Study On Real-time Object Detection Based On 3D LiDAR

Autonomous driving technology can fundamentally improve the traffic safety,travel efficiency,and driving comfort.One of the key components for the safe operation of autonomous vehicles is the reliable environmental perception system.The perception system is usually a combination of hardware and software.Different sensors are the main hardware components of the environmental perception system of autonomous vehicles,and the 3D Li DAR is one of the most prevalent components.3D Li DARs are robust,whether it is day or night,with or without glare and shadows.There are many problems in the object detection methods of autonomous vehicles that collect images based on sensors such as cameras.For example,due to the influence of light,those methods cannot achieve ideal recognition accuracy at night.Most of the object detection methods are limited to the view in front of the sensor and cannot detect the moving objects behind.This thesis studies real-time object detection based on 3D Li DARs,and proposes a complete set of efficient and robust algorithms for object detection based on 3D Li DAR point clouds.Based on the above algorithms,this thesis also proposes a semantic segmentation method of road object based on convolutional neural networks.The main study contents of this thesis are as follows:(1)The transformation relationship between 3D point cloud data and panoramic depth map are established,and the panoramic depth map is enhanced.In view of the unstructured nature of 3D point cloud data,which aggravates the difficulty of neighbor search,the time-consuming problem of dealing directly with 3D point cloud is solved.The code of neighborhood relation implied in the panoramic depth map is discussed,the 3d point cloud data is converted into the panoramic depth map through the projection model of spherical coordinate system,the complexity of calculation is reduced,the real-time performance of object detection is guaranteed,and the high efficiency of directly processing the panoramic depth map is verified.The enhancement processing algorithm of interpolation and bilateral filtering for panoramic depth image is applied to solve the problem of poor image quality caused by noise and remote scanning,and the effectiveness of the algorithm in improving image quality is verified.(2)The algorithms of removing the ground and filtering the 3D point cloud,clustering and secondary filtering the remaining point cloud,and projecting the clustering point cloud onto the panoramic depth map are illustrated.The improved Ground Plane Fitting(GPF)algorithm is used to established a Ground model to remove the Ground and filter the point cloud.The 3d clustering of the remaining point cloud is realized by treating the remaining 3d points after removing the Ground as the pixel points of the image and using the two-run connected component labeling technique of the binary image.A series of algorithms for performing secondary filtering on the point cloud after clustering according to the characteristics of the detection object are proposed,and the remaining point cloud after secondary filtering is projected onto the panoramic depth map through the spherical coordinate system projection model,which verifies the effectiveness of the algorithm on highlighting the detection object.(3)After clustering and filtering,the panoramic depth map is spliced,and the panoramic depth maps with different sizes are detected by the convolution neural network for object pedestrians,and then the results are compared with the classical convolution neural network such as PVAENT and Faster R-CNN.After splicing,the size of the dataset consisted of the new panoramic depth map is increased,which is more suitable for convolutional neural network such as PVANet and Faster R-CNN.By improving the network structure of PVANet,it is more applicable to detect pedestrians and other "small objects".Compared with PVANet,the average accuracy of the proposed network structure for object detection is increased by 2.8-5.1%,and compared with Faster R-CNN,the average accuracy of the proposed network structure for object detection is increased by 40%-45%.(4)Based on the analysis of the lightweight network Squeeze Net,a faster running time algorithm is proposed on the premise of ensuring the accuracy of object segmentation,and the segmentation results are compared with the deep-based methods such as Squeeze Seg and Point Seg,as well as the traditional point cloud segmentation methods.Squeeze Net is improved by adding domain transforms to align the segmented objects precisely with their boundaries,so as to obtain better point-to-label mapping as output,which can improve the detection accuracy.At the same time,the visualization differences between the proposed point cloud segmentation results and the traditional point cloud segmentation results are analyzed.