A Hierarchical Target Vehicle Pose Detection Framework in Ro-Ro Terminal Environment

In light of the urgent need to automate and upgrade roll-on/roll-off (Ro-Ro) transportation methods in ports, many ports have started using transfer automated guided vehicle (AGV) to replace traditional manual operations. Automatic docking is a critical aspect of transportation, which relies heavily on accurately identifying and localizing target vehicles. However, frequent issues such as point cloud feature loss and truncation of point clouds—often due to perception dead zones during the docking process—can significantly hinder existing vehicle recognition methods. This article introduces a hierarchical vehicle pose detection approach. During the early and mid-stages of docking, we employ two branches for identifying target vehicles to reduce the effects of feature loss and the frequent changes in shape: Voxel R-convolutional neural network (CNN) and an improved search-based optimal bounding box (BBox) fitting algorithm, which are then combined through result fusion. Additionally, to tackle feature loss caused by point cloud truncation in the late docking phase, we have developed a method to detect wheel poses and calculate the target vehicle pose inversely. Building on this, we established a set of composite evaluation metrics for method switching, ensuring the stability and robustness of the results. Our hierarchical vehicle pose detection method has been successfully implemented in transfer AGVs and applied to port Ro-Ro logistics transportation. In datasets collected during actual docking processes, the recognition performance of this framework has surpassed that of the most commonly used 3-D object detection methods.