基于单目视觉的集装箱角件自动对位

To enhance the efficiency of container loading and unloading and realize automated operations，this study investigates monocular-vision-based automatic alignment between spreader and container，integrating deep learning and image-processing techniques. Monocular images of container hoisting conditions were analyzed with emphasis on the regional characteristics of corner components. To address the low pixel proportion of corner regions in high-resolution images，a two-stage "coarse positioning-fine segmentation" strategy was proposed. Based on the segmentation results，key feature points were detected，2D-3D point correspondences were established，and pose estimation was performed using the Levenberg-Marquardt algorithm. Validation was conducted on an AUBO-i10 manipulator alignment platform in laboratory settings. Experimental results demonstrate that the mean average precision（mAP）for detection of container corner components exceeds 95% in both laboratory and real-scene environments. Mean intersection-over-union（mIoU）for corner segmentation reached 98. 15% and 93. 89%，respectively-improvements of 1. 24% and 1. 64% over the baseline SegFormer-B0-while model computational cost was reduced by approximately 23. 2%. At a camera-to-corner distance of about 2 m，the aiming error of the alignment position was below 1. 0 mm. Absolute translation errors on the X，Y，and Z axes were all below 5. 0 mm，and absolute rotation error was below 0. 5°. These results indicate that the proposed method achieves reliable accuracy and satisfies the requirements for automatic alignment of single-angle components.