Vehicle-Infrastructure Cooperative Perception Based Intelligent Vehicle Driving Map Construction

Intelligent and Connected Vehicle(ICV)can effectively reduce road traffic safety accidents and improve road traffic efficiency.As the core technology of ICV,traffic environment perception is the basis for decision-making.However,the range of single roadside perception and on-board perception is limited,while Cooperative Vehicle Infrastructure System(CVIS)technology can provide a larger and more comprehensive information of road trafific environment.In CVIS,intelligent roadside perception and vehicle on-board perception are both very important means of environmental perception.However,the road traffic environment often is complex and diverse.How to collect,process,fuse and construct the traffic environment according to the actual needs to realize the driving environment perception and representation is the key research of intelligent vehicle technology.Therefore,this paper addresses the needs of CVIS perception,and carries out researches in three aspects: roadside cross-field object perception,object fusion perception of vehicle sensors and construction of driving maps based on vehicle-road perception information.The main research contents are as follows:(1)Vision-based roadside cross-field object perception method.To address the problem of cross-roadside object rerecognition issue,this paper constructs a visionbased roadside cross-field object perception method,which first uses the object detection and tracking algorithm to perceive single-roadside environment information,and then proposes an improved object re-identification algorithm to extract and measure the features of cross-field object information to complete the matching of cross-field targets.To further improve the discrimination ability of the Re-ID model,an adaptive label smoothing method is proposed,which can effectively alleviate the overfitting problem of the Re-ID model,improve its accuracy and robustness.Tested on the Market-1501 and Duke MTMC-Re ID datasets,the proposed method achieves87.9% and 94.8% in rank-1 accuracy,respectively,which are highest among other advanced algorithms.The experimental results demonstrate that the proposed method can effectively improve the accuracy and robustness of the re-identification model and achieve the re-identification of cross-field targets.(2)Object fusion perception method based on on-board sensors.In order to improve the sensing capability of intelligent vehicles,this paper proposes a road environment sensing method based on Li DAR and camera fusion.To efficiently fuse camera and Li DAR sensing data,this paper uses a decision-level multi-sensor information fusion method,constructs a sparse tensor based on semantic information consistency and geometric information consistency constraints on the LIDAR and camera object detection results.Then the information fusion model is built using convolutional neural networks,and the information fusion is performed on the sparse tensor.Finally,the fusion results are output using Max Pooling.The proposed fusion algorithm is trained and tested on the KITTI dataset,and achieves detection accuracy of 95.52% and 82.76% on Easy and Moderate difficulty of vehicle targets,respectively,which is very competitive compared with other algorithms.The proposed method can improve the intelligent vehicle perception accuracy.(3)Vehicle-road perception fusion-driven driving map construction.Based on the roadside and on-board traffic information perception results,this paper proposes an occupancy grid map-based driving map construction method to fuse and construct vehicle-road perception information.A 2D occupancy grid map-based driving map model is established,and the D-S evidence theory is appliedto fuse on-board perception information and roadside perception information,and evaluate the state of each cell in the constructed driving map,and obtain the confidence levels of "Free","Occupied",and "Dangerous" for each cell.The state corresponding to the maximum value is the final state of the cell.This paper proposes a D-S evidence theory information fusion method based on the importance of data sources,which can process some conflicting evidence to obtain more reliable fusion results.Finally,this paper conducts experiments on the fusion algorithm and driving map.The results demonstrate the effectiveness of the driving map construction method proposed in the paper,which can provide a basis for ICV’s decision-making.The researches on roadside cross-view object perception,vehicle sensor fusion perception,vehicle-road information fusion and driving map construction methods in this paper can improve the development of cooperative vehicle and roadside perception and autonomous driving technology.