【作者】 王斌；

【导师】 赵佳琦；

【作者基本信息】 中国矿业大学 ， 计算机科学与技术， 2024， 硕士

【摘要】 三维目标检测技术是指利用传感器采集的数据,如点云、RGB图像等,对周围环境中的目标进行识别和定位的技术,在自动驾驶、车路协同以及机器人导航中发挥着重要作用。根据使用传感器的不同,三维目标检测技术主要可以分为基于激光雷达的方法、基于双目相机的方法、基于单目相机的方法等,由于单目相机价格低廉、使用广泛,近年来吸引了众多研究者的兴趣。然而,由于缺乏深度信息,基于单目相机的方法的检测准确率相比于基于激光雷达和双目相机的方法有较大的差距。本文分别针对基于路侧和车载相机的单目三维目标检测技术进行了研究,主要工作如下:(1)针对单目深度估计的不适定性,以及路侧相机光轴不平行于地面的问题,本文提出一种基于地面深度与视觉特征融合的路侧单目三维目标检测方法。首先,该方法将目标三维中心深度估计问题转化为目标接地中心深度估计问题,并利用相机坐标系下的地面方程生成像素级地面深度图。其次,该方法利用可变形注意力机制将地面深度图与视觉特征融合。最后,该方法在相机光轴倾斜场景下重新定义关键点的计算方法,使得目标高度与深度关系适用于路侧相机。实验结果表明,在路侧数据集Rope3D上,该方法的平均精度超越了简单移植到路侧的车载单目三维目标检测方法,并与最新的针对路侧相机设计的方法接近。(2)针对单目图像缺少深度线索、数据标注昂贵的问题,本文提出一种基于两视图的半监督车载单目三维目标检测方法。首先,该方法利用标注数据训练激光雷达检测器,并使用训练好的检测器和未标注数据生成伪标签,使单目检测器受益于无标注数据同时减少人工标注成本。其次,该方法利用两帧时序图像以及相机在两帧间的运动构建运动感知代价体,并引入注意力权重生成模块提高特征显著性。最后,该方法使用一种鸟瞰图下可伸缩的非极大值抑制方法,有效减少了冗余的预测框。实验结果表明,在车载数据集KITTI上,该方法提升了检测的平均精度。综上,本文聚焦于车路协同系统中的单目三维目标检测技术,充分发挥各自优势以提高检测平均精度。在路端,以结构感知为主线,将固定相机视角下的地面深度引入到目标深度预测过程中;在车端,以运动感知为主线,将两帧时序图像与相机在三维空间中的运动信息结合。实验表明本文方法可以在一定程度上解决现有方法的问题,对于未来单目三维目标检测的研究具有一定的理论意义和实用价值。最后,本文设计了一个单目三维目标检测原型系统,整合了本文方法和相应的基线方法,方便用户使用模型进行推理、可视化以及进行不同方法的对比。更多还原

【Abstract】 3D object detection technology refers to the technology that uses the data collected by sensors,such as point clouds,RGB images,etc.,to identify and locate the object in the surrounding environment.It plays an important role in automatic driving,vehicle road coordination and robot navigation.According to the different sensors used,the 3D object detection technology can be divided into the methods based on Li DAR,binocular camera and monocular camera.The monocular camera based scheme is cheap and universal,and has attracted the interest of many researchers in recent years.However,due to the lack of depth information,the detection accuracy of the method based on monocular camera is much lower than that based on Li DAR and binocular camera.In this thesis,the monocular 3D object detection technology based on roadside and vehicle camera is studied,and the main work is as follows:(1)Aiming at the ill posed nature of monocular depth estimation and the problem that the optical axis of roadside camera is not parallel to the ground,this thesis proposes a roadside monocular 3D object detection method based on the fusion of ground depth and visual features.Firstly,the method transforms the problem of object 3D center depth estimation into the problem of object grounding center depth estimation,and generates pixel level ground depth map by using the ground equation in the camera coordinate system.Secondly,the method uses deformable attention mechanism to fuse the ground depth map and visual features.Finally,the method redefines the calculation method of key points in the scene where the camera optical axis is tilted,so that the relationship between object height and depth is suitable for roadside cameras.The experimental results show that on the roadside dataset Rope3 D,the average precision of this method exceeds the vehicle monocular 3D object detection method simply transplanted to the roadside,and is close to the latest method for roadside camera design.(2)Aiming at the problems of monocular image lack of depth clues and expensive data annotation,this thesis proposes a semi supervised vehicle monocular 3D object detection method based on time series two views.Firstly,the method uses labeled data to train the Li DAR detector,and uses the trained detector and unlabeled data to generate pseudo tags,so that the monocular detector can benefit from unlabeled data and reduce the cost of manual labeling.Secondly,the method uses two time-series images and the motion of the camera between two frames to construct the motion perception cost volume,and introduces the attention weight generation module to improve the feature saliency.Finally,the method uses a scalable non maximum suppression method under the aerial view,which effectively reduces the redundant prediction frame.Experimental results show that this method improves the average precision of detection on the vehicle dataset KITTI.To sum up,this thesis focuses on the monocular 3D object detection technology in the vehicle road cooperation system,and gives full play to their respective advantages to improve the average detection accuracy.At the end of the road,the ground depth under the fixed camera perspective is introduced into the object depth prediction process with the structure perception as the main line;At the vehicle end,the motion perception is taken as the main line,and the two time-series images are combined with the camera’s motion information in 3D space.Experiments show that this method can solve the problems of existing methods to a certain extent,and has certain theoretical significance and practical value for the future research of monocular 3D object detection.Finally,a monocular 3D object detection prototype system is designed,which integrates the method in this thesis and the corresponding baseline method,and is convenient for users to use the model for reasoning,visualization and comparison of different methods.更多还原