| In automatic driving,it is very important to accurately perceive the surrounding environment.With the development of science and technology,the traditional 3d object detection algorithm has been replaced by the algorithm based on deep learning because of its slow speed and low accuracy.At present,due to the rapid development of computing power and the advantages of lidar that can accurately give the three-dimensional appearance and position of the object,most of the 3d object detection algorithms choose lidar as the data source.Each point cloud contains a huge amount of data,which has a completely different format from image data.How to use depth learning to effectively process point cloud data is the focus and difficulty of research.As the ”eyes” of automatic driving,it is of great significance to study the 3d object detection algorithm.All experiments in this paper are conducted on the Kitti data set,which contains point cloud data collected by velodyne-hdl-64 e lidar and corresponding image data collected by two cameras.The purpose of this paper is to find a real-time high precision vehicle detection algorithm for automatic driving.This paper mainly studies three kinds of 3d object detection algorithms,and puts forward: a scheme of embedding deep learning into the traditional three-dimensional target detection framework;a 3d object detection algorithm integrating image information;and an anchor-free 3d object detection algorithm based on voxel for the first time.The main work of this paper is as follows:1.3d object detection framework is studied: in the ground point filtering task,a ground point filtering algorithm combining the front view and convolution neural network is proposed;in the point cloud clustering task,the characteristics of lidar data are studied,and a hierarchical Euclidean clustering algorithm is used;in the object recognition process,the characteristics based on the bounding box are proposed and then the object is identified with deep learning based object detection algorithm.Experiments show that the above algorithm can improve the effect and speed of traditional algorithm.2.Point-based 3d object detection is studied,and a method to fusion image data is proposed: a new point cloud classifier is trained by using the point cloud in the obtained prediction bounding box,and the prediction bounding box is projected back to the image to train an image classifier,and the classification confidence of the algorithm is improved by combining the above two classifiers which can improve the detection accuracy.Experimental results show that the above algorithm can effectively improve the detection precision of the original algorithm,but the speed is reduced.3.Voxel-based 3D object detection is studied,and a fast voxel generation algorithm is proposed.In addition,anchors are widely used in this kind of algorithm,but more super parameters are introduced due to anchors.This paper presents an algorithm to remove anchors,which proves that anchors are unnecessary.Experimental results show that the above algorithm can effectively improve the speed of voxel generation and the precision of the algorithm. |
