Research On 3D LiDAR Based Urban Road Segmentation

In the autonomous driving task,the reliability of the road segmentation task is an important prerequisite to ensure the safe vehicle driving.The main purpose of the road segmentation task is to divide the drivable road area in the scene,and define a safe driving range for the path planning task.Current road segmentation methods have some problems,such as the inability to balance computational complexity and accuracy,the inability to adapt to the lighting changes of the scene,and the fluctuation of the road slope.This paper mainly focuses on the existing problems.According to the point cloud information obtained by the 3D LiDAR sensor,the road segmentation in the urban scene based on the LiDAR point cloud is studied and verified on the public data benchmark.The main research contents of this article include the following three parts:(1)Aiming at the problem of low accuracy of current road segmentation methods,this paper proposes a road segmentation method based on histogram statistics and point cloud image traversal.First,the 3D disordered LiDAR point cloud is projected to the camera image plane,named point cloud image.Because the directly projected data is sparse,it is filled into dense point cloud image by up-sampling,and road segmentation is realized on this basis.For the dense height point cloud image,first perform histogram statistics on them,and make a preliminary estimate of the ground based on the characteristics of the road in the histogram.Then,the points that do not belong to the actual road are eliminated by traversing the image,and the final road estimation result is obtained.Experimental results show that this method can obtain accurate road areas.(2)Aiming at the problem of poor real-time performance of road segmentation methods,this paper proposes a road segmentation method based on LiDAR images.The representation of the LiDAR image can represent the neighborhood relationship of the 3D LiDAR points in a small-sized two-dimensional space,thereby improving the execution efficiency of the algorithm.For road segmentation tasks,it is mainly divided into RANSAC-based ground estimation and LiDAR image traversal-based road result optimization.After obtaining the road segmentation result under the LiDAR image plane,it is projected to the camera image plane for up-sampling processing to obtain an accurate road segmentation result in the camera image.Experimen-tal results show that this method can significantly reduce the time cost of the algorithm while ensuring the accuracy of the results,and is suitable for real-time road segmentation tasks.(3)Aiming at the problem that traditional road segmentation methods need to set more artificial thresholds,this paper proposes a road segmentation network based on attention mech-anism.Embedding the attention mechanism in the previous layer of the classifier in the road segmentation network can improve the performance of the road segmentation network.The at-tention mechanism proposed in this paper mainly includes two modules: pixel-wise attention mechanism and channel-wise attention mechanism.Experimental results show that the proposed network can obtain high road segmentation accuracy while being sufficiently lightweight.