Research On 3D Object Detection Technology Based On Deep Learning

3D object detection is based on the measurement results of one or more sensors to locate the object in 3D space and detect the boundary box.For the perception of outdoor scene,3D object detection is very important for the safety and reliability of automatic driving,because3 D object detection can perceive the absolute position and size of vehicles,pedestrians,non motor vehicles,obstacles,road edges,etc,compared with 2D object detection and semantic segmentation,and can even obtain the orientation information of objects,so as to serve as the object behavior prediction,target action planning,vehicle control and simulation provide necessary perception information.However,in the scene of automatic driving,the point cloud has the characteristics of large scale,disorder,uneven distribution,and because of occlusion and self occlusion,the object can only collect part of the point cloud information,which makes it difficult to effectively process the point cloud,and there is a big gap between the original point cloud and the 3D object detection results.For these problems,this paper has done the following work:1.A large number of research on one or more of the 3D object detection technologies,such as image,binocular stereo vision and point cloud,and the deep learning methods related to 3D object detection are carried out.The point cloud 3D object detection algorithm based on deep learning is introduced in detail.The advantages and disadvantages of various detection methods are analyzed and summarized,and the future research trend of 3D object detection is predicted.2.According to the characteristics of point cloud in automatic driving scene,put forward scale FPS to improve the origin farthest point sampling(FPS)method,which improves the sampling rate of front points,makes the subsequent extracted point cloud features richer,and improves the accuracy of proposal generation stage.3.Aiming at the problem that the 3D object detection algorithm with point cloud as input does not make full use of the supervision information,and there is a semantic gap between the original point cloud and the detection result,this paper proposes the structural aware feature,which is greatly improved in the proposal generation stage,and the final 3D object detection accuracy is also improved compared with its benchmark.