Dynamic Object Detection And Tracking With Vehicle-mounted Multi-sensor Fusion

Multi-sensor information fusion technique of vehicle is a hot issue of autonomous driving research area,reliable environment perception information is a necessary condition for the path planning of autonomous vehicles.In this paper,to combine the precise range information of lidar and high resolution of camera,we studied the dense depth image methods of how the depth information of lidar and the high-resolution image works.And based on the range data in dense depth image and color data in RGB image,combining 3D scene flow to detect and track dynamic target.the work and contributions of the thesis includes:1?Considering the information acquisition way of lidar and analyzing the egomotion in a single frame of lidar,we accomplished the data compensation of lidar in two moving components of translation and rotation transform.2?In the study of sensor information fusion,we generate a new approach about dense depth image.By project the point cloud of lidar into the image,combine the depth information of sparse point cloud and color information of dense image to build KD feature space in the Bilateral Filter framework,and get the dense depth image by ANN way we build MRF framework in the dense depth image next.Under the optimization of MRF framework,we further smooth the weight value of color channel and distance channel in Bilateral Filter to get the new dense depth value.3?In the case of computing the ego-motion of vehicle,as it guarantee the efficient enough.This paper designs the bucketing of image and combine the dense depth image and RGB image to segment the ground static feature points to improve the robust of egomotion of vehicle.To construct the 3D sparse scene flow model in the RGB image and dense depth image which generated by the lidar,use Delaunay triangulation to connect adjacent feature points,and combine the dense range information from lidar to remove feature points.By classify feature point to detect and track the dynamic targets,experiments on the real road scene demonstrate the effectiveness of the proposed algorithm.