Ervin Lauren, Eastepp Max, McVicker Mason, Ricks Kenneth
Electrical and Computer Engineering Department, The University of Alabama, Tuscaloosa, AL 35487, USA.
Sensors (Basel). 2025 Jan 10;25(2):370. doi: 10.3390/s25020370.
Discretely monitoring traffic systems and tracking payloads on vehicle targets can be challenging when traversal occurs off main roads where overhead traffic cameras are not present. This work proposes a portable roadside vehicle detection system as part of a solution for tracking traffic along any path. Training semantic segmentation networks to automatically detect specific types of vehicles while ignoring others will allow the user to track payloads present only on certain vehicles of interest, such as train cars or semi-trucks. Different vision sensors offer varying advantages for detecting targets in changing environments and weather conditions. To analyze the benefits of both, corresponding LiDAR and camera data were collected at multiple roadside sites and then trained on separate semantic segmentation networks for object detection. A custom CNN architecture was built to handle highly asymmetric LiDAR data, and a network inspired by DeepLabV3+ was used for camera data. The performance of both networks was evaluated, and showed comparable accuracy. Inferences run on embedded platforms showed real-time execution matching the performance on the training hardware for edge deployments anywhere. Both camera and LiDAR semantic segmentation networks were successful in identifying vehicles of interest from the proposed viewpoint. These highly accurate vehicle detection networks can pair with a tracking mechanism to establish a non-intrusive roadside detection system.
当车辆在没有高架交通摄像头的主干道以外的道路上行驶时，离散地监测交通系统并跟踪车辆目标上的载荷可能具有挑战性。这项工作提出了一种便携式路边车辆检测系统，作为沿任何路径跟踪交通的解决方案的一部分。训练语义分割网络以自动检测特定类型的车辆，同时忽略其他车辆，这将允许用户跟踪仅存在于某些感兴趣车辆上的载荷，例如火车车厢或半挂车。不同的视觉传感器在变化的环境和天气条件下检测目标具有不同的优势。为了分析两者的优势，在多个路边站点收集了相应的激光雷达和摄像头数据，然后在单独的语义分割网络上进行训练以进行目标检测。构建了一个自定义的卷积神经网络架构来处理高度不对称的激光雷达数据，并使用受DeepLabV3+启发的网络来处理摄像头数据。对两个网络的性能进行了评估，结果显示出相当的准确性。在嵌入式平台上运行的推理显示实时执行与在任何边缘部署的训练硬件上的性能相匹配。摄像头和激光雷达语义分割网络都成功地从提议的视角识别出感兴趣的车辆。这些高精度的车辆检测网络可以与跟踪机制配对，以建立一个非侵入式的路边检测系统。
