Road Vehicle 3D Detection And Tracking Based On RGBD Information

With the rapid development of artificial intelligence technology,smart cars with ADAS and driverless technology have become one of the key directions of future automobile developing.As one of its key technologies,object detection and tracking has always been a research hotspot in this field.The vehicle's target detection and tracking is a technology that finds targets in the iamge captured by cameras and extracts its motion trajectory to perceive the surrounding environment.As the most imaportant sensing module of the smart car,its results directly affect the follow-up decision-making and planning process,therefore,it's of great research significance and application prospects to propose state of art detection and tracking methods.However,due to the complex real-life traffic scenarios such as mutual occlusion,uneven illumination,truncated targets,and changes resulting by different viewing angles,problems like misdetection,missed detection,mistracking,and redundant tracking are caused.This has largely constrained the development of unmanned technology and seriously affected the process of smart cars from concept to practical.In view of the above problems,this paper has made full study of the existing object detection and tracking algorithms and proposed a series of improvement methods for its defects and verified the feasibility of the improvement through experiments.The main research contents and contributions of this thesis are as follows:Firstly,we fully studied a kind of object detection algorithm based on 3D object proposals and verified its performance through numerous comparative experiments.For its defect that the 3D proposals are not fully utilized,we combine 2D detection box and3 D proposals to generate 3D detection box from 2D bounding box.Secondly,a multi-target tracking method based on 3D triangulation is deeply studied.Through a large number of experiments,we found that the output 3D bounding boxes have great dimensional error causing by its simple use of scene flow cluster method.We then proposed a post-processing method nased on K-means and depth-first to solve the defect.To achieve better results,we also adjusted the original algorithm framework by adding a object detection module handling obeject detection and the scene flow cluster module processing ego motion instead.Finally,We proposed a multi-target tracking algorithm based on 2D-3D combinedfeatures,which has expand the original MDP tracking from image space to world space.The observation set combining 2D and 3D information is obtained through 3D object detection algorithm.During the objects associaion part,the 2D-3D observation set is used to extract 2D and 3D features separately for more accurate results.Considering the advantages that binocular camera inter-frame images are more than monocular ones,we structured a new feature descriptor called multi-image optical flow replacing the original multi-aspect optical flow.According to the experiments,the improved MDP tracking method outperforms the original MDP and has advantages over other state-of-art tracking methods.