Research On Point Cloud Based On Object Detection Algorithm

With the rapid deployment of fifth generation(5G)mobile wireless communication technology in the world,the 5G makes it a reality for car to communicate with the other cars in autonomous driving,and it applies conditions for the autonomous driving universal applications.The perception of the surrounding environment is a key step in the autonomous driving system,in addition,the dynamic object detection,as the crucial part of the perception system,is significantly important.Since Light Detection and Ranging(LIDAR)can provide high precision and wide range environment depth information,and it can work usually under weak light conditions,the point clouds object detection based on the LIDAR becomes a hotspot topic.The previous point cloud object detection technology can be divided into the object detection methods based on deep learning and the object detection methods based on traditional technology.However,in the traditional object detection methods,there are some shortcomings need to be improved,such as that the accurate ground estimation is timeconsuming,and the pose estimation of the object is inaccurate.In the deep learning methods,there is still much room to improve in point cloud object detection performance.Based on the aforementioned problems,we search methods to solve the above problems,so as to improve the performance of the object detection.Firstly,to solve the time-consuming problem in the accurate ground estimation,we propose a point cloud ground removing method based on fast segment line fitting technology(FSLF-GR).We segment the point cloud into multiple grid regions according to imaging characteristic of the point cloud,each grid region has some points.Then we extract lines fast in each sub segmentations,we fit the ground line in the current segmentation by removing several incompetent lines and connecting the lines which need to be connected.The point cloud can be separated into ground points and nonground points by calculating the distance from the ground line in the end.FSLF-GR demonstrates a good performance in urban environment and on uneven road in suburban environment.Secondly,in order to cope with the problem that the pose estimation of the object is inaccurate in the bounding box methods,we present an optimal contour-based L-Shape bounding box fitting method using the random consistency sample detection(OCL-BBox).This method uses the outermost contour point algorithm to extract the contour points of the point cloud objects then uses L-shape algorithm to extract the two main sides of the point clouds facing the LIDAR.Finally,OCL-BBox uses the contours points within a certain angle range of the optimal main side to obtain the best target cloud orientation.This method greatly improves the orientation accuracy of point cloud optimal bounding box and it takes less than 1ms for every 100 targets.Finally,to solve the time-consuming problem of the 3D convolutional neural network for the point object detection,we propose an object detection method which can capture the spatial feature under the condition of ensuring operation efficiency.We construct a Multi-view Semantic Learning Network(MVSLN)for 3D object detection.We firstly use the Multiple Views Generator(MVG)module to generate multiple point cloud pseudo images from four views of the point cloud.Then,the Spatial Recalibration Fusion(SRF)operation is utilized to adjust and fuse the locations of features for these four views,which can extract the semantic features of the pseudo images.Last but not least,the semantic features are utilized by 3D convolutional neural network to classify the point cloud object and fit the optimal bounding box.The experiments on KITTI demonstrate our proposed methods dramatically improve the point cloud object detection performance compared with the previous methods.