3D Object Detection In Electric Multiple Units Scene

With the rapid expansion of the operation scale of Electric Multiple Units（EMUs）in China,the maintenance of EMUs is also facing increasingly severe challenges.The maintenance method based on man-labor has low efficiency and high missing rate; The maintenance method based on machine vision has the advantages of high efficiency and low missing rate,which provides a new strategy for EMUs maintenance.Common visual information includes 2D images and 3D point clouds.However,there remain the following challenges in the maintenance of EMUs based on machine vision: 1.Lack of benchmark dataset; 2.The field of view of the 3D scanner is inconsistent with that of the camera which leads to an information mismatch between the image and point cloud; 3.The interference of non-object point clouds seriously makes it difficult to locate small objects on dense point clouds.The following works are carried out to solve these problems respectively:1.In response to the lack of data,we constructed a 3D object detection dataset in EMUs scene which contains a total of 25161 2D objects and 9117 3D objects of three categories,and the 3D bounding box of each object has any degree of freedom.Our work fills the blank of the 3D object detection dataset in the EMUs scene.2.Aiming at the problem of mismatch between image and point cloud information,a depth completion algorithm of multi-scale context information aggregation based on codec structure is designed by using deep learing.The encoder structure combining attention mechanism module and global average pooling module.The decoder structure is based on convolutional neural network.Our algorithm can efficiently recover depth in image edge area.3.Aiming at the problem of the interference of non-object point clouds,a 3D object detection algorithm based on two-stage object segmentation and bounding box esitmation is designed using deep learning.The two-stage object segmentation module is constructed with 2D object detector and 3D object segmentation module,and the bounding box estimation module is constructed with 6D degrees of freedom.Our algorithm can effectively and efficiently detect small objects on dense point clouds.The experimental results show that compared with the Frustum object detection algorithm,our object detection algorithm designed in this thesis can improve the accuracy of 3D object detection in the EMUs scene to 84.1%,an increase of 14.6%.Thus,the feasibility of the machine vision-aided maintenance method in EMUs scenario is proved.