Ground Target Detection Method And Application Based On 3D Lidar

Autonomous driving system is an artificial intelligence platform that integrates map engine,highprecision positioning,environment perception,prediction,planning and decision-making and motion control.It has broad application scenarios in alleviating traffic pressure and improving traffic environment.With the deepening of people's research on autonomous driving,when smart cars are driving in complex road environments,they need to continuously analyze and process the surrounding environment to obtain accurate obstacle information.Due to the characteristics of sensitive response speed,strong anti-interference ability and high resolution,lidar plays a vital role in the automatic driving system.This article focuses on vehicle laser radar and roadside laser radar,focusing on the detection and recognition of pedestrians,vehicles and other targets in road traffic scenarios.The main research contents are as follows:(1)A method of making offline grid maps to filter out point clouds of invalid areas is proposed.In order to eliminate the influence of invalid driving area on target detection,an offline raster map can be made to filter out point cloud data that has nothing to do with driving area.In order to solve the problem of inaccurate location information collected by RTK in some areas affected by building occlusion,this paper uses a univariate polynomial regression method to determine the corresponding relationship between the Gaussian projection x-axis coordinates and the Gaussian projection y-axis coordinates,fitting to obtain complete road boundary information.Rasterize the area according to the road boundary to get an offline grid map.This method is more accurate than the grid map made by ear-clipping.Finally,according to the "and" operation of the offline grid map and the radar scanning area label,the effective area point cloud data is extracted.(2)A vehicle lidar point cloud target detection system based on multi-level European clustering is designed.First,the lidar is calibrated,and the point cloud data of the two 16-line lidars are fused;then,according to the offline raster map made in this paper,the invalid areas such as the green belt outside the road are filtered out;then the effective area is used to divide the effective area Intra-ground point cloud filtering to reduce the impact of ground point clouds on target detection;for non-ground point clouds,the hierarchical European clustering method is used to cluster the targets according to the spatial point cloud distribution;finally,the clustering targets are drawn The smallest circumscribed rectangle,classify the target according to its geometric characteristics,and obtain the final recognition result.Experimental results show that the vehicle-mounted radar target detection system designed in this paper can meet the requirements of accuracy and real-time in the automatic driving system.(3)A roadside lidar point cloud target detection system based on Squeeze Seg and Voxel Net is designed.In order to obtain information about obstacles around the road,a lidar is deployed on the side of the road to detect and analyze the road environment.After the original point cloud data is cut by dimensionality reduction,they are input into the Squeeze Seg and Voxel Net models respectively.Squeeze Seg uses a combination of convolutional neural network and conditional random field to segment point cloud data point by point.Voxel Net uses a feature learning layer,a convolutional intermediate layer,and a region generation network.It combines point-by-point features with local features to accurately predict the three-dimensional bounding box of obstacles.The experimental results show that Squeeze Seg has a simple structure and fast calculation speed,but has low detection accuracy,and is suitable for high-speed scenes such as straight travel,while Voxel Net has a complex structure and high detection accuracy,but has low calculation speed and is suitable for low-speed scenes such as intersections or parking lots.