| In recnt years,with the rapid development of information fusion and artificial intelligence technology,the intelligence level of Autonomous Vehicle(AV)is also improved.The application of AVs is gradually penetrating from the relatively simple traffic environment,such as closed parks and highways to the more complex urban environment.Given the complex urban road conditions,mixed traffic between peope and vehicles and large number of GPS-blind areas,many dynamic objects are detected around the AV,and occlusion is easily formed.All of these factors bring considerable challenges to the high-precision environmental perception of AVs.To solve these problem,on the basis of the point cloud and image data collected by multi-source sensors,such as three-dimensional light detection and ranging(3D LIDAR)and RGB camera,in depth research on key issues such as ground point cloud segmentation,road boundary detection and object recognition,of AV driving in the urban environment is conducted.The specific research contents are as follows:1.Designing the driving environmental perception system framework.Through in-depth analysis of the requirements for the safe navigation of AVs in the urban environment,with Chang'an University's ‘Xinda' AV as the research platform,the driving environmental perception system framework based on multi-source sensor information fusion is proposed.The framework is based on on-board LIDAR,millimetre-wave radar and camera.Moreover,the V2 X vehicle-road collaborative perception terminal is used to enhance the capability of the AV to sense the super-view distance and share information and to realise the goal of all-round perception of the remote,medium and near driving environments of AV.2.Research on the real-time robustness of ground point cloud segmentation algorithm based on 3D LIDAR.In view of the undulations of road elevation in the urban environment,two robust 3D LIDAR ground ground point cloud segmentation methods are proposed.Firstly,on the basis of the analysis of the distribution characteristics of 3D LIDAR point cloud,an adaptive polar grid map is constructed and the data of LIDAR point cloud are mapped onto the grid.The polar grid is classified into ground,obstacle and boundary grids according to the height distribution characteristics of LIDAR point cloud in the grid.The points in the ground grid are marked as the ground points,and the ground points in the boundary grid are devided using the local continuity of the ground height variation.The experimental results show that the average processing time of each frame is only 17 ms,which can realise the real-time segmentation of relatively flat road,and the segmentation accuracy is more than 95%.Secondly,for the pavement with large elevation fluctuation,on the basis of the in-depth analysis of the characteristics of the radial distance change between adjacent scan lines and LIDAR points at the same azimuth of the ground and the non-ground parts,a ground point cloud segmentation method based on the analysis of the radial distance between adjacent scan lines is proposed.The disorder points are serialised,and then the expected distance between two points of the adjacent scan lines at the same azimuth is calculated according to the 3D LIDAR scanning parameters and vehicle attitude.The ground point is segmented on the basis of the change relationship between actual distance and expected distances.The proposed method is tested using the typical undulating road in Chongqing.This method can segment the undulating road surface robustly.The processing time of each frame is approximately 34 ms,which meets the real-time,precision and robustness requirements for the AV algorithm.3.Research on robust detection algorithms for urban road boundary based on 3D LIDAR.To precisely extract the irregular road boundaries or those blocked by obstructions on the road from the 3D LIDAR data,a novel and robust road boundary detection algorithm based on 3D LIDAR is proposed.Firstly,the ground point cloud segmentation method based on adaptive grid was used to quickly separate the ground part from the non-ground part of the scene,which not only reduces the number of 3D LIDAR points to be processed,but also reduce the interference of road obstacles.The candidate points of the road boundaries are extracted from the ground point cloud by using the maximum and minimum height difference,the angle between adjacent points of the same scan line and the gradient feature between two points at the same azimuth of the adjacent scan line.The position of the boundary points is predicted according to change trend of the road shape.The non-road boundary feature points are filtered,and the boundary feature points blocked by occlusion are regarded as complements.Finally,the remaining candidate boundary points are fitted to the road boundary by the spline model.The results show that the average processing time of each frame is about 36.5 ms and the average accuracy of the algorithm is approximately 93% and,which can robustly detect various typical road boundaries in real time even if the boundary is partially blocked by obstacles on the road.4.Research on object detection algorithm based on 3D LIDAR and image data fusion.3D LIDAR sensors can directly obtain the position and geometric structure of an object within its detection range,whereas the use of vision cameras is most suitable for object recognition.Accordingly,in this paper,we present a novel object detection and identification method that fuses the complementary information obtained by two types of sensors.First,we utilise 3D LIDAR data to generate accurate object-region proposals.Then,these candidates are mapped onto the image space from which regions of interest(ROI)of the proposals are selected and input to a convolutional neural network(CNN)for further object recognition.To precisely identify the sizes of all the objects,we combine the features of the last three layers of the CNN to extract multi-scale features from the ROIs.The evaluation results obtained on the KITTI dataset demonstrate that unlike sliding windows that produce thousands of candidate object-region proposals,3D LIDAR provides less than a hundred of real candidates per frame and the minimal recall rate is better than 95%,which greatly decreases the extraction time.The average processing time for each frame of the proposed method is only 66.79 ms,which meets the real-time demand of autonomous vehicles.While obtaining object depth information,the average identification accuracies of our method for cars and pedestrians at a moderate level of difficulty are 89.04% and 78.18%,respectively.All of the algorithms investigated in this study have been successfully applied in the ‘Xinda' AV,which was independently developed in Chang'an University.The three environmental perception methods proposed in this study are systematically tested using the urban environment data collected by the 3D LIDAR and RGB camera mounted on the ‘Xinda' AV.The qualitative and quantitative experimental results show that the methods proposed in this study are real-time,accurate and reliable,and can enhance the driving environmental perception system of AVs.The AV developed on the basis of the environmental perception method proposed in this study performed well in the ‘World Intelligent Driving Challenge',‘i-Vista Intelligent Driving Challenge' and ‘China Intelligent Vehicle Future Challenge',and won many important awards. |
PDF Full Text Request