Research On 3D Object Detection For Traffic Scenes

3D object detection technology has been widely used in intelligent traffic scenes.The research of 3D object detection technology for traffic scenes has important scientific value and application value.At present,the main research objects of 3D object detection in traffic scenes include visual data and 3D laser-point cloud data.According to different data types,3D object detection technology can be divided into methods based on pure vision,laser point cloud(hereinafter referred to as point cloud)and the fusion of vision and point cloud.This paper mainly focuses on laser-based point cloud and fusion-based methods,researches on the difficulty in learning features caused by the disorder of point cloud data and the difficulty in fusion of point cloud and visual images,and designs and develops a set of track obstacle detection system based on the scene of rail transit.The specific research contents of this paper include:(1)A point cloud feature enhancement algorithm based on self-attention mechanism is proposed.To disorder caused by the 3D point cloud data of local point cloud characteristics to study the problem of inadequate,the algorithm through the attention mechanism,in the local point cloud creates a link between each two points,make full use of the location of the relationship between,to enhance the local characteristics of point cloud learning ability,provide more information for the downstream tasks,by raising the precision of the model.Experimental results on the open data set Kitti show that the proposed algorithm significantly improves the detection accuracy of the original detection model in the 3D object detection task,and the average detection accuracy is improved by about 2%.(2)A multi-sensor feature fusion algorithm based on point cloud transformation is proposed.Visual sensors can make up for the lack of color,texture and other features of point cloud data and improve the differentiability of point cloud data.It is very difficult to directly fuse the data due to the great difference between the point cloud and the modal form of the image.To solve this problem,this paper first transforms the presentation form of point cloud to make point cloud and RGB image span the difference of data presentation form.Then,the feature fusion module is designed based on the attention mechanism to fully integrate the feature of point cloud and RGB image,and effectively enhance the feature information of point cloud,so as to improve the detection accuracy.This scheme can be applied to any current 3D object detection framework of pure point cloud and fuse RGB features into the point cloud without changing the structure of the point cloud.Experimental results show that the enhanced point cloud can improve the detection accuracy of objects in 3D object detection tasks,especially for small objects such as pedestrians.The average accuracy of humans increases by about 5% in the Kitti data set.(3)The urban rail obstacle detection system is designed and implemented.Aiming at the problem of obstacle intrusion in the scene of public rail transit,based on the QT platform,this paper designs and implements an obstacle detection system with convenient operation and visible results.In order to realize the information interchange of multisensors,the multi-machine communication mechanism of ROS system is used.In order to deal with the problem that various detection categories are needed and difficult to distinguish in the actual scene,the unsupervised learning method is adopted,and the traditional clustering algorithm is specifically used to improve the security and reliability of the system.Further,in order to solve the problem of low operating efficiency and high false positive rate of the system,track line segmentation is carried out in the visual image in the system,and the point cloud is filtered with the result of image segmentation,so as to reduce the redundant calculation and improve the detection efficiency.