN as the benchmark.By considering the limitation of current algorithms that are difficult to migrate to unmanned driving scenarios,an algorithm that fully involves map features and is suitable for trajectory prediction of multiple types of traffic participants is proposed.As for the huge difference in different types of interactions,a heterogeneous interaction feature extraction module is designed to realize the interaction modeling between different types of traffic participants.Taking into account the impact of road structure on unmanned driving and surrounding traffic participants,a map feature extraction module is suggested to achieve the extraction of lane line features and driveable area features.Through ablation experiments on the pre-processed unmanned driving dataset Argoverse,the effectiveness of the above modules and the overall effectiveness of the algorithm are verified.Finally,to overcome the defects of the pooling layer modeling in current trajectory prediction architecture that is not intuitive and prone to information leakage,the large amount of recurrent neural network parameters and the limitations of long inference time,the architecture based on graph convolution is selected.And we choose the classic method Social-STGCNN as the benchmark.The adjacency matrix is modified according to the blind area of the traffic participants’ visual field,and the adjacency matrix with category information is constructed to affect the extraction of spatiotemporal interaction features.The map feature extraction module in the previous chapter is utilized to obtain map features and combine them with spatio-temporal interactive features,and jointly determine the characteristics of the future trajectory.Through experiments on the dataset produced in the previous chapter,the advantages of the algorithm in prediction accuracy,parameter amount and inference time are verified.