Research On 3D Object Detection System For Airport Cargo Environment

3D Object Detection plays a vital role in the field of automatic driving.Accurate mapping and localization,safe obstacle avoidance and planning tasks in intelligent vehicles are inseparable from accurate test results.The 3D object detection technology in the largescale scene environment mainly completes the classification and positioning tasks for the object of interest.On the other hand,the cargo transportation throughput of major domestic and foreign airports is huge,especially the Hong Kong airport,which is dominated by airport cargo,in which achieving self-driving tractor to assist in the high-quality haulage of goods has great significance in terms of safety and economic development.This thesis is based on the project of implementing autonomous driving in the cargo scenario of Hong Kong Airport.Aiming at the problem that it is difficult to achieve a robust and high-precision 3D obeject detection effect with a single sensor in this special scene,the sparseness of the lidar sensor,and the need to detect a variety of objects that lack training sets in the public data set,etc.,this thesis realizes the research of 3D object detection system based on the fusion of camera and lidar information,including the design of the hardware platform,the software framework and the algorithm flow for 3D object detection.The research is as follows:In order to facilitate the actual testing of the data acquisition and detection algorithms,a sensor system information platform was built.The selection and layout of six cameras and a lidar and the drive under the ROS framework have been completed.The internal and external parameters of the camera and lidar were calibrated.For the sparsity problem of the 16-line lidar point cloud,the EPn P algorithm is used to complete the convenient and high-precision joint calibration method of the camera and the lidar.A framework of 3D object detection algorithm based on camera and lidar information fusion is proposed.The algorithm framework utilizes the idea of 2D object detection results to drive the 3D frustum region: First use the YOLOv3 algorithm with great speed and accuracy to detect the object in 2D image.The collection of images that do not exist in the public datasets such as airport trailers and custom containers are collected and marked,and the training of the network is completed.Then,according to a fusion data storage structure designed in this paper,the quadrilateral pyramid region containing the target 3D point cloud is generated from the 2D detection bounding boxes.Finally,theLocation of the 3D object point cloud is realized by a series of 3D point cloud processing methods without deep learning in the frustum region,and avoid making expensive 3d point cloud detection training set.The 3D point cloud processing flow designed in this paper includes the lidar point cloud ground segmentation,3D point cloud clustering,and using a novel scoring mechanism to segment the object point cloud.Thereby completing the classification and localization tasks for the 3D object detection.At the end of this paper,the experimental analysis of the designed system is carried out in the actual airport cargo scene,and robust real-time 3D object detection is realized,which can effectively detect the interested objects and obtain their 3D information.The effectiveness of the algorithm in different scenarios is also verified on the KITTI dataset.