Research On Dynamic Obstacles Detection And Recognition Of Driverless Vehicles Based On Lidar

In recent years,with the development of artificial intelligence technology and improving level of requirements for travel safety and convenience,driverless technology has become a research hotspot in universities and enterprises.Although the current unmanned driving technology has made great progress,it is still a long way from large-scale driverless product going to the market.Environmental awareness has always been the key to driverless technology,and it is also the bottleneck for driverless vehicles to break through in terms of safety.Because the camera is obviously affected by the light when it perceives,and it has limitations when it is used in the extreme light environment such as night,lidar is selected to realize the detection and recognition of obstacles.The driving environment of high-speed road is simple,the types of obstacles are single,and the driving speed of vehicles is fast.A real-time dynamic obstacle detection algorithm based on point cloud clustering is established for high-speed road condition.The point cloud is voxelized and filtered to realize the subsampling.Through the experiment,the effect of different data filtering algorithms is compared,and ray algorithm is selected to filter the ground point cloud.Aiming at the problem that the ray algorithm is sensitive to the running slope,the slope gradient composed of the ray data from the two adjacent point clouds in front of the target point cloud is calculated to replace the slope threshold,which makes the ray algorithm no longer sensitive to the slope of the ground and improves the robustness of the ground point filtering process.The simulation time and contour coefficient of the same data by different clustering methods are compared,and Euclidean algorithm is chosen to cluster the data after filtering the ground point cloud.Aiming at the problem of uneven density of point cloud,the distance threshold of point cloud clustering is designed as a linear function of distance,which improves the robustness of clustering algorithm.The driving environment in urban road condition is complex,and obstacles are of all kinds.Merely detection of obstacles in the driving process can not meet the needs of perception,so it is necessary to obtain the semantic information of driving environment.The traditional method is difficult to extract the features of point cloud,so a deep learning target detection network based on lidar data is established to realize the recognition of dynamic obstacles under urban road conditions.According to the characteristics of lidar data,the point cloud is voxelized,the input feature vector of lidar data is established,and the features of point cloud are transformed into voxel features through the preprocessing of feature data.In order to solve the problem of large computation and sparse features of 3D point cloud,the feature of voxel is maximized in Z coordinate direction.The feature of voxel is further transformed into the feature of 2D grid map,and the 2D feature tensor of point cloud convolution is established.It not only avoids the influence of three-dimensional convolution on convolution efficiency,but also eliminates a large number of empty voxels,which makes the feature information more abundant.The feature extraction network of point cloud is established,and the features of different convolution layers are united by deconvolution operation,then the feature dimensions are spliced.The detection head of single-stage target detection is used for classification and regression,and four prior boxes are designed on each pixel of the feature map.In the process of training,each feature map extracts 200 positive samples and 600 negative samples through IOU value of calibration boxes to prevent excessive imbalance of positive and negative samples.Finally,the prior boxes of each sample are classified and regressed,and repeated targets are filtered by non maximum suppression method.In the system of ROS based on Ubuntu,the algorithm of obstacle detection and recognition established in this paper is verified by experiments.In the clustering detection algorithm,the interface of receiving radar topic is established,and the clustering algorithm processes the original point cloud data,and publishes the topic of obstacle information for the planning layer.The experimental results show that the Euclidean algorithm with improved distance threshold has better clustering effect and can detect dynamic obstacles in real time.The point cloud depth learning algorithm is trained,and the accuracy rate of the algorithm for cars,pedestrians and bicycles in the Kitti data set is 59.21%,0.96 percentage points higher than Voxelnet,and the detection efficiency is improved from 0.23s/frame to 0.09s/frame.Real-time and accurate information of obstacles location and size for driverless vehicles in high-speed road conditions are available in this research.Also,semantic information of obstacles in complex environment in urban road conditions can be identified.Lidar data is not sensitive to light conditions,which can make up for the limitations of camera in the field of perception.Therefore,the research results of this paper are of great significance.