3F LIDAR-based Dynamic Vehicle Detection And Tracking

Object detection and tracking is one of the most essential problems in mobile robotic-s.As an important component of object detection and tracking,dynamic vehicle detec-tion and tracking also plays a key role in improving the environmental perception for an Autonomous Land Vehicle(ALV).Under the background of autonomous navigation in urban environments for an ALV,this dissertation focuses on the issue of dynamic vehi-cle detection and tracking with a three-dimensional LIDAR.The main contributions and innovations of this dissertation are as follows:Firstly,two novel ground segmentation algorithms were proposed for segmenting three-dimensional scans of various terrains.One is the ground segmentation algorithm based on the sparse Gaussian Process Regression(GPR)and the other is the segmented real-time ground segmentation algorithm based on the recursive GPR.The ground seg-mentation algorithm based on the sparse GPR constructs a three-dimensional grid map which will be partitioned into many cells,and then exploits a two-dimensional GPR with sparse covariance function to model the ground in the whole grid map with the lowest 3D points in these cells directly.This algorithm achieves 97.90%precision on the pub-lic Boston dataset.The segmented real-time ground segmentation algorithm based on the recursive GPR splits the complex,large-scale two-dimensional ground segmentation problem in the three-dimensional grid map into many one-dimensional regression prob-lems with lower complexity in the polar grid map.For each segment of the polar grid map,a one-dimensional recursive GPR with the non-stationary covariance function is used to model the local ground in it.The accuracy of the ground segmentation algorithm based on the recursive GPR is 97.67%on the same Boston dataset.Furthermore,this algorithm can also get real-time performance,which is essential for an ALV.Secondly,a novel curb detection algorithm based on the iterative GPR was proposed to obtain the Region Of Interest(ROI)for an ALV.The curb detection algorithm takes each scan line of the three-dimensional LIDAR as a processing unite and extracts feature points from the individual scan lines directly.Then an iterative GPR is utilized to repre-sent both straight-line and curved curb models.The proposed curb detection algorithm based on the iterative GPR can detect curbs up to 50 meters away while meeting the ac-curacy requirement.Moreover,in order to evaluate the performance of the proposed curb detection algorithm quantitatively,we have labelled some curbs from some Velodyne L-IDAR scans manually to form a new public dataset,and the detection accuracies of the proposed curb detection algorithm based on the iterative in the aspects of the left and right curbs on the dataset reach 78.74%and 81.96%,respectively.Thirdly,a novel Global Cylindrical Coordinate Histogram Descriptor(GCCHD)was proposed to recognize vehicles in urban environments.This descriptor is created for the centre point of each object in the ROI.In order to achieve independence with respect to the rotation around z axis,a global reference frame is introduced in GCCHD,and all the three-dimensional points in the cylindrical support region are projected into the three-dimensional histogram according to their three cylindrical coordinates.Experiments on the Sydney Urban Object dateset and the dateset we prepare and label manually verify the performance of the GCCHD in vehicle recognition.Fourthly,a novel dynamic vehicle detection and tracking algorithm based on the like-lihood field model was proposed.The algorithm exploits the novel likelihood-field-based vehicle measurement model coupled with our newly modified Scaling Series algorithm to estimate the poses of the vehicles in the ROI.In the phase of dynamic vehicle detection,the three-dimensional LIDAR data representation based on the two-dimensional virtual scan is improved to detect the dynamic vehicles,even they are occluded by other objects in the xy plane.For vehicle tracking,a novel varying extent vehicle tracking algorithm based on the Bayesian filter was proposed.Because of the anchor point,this tracking al-gorithm not only can update the pose and the velocity of the target vehicle in the dynamic background scenes,but also can estimate its extent according to the associated measure-ments in the tracking process.Both the quantitative and qualitative experimental results validate the performance of our novel dynamic vehicle detection and tracking algorithm on the KITTI datasets and the Velodyne LIDAR scans collected by our ALV in dynamic urban environments.All the above algorithms have been successfully used in our ALV,which won the third place in the seventh Chinese Future Challenge in 2015.