Research On The Algorithm Of Road Detection And Map Generation Based On Lidar

Autonomous driving is an important branch of artificial intelligence.Nowadays,more and more companies and scientific research institutions are engaged in the research and application of autonomous driving technology.Autonomous driving consists of key technologies of environment perception,position fix,route planning and decisionmaking control.Among them,environment precision is not only the key point of autonomous driving technology but also the first step of the entire autonomous driving technology system.Its performance affects the performance of route planning and decision-making control.Traditional road information extraction solution is using camera and 2-D image processing algorithm to achieve the recognition and extraction of road boundaries information.Its advantage is that it can obtain environmental information in high real time.But the disadvantage is that it is obviously affected by the light intensity changes in the environment.In harsh climate environment(rain,snow,fog,etc.)or night,etc.The recognition accuracy of the algorithm will fall rapidly even unable to recognize.Therefore,the point cloud collected by the lidar sensor is used as the input in this paper to overcome the traditional algorithm’s disadvantages.What’s more,the Velodyne HDL-32 E lidar has a better viewing area than traditional cameras.And the captured information which is improved from 2-D to 3-D become more abundant.The main steps of this algorithm is to extract the road boundary information frame by frame by collecting the three-dimensional point cloud information of driving environment by lidar.After that,the boundary information is used to segment each frame of the road point cloud data.Finally,point clouds containing road segmentation information in each frame are stitched together to generate a complete point cloud map.The main content and innovation of this paper are as follows:(1)In terms of the selection of data sets,traditional papers use public data sets to verify the performance of the algorithm.In this paper,we collect data on the campus environment through the experimental platform independently built by the laboratory.We fix lidar,aluminum profile fixator and the experimental car,collect school environment’s3D point cloud data frame by frame.(2)For the problem of road boundary feature extraction,this paper uses the idea of rasterization to preprocess the discrete point cloud for dimensionality reduction firstly.After that,grid was used as the clustering unit for preliminary screening through gradient conditions.Then,according to the geometric relationship and spatial distribution feature of the road boundary points,the innovation points of the geometric feature screening conditions and the single laser data point distribution feature condition are proposed.The geometric feature screening condition realizes the rough clustering of road boundary points.The single laser data point distribution feature condition filters out the noise points(low vegetation,stone pier,etc.)which have similar features with the road boundary points,and achieves the affect of fine clustering of road boundary points.(3)For the problem of road boundary fitting.This paper combined RANSAC(Random Sample Consensus)fitting and polynomial least square fitting to achieve the affect of road boundary grid point denoising and curve generation.In the RANSAC quadratic curve fitting,the time complexity of the process of each fitting is optimized.Finally,the segmentation and extraction effect of single frame point cloud road is achieved through the fitting boundary,which provides the road segmentation information of point cloud data in each frame for the later point cloud registration and map generation.