Research On Target Detection And Recognition Of Intelligent Vehicle Based On Lidar Point Cloud

With China’s economy and society continues to improve,traffic accidents caused by huge car ownership still remain high,and the traffic congestion also has become the pain point of urban governance.Therefore,the development of intelligent vehicles has become an effective means to manage the above two problems.As a critical technical to the intelligent vehicle,environment perception is one of the research hotspots in the field.This research is based on the National Key Research and Development Program(No.: 2020YFB1313400)and the key project of the National Natural Science Foundation of China(No.: U1864204).In this paper,multi-line lidar sensor is used to complete obstacle detection in the surrounding environment of intelligent vehicles.The research contents of this paper are as follows:(1)According to the characteristics of lidar 3D point cloud data,a combined filtering method is proposed.It is difficult for a single filtering method to reduce the number of point clouds,filter noise points and discrete points at the same time.This paper proposes a combined filtering algorithm considering the advantages of each filtering method.Firstly,through filtering algorithm is used to cut the invalid point cloud to reduce the invalid input value of the network;Secondly,the voxel filtering algorithm is used to downsampling the point cloud data to further simplify the point cloud data and reduce the network input;Finally,the improved radius filtering algorithm is used to remove the obvious outliers in the point cloud data and optimize the network input data.The combined filtering method can effectively remove outliers on the premise of preserving the integrity of the target contour.(2)Based on Yolo v4 two-dimensional target detection network,a one stage obstacle detection network is proposed.The one stage target detection scheme which is used in image detection is transferred to the point cloud data.In order to void the of three-dimensional convolution calculation and to ensure the speed of model reasoningand,the aerial view feature map of LIDAR point cloud data is used as the model input.At the same time,in order to ensure that the real-time performance,a simplified network structure is adopted.Different from the Yolo v4 target detection model,the structure of the model is simple,and the model can complete the target detection efficiently with fast calculation speed.The test results on the same test platform and Kitti data set show that the detection speed of the model proposed in this paper is6.98 times faster than that of Yolo v4 model,which can meet the needs of real-time detection of intelligent vehicles.(3)Based on the data collected from real vehicles,the point cloud data set is constructed for filtering,point cloud segmentation and obstacle detection experiments of intelligent vehicle perception system.In order to apply the model to actual scene project,this paper collects the scene information of campus,then preprocesses and labels the data,and finally arranges it into real vehicle data set.Through model training and result verification,it is proved that the proposed algorithm can be used in practical application.