基于Pseudo-LiDAR和立体相机的道路移动目标检测研究

Study on Moving Object Detection on Road Based on Pseudo-LiDAR and Stereo Camera

面向自动驾驶的道路移动目标检测方法在很大程度上依赖于昂贵的LiDAR传感器来获取准确的3D边界框,由双目立体相机图像转换得到的Pseudo-LiDAR则使低成本获取目标深度信息成为可能。首先综述了基于Pseudo-LiDAR的单目视觉和双目立体视觉目标检测算法,对基于Pseudo-LiDAR的相关方法所采用的网络结构、深度估计及修正、损失函数设计进行了对比总结。然后分析了Pseudo-LiDAR目前存在的局限性,如点云生成结果易受天气条件影响且存在长尾现象、由图像估计物体深度仍是不适定问题等。最后提出了几点改进措施,如合理利用路侧双目摄像头预训练Pseudo-LiDAR生成模型并迁移部署到车端,以解决车端算力不足的问题;或充分使用路端已部署的多视角相机,研究多视角下的伪激光雷达点云生成方式等。

Automated driving-oriented road moving target detection methods rely to a large extent on expensive LiDAR sensors to obtain accurate 3D bounding boxes.Pseudo-LiDAR,which is converted from binocular stereo camera images,makes it possible to obtain depth information at low cost.First,the object detection algorithms of monocular vision and binocular stereo vision based on Pseudo-LiDAR are reviewed.The network structure,depth estimation and correction,and loss function design adopted by Pseudo-LiDAR-based methods are compared and analyzed.Then,the current limitations of Pseudo-LiDAR are analyzed.For example,the point cloud generation results are easily affected by weather conditions with long tail phenomenon,and the estimation of object depth from images is still an ill-posed problem.Finally,several improvement measures are proposed,such as the rational use of roadside binocular cameras to pretrain the Pseudo-LiDAR generation model and transfer it to the vehicle,so as to solve the problem of insufficient computing power of ego-vehicle.It is also possible to make full use of the multi-view cameras deployed at roadside to explore the generation approach of Pseudo-LiDAR point clouds.