Research On Binocular Stereo Matching Method For Road Vehicle Detection

As the core component of the self-driving perception technology,the three-dimensional object detection technology provides the input information for the entire self-driving driving system and is responsible for the safety of autonomous vehicles.Three-dimensional object detection of vehicles in the road traffic environment is taken as the research object in the thesis.The existing three-dimensional object detection methods based on Lidar are too expensive,while those methods based on vision need too much computational resources.In the thesis,a low-cost real-time three-dimensional object detection method based on binocular camera is proposed.The technical route in the thesis is divided into two steps.The first step is using stereo matching algorithms to transform the stereo image into a pseudo point cloud with a threedimensional spatial structure,and the second step is using the advanced point cloud threedimensional object detection algorithm for vehicle detection.The main work of the thesis is as follows:(1)To alleviate the problem of too much parameters and too heavy calculation burden of the existing stereo matching network,a lightweight real-time stereo matching network based on the attention mechanism is proposed.The proposed stereo matching network is based on an efficient network architecture of shared features,low-resolution cost volume and superresolution module.The proposed stereo matching network first generates a low-resolution disparity map,and then uses a super-resolution module to upsample the low-resolution disparity map.A super-resolution module based on the spatial attention mechanism is proposed to alleviate the problem of edge blur and at the same time control the overall calculation burden.Furthermore,Tesor RT,an inference framework,is used to optimized the proposed network and improve inference speed.The proposed method is verified by comparative experiments on multi public datasets and multi devices,which demonstrate the proposed method is valueable in practice.In Scene Flow dataset,the proposed method achieves 0.9px EPE and 69 Fps inference speed.And in KITTI dataset,the proposed method achieves 3.44% error rate and 68 Fps inference speed.On embedded AI devices,the proposed method is able to generate high-quality disparity map in real-time.(2)In order to deal with the low-precision of the stereo matching network in the depth estimation task,the depth cost volume and depth loss function are proposed to modify the stereo matching network and a depth estimation network based on stereo vision is proposed.Using public datasets,it is verified that the proposed method improves the depth estimation accuracy of the stereo matching network,and the generated pseudo pointcloud can replace the real point cloud collected by lidar to a certain extent.The network structure and detection principle of Point Pillars,a three-dimensional object detection algorithm based on pointclouds,is studied.Point Pillars is used to process the generated pseudo pointcloud to generate a three-dimensional object detection bounding box.Using public datasets,the technical route proposed in this thesis is verified.(3)The construction methods of the stereo experiment platform are studied,and the collection and labeling of the local dataset are compeleted.Aiming at solving the problem of camera and lidar time synchronization,a time synchronization data acquisition method based on camera external triggering,GPS timing and PTP technology is proposed.Based on the local dataset,the algorithms proposed in this thesis is verified.