毕业生知识库

当前货运编组站正面临基础设施老化、作业过分依赖人工的问题。出于作业需求，工作人员需频繁出现在轨道周边开展作业，如调车人工领车、列检、货检、摘钩、摘风管及装卸作业等，存在较大的安全隐患。本文针对当前货运编组站远距离目标检测能力弱、远距离深度信息稀疏以及侵限风险难以定量评估等问题，提出了基于传感信息决策级融合的目标检测与定位方法、基于多任务学习的环境深度估计方法和基于电子地图及目标实时定位的侵入检测方法，实现了面向铁路货运编组站的高精度、强鲁棒性的人员侵入检测系统设计。主要研究内容如下：

第一，面向场站远距离目标感知需求，针对RGB图像缺乏深度信息、激光雷达在远距离处信息稀疏等弊端，本文提出了一种基于传感信息决策级融合的目标检测与定位方法。首先，提出了一种面向远距离场景的目标检测方法，并通过逆透视变换获取地面目标的三维定位；同时，使用Pointpillar基于激光雷达点云实现三维目标检测；最后，基于Wasserstein距离度量对视觉和激光雷达的检测结果进行决策级融合。该方法既能保证检测目标的定位精度，又提高了远距离目标的检测准确率。方法的有效性和性能在现场数据集上进行了验证。

第二，针对激光雷达点云在远距离处深度信息稀疏，而传统的利用双目视差估计获得深度信息的方法对于远距离目标估计存在较大误差的问题，本文提出了一种基于多任务学习的环境深度估计方法。该方法通过将视差代价体三线性插值重构为深度代价体，降低了远距离目标的深度估计误差。此外，针对当前有深度标注的双目数据集深度标注获取成本高昂、标注过于稀疏等问题，本文提出了一种多任务学习网络和半监督的训练策略，通过分支间自监督提供稠密的伪标签降低了标注成本。上述方法在公开数据集进行算法效果和性能的验证。

第三，针对边缘计算设备的算力限制与定量评估侵入风险需求，本文提出一种基于电子地图及目标实时定位的侵入检测方法。首先，针对货运编组站铁轨密集、咽喉处道岔等特点对电子地图匹配过程中轨道占用判断精度带来的挑战，通过匹配先验行驶轨迹而非常规地匹配所有轨道，排除了无关轨道信息的干扰，实现了高精度的列车实时定位。其次，基于列车行驶轨迹与前后转向架的强约束关系，提出了一种针对铁路场景的实时姿态估计方法，实现了目标在地图中的实时定位。相较于直接移植面向通用公路场景的行车位姿估计方法，该方法可靠性较高且无累积误差。上述方法在直道和弯道两类场景下进行了人员侵入检测，实验验证了算法的可靠性。

The current freight marshalling yards are facing issues such as aging infrastructure and excessive reliance on manual operations. Due to operational requirements, personnel need to frequently work around the tracks, engaging in tasks such as manually leading trains, train inspection, cargo inspection, uncoupling, removing windbreaks, and loading and unloading operations, which pose significant safety hazards.In response to the weak long-distance target detection capabilities, sparse depth information at long distances, and the difficulty of quantitatively assessing intrusion risks in current freight marshalling yards, this paper proposes methods for target detection and positioning based on hierarchical fusion of sensor information, environmental depth estimation based on multitask learning, and intrusion detection based on electronic maps and real-time target positioning. This achieves the design of a high-precision and robust personnel intrusion detection system tailored for railway freight marshalling yards. The main research contents are as follows:

First, to meet the perception requirements of distant targets in the station area, addressing the lack of depth information in RGB images and the sparse information of lidar at long distances, a three-dimensional obeject detection method based on target-level sensor information is proposed. Firstly, an object detection method tailored for distant scenes is proposed, and the three-dimensional positioning of ground targets is obtained through inverse perspective transformation. Next, using PointPillar based on lidar point clouds to achieve three-dimensional object detection. Finally, based on the Wasserstein distance, target-level fusion of visual and lidar detection results is conducted. This method ensures both the positioning accuracy of detected targets and improves the detection accuracy of distant targets. The effectiveness and performance of the method are validated on-site data sets.

Second, Lidar suffers from sparse depth information for distant targets, while traditional methods that use stereo disparity estimation to obtain depth information result in significant errors for distant target estimation. This paper proposes an end-to-end environmental depth estimation method. By converting the stereo disparity cost volume into a depth cost volume, this method reduces the depth estimation errors for distant targets. Additionally, addressing the issues of small-scale datasets and sparse annotations in current stereo datasets, this paper proposes a multitask learning network and a semi-supervised training strategy, significantly reducing annotation costs and improving annotation quality, training efficiency, and accuracy. The effectiveness and performance of the above methods are validated on publicly available datasets for algorithm evaluation.

Third, In response to the computational constraints of edge computing devices and the demand for quantitatively assessing intrusion risks, this paper proposes a constraint detection method based on electronic maps and real-time target positioning. Firstly, to address the challenges brought by the characteristics of dense railway tracks and switch tracks at throat sections of freight yards to the accuracy of track occupancy judgment during the electronic map matching process, the paper matches prior driving trajectories instead of conventionally matching all tracks, eliminating the interference of irrelevant track information and achieving low computational cost and high-precision real-time train positioning. Secondly, based on the strong constraint relationship between the train's driving trajectory and the front and rear bogies, a real-time attitude estimation method tailored for railway scenes is proposed to achieve real-time positioning of targets on the map. Compared with directly transplanting vehicle pose estimation methods aimed at general highway scenes, 
the method is highly reliable and has no cumulative errors. The above methods are applied to pedestrian intrusion detection in both straight and curved track scenarios, and experiments verify the reliability of the algorithm.