Research On Complex Traffic Environment Perception Based On LiDAR And Camera Vision Fusion

The target detection technology based on sensor is an important prerequisite for achieving autonomous driving environment perception.Whether the target detection is realtime and accurate under vehicle operation is directly related to the correctness of decision and results of control.Therefore,in the face of complex and various traffic environment,how to use vehicle sensors to identify targets and achieve vehicle-level detection requirements is still a major difficulty in current environmental perception research.To solve the problem of poor portability of the algorithm due to the limitation of specific scenes in the current target detection research,we use the information data obtained by Li DAR and camera in a variety of complex traffic scenarios to study the single sensor and fusion target detection method.The main research contents are as follows:(1)By comparing and analyzing the performance characteristics of various vehicle sensors,we choose Li DAR and camera to build a set of autonomous driving data acquisition and target detection system.Furthermore,we propose a complex traffic environment sensing scheme based on point cloud and visual information fusion.(2)Aiming at the problem of low accuracy of Li DAR extraction of structured road boundary based on single feature information of point cloud,we propose a multi-feature window search algorithm based on 32-line Li DAR to detect road boundary.Firstly,we dynamically collect point cloud data which is acquired by the Li DAR scan based on the theory of safe vehicle distance control and the foreground point cloud is extracted.Then,the geometric characteristics and spatial distribution characteristics of the point cloud are used to search the boundary candidate points at different scales.Finally,the boundary candidate points are fitted to realize the extraction of road boundaries,and the algorithm is verified by road scenarios in a variety of complex situations.(3)To address the shortcomings of traditional point cloud clustering methods that still require manual intervention to set parameters,we propose an adaptive 3D spatial clustering algorithm.Firstly,the clustering radius of different distance targets are dynamically determined based on the relationship between the scanning distance and the point cloud density.Subsequently,a three-dimensional virtual segmentation network is established,and the three-dimensional data is divided into multiple equal volume cubes.Finally,the target feature points are determined according to the point cloud density characteristics in the cube,thereby completing the target clustering.(4)Aiming at the problem of lack of classification of training sample scenes of current deep detection-based visual detection algorithms,we collect sample data of vehicles and pedestrians in different weather conditions and complex driving conditions.Meanwhile,we innovatively introduce night-time infrared data into the sample set,and modify the classifier in YOLOv3 to a Softmax classifier,as well as the algorithm performs better in multi-class detection of targets in complex traffic environment.Moreover,in order to prevent the loss of the target during the detection process,the target matching function is used to complete the accurate tracking of the detection object.(5)In order to further improve the accuracy and stability of target detection,we study the target detection algorithm based on the fusion of Li DAR and visual information.Firstly,the Li DAR and camera are synchronized in time and space through joint calibration,and then the D-S evidence theory fusion algorithm is used to match the reliability values of the detection results of different sensors.Through a variety of complex traffic scene verifications,the fusion detection result is more accurate and stable than the detection result of a single sensor.