Research On Fast Ground Point Segmentation Method For Autonomous Driving Based On Lidar Data

Autonomous driving is attracting more and more attention from the society because of its innovative impact on society and transportation,with the wide detection range,accurate measurement,the Lidar in the current autonomous driving areas are deeply research and widely used.High-precision Lidar sensors receive millions of points every second,with half of the Lidar data reflected off the ground.In order to ensure the real-time information processing of autonomous driving vehicles,it is not necessary to directly use the original data.We use ground segmentation to remove ground points,which greatly improves the speed of subsequent tasks such as clustering,identification,tracking and control.This thesis studies ground segmentation for automatic driving based on Lidar data,including the following contents:1.In view of the influence of Lidar sensor hardware itself,the distribution of Lidar data collected is not uniform,and dynamic mesh division is introduced.The size of the grid varies with the distance from the Lidar,so that the number of Lidar data in each grid is roughly uniform.This method makes the partial feature heuristic ground segmentation method get better performance without increasing the computational complexity.2.A simple ground segmentation method for Lidar 3D point clouds is proposed.In order to solve the problem of too many linear regressions when building ground skeleton,the idea of dichotomy is introduced to reduce the number of linear regressions.The region of interest(ROI)was divided according to the computing power of the processor,and the lowest point in each grid was used as the seed point.A conditional filter is used to filter the noise in the seed points according to the height characteristics and slope.The key points were found through the idea of dichotomy,and the ground skeleton model was built by using the least square linear regression.Finally,the ground is segmented according to the ground skeleton model.Experimental results show that our method has high speed and good accuracy in different scenarios.3.A downsampling method based on distribution features is proposed to deal with the influence of laser radar sensor hardware and processor hardware itself.Under the condition that information is not lost as far as possible,the sampling is carried out to adjust the number of Lidar data in a reasonable range.4.To solve the problem of ground point segmentation on low power edge computing unit,GSECnet is proposed.GSECnet converts the original Lidar data into discrete representation through Pillar Grid Map,and Pointnet extracts the features of Lidar data in the grid,and then obtains the corresponding pillar feature map.Based on the U-Net neural network framework,GSECnet uses deep separable convolution to greatly reduce the amount of calculation and the number of parameters,and uses Focal Loss to solve the imbalance of positive and negative samples,and applies attention mechanism to capture important features.Finally,the GSECnet learns the pillar grid classification from the pillar feature map.