Research On Calibration Algorithm And Vehicle Detection Application Based On Stereo Camera And LiDAR

Autonomous driving is one of the most popular technologies at present,and it has received great attention from academia and industry.Sensor fusion algorithms enable autonomous vehicles to have a more comprehensive and accurate understanding of their surroundings,enabling better decisions and safer driving.For sensor fusion algorithm research,we have studied two sub-fields in this thesis,which are extrinsic calibration of sensors and object detection.At present,in the research of extrinsic calibration of sensors,stereo camera calibration and camera-LiDAR calibration are separate processes,which will lead to accumulation of calibration errors.In the research of object detection with fusion of camera and LiDAR,the algorithms based on feature layer fusion have complex network structure but the output results are not good.In view of the above-mentioned shortcomings,this thesis proposes a calibration method and a vehicle detection algorithm based on stereo camera and LiDAR.Our contribution is summarized as follows:Firstly,we collect the calibration data including stereo images and LiDAR point clouds with our vehicle-mounted experimental platform.And we used several calibration toolboxes to conduct stereo calibration and camera-LiDAR calibration respectively.In the analysis of the calibration results,we propose that using the closed-loop pose to evaluate the calibration results.Then,we develop a joint extrinsic calibration algorithm of stereo camera and LiDAR which satisfies the pose closed-loop constrainst.The algorithm combines the process of stereo extrinsic calibration and camera-LiDAR extrinsic calibration by using the pose closed-loop constraint,and combines point and plane features to improve the calibration accuracy.For camera-LiDAR calibration,our algorithm has a positional accuracy of within 2.3cm and an angular accuracy of within 0.37 degrees on the simulation dataset.In the real vehicle calibration experiment,our algorithm also has smaller error and faster speed than the Matlab LiDARCamera Calibrator.Finally,we propose a vehicle detection algorithm,which fuses the stereo images and cloud data in decision level.With the sensor calibration parameters,the algorithm uses handcrafted features to characterize the geometric and semantic consistency of a set of 2D object detection results from stereo images and 3D object detection results from point cloud,and then obtains higher-precision 3D object detection results through a last-fusion network.For KITTI validation dataset,our algorithm has 4.62% improvement and 1.93%improvement under moderate difficulty compared to the original SECOND network,and the improvements are more significant in the distances range above 20 m.