CBAM-YOLOv11 and Geometric Constraint-Enhanced PnP for High-Precision EV Charging Port Pose Estimation

Abstract

The precise detection and pose estimation of electric vehicle (EV) charging ports in unstructured outdoor environments remain challenging due to small sizes, variable illumination, and stringent tolerance requirements for robotic plug-in operations. To address these issues, this paper presents a hybrid perception framework that integrates an attention-embedded detection network with geometrically constrained pose optimization. For robust detection, CBAM-YOLOv11 is proposed, which incorporates a sequential channel-spatial attention module into the backbone network to enhance feature representation of texture-less small targets while suppressing background clutter and glare. Then, a topological geometric constraint-based method is developed for accurate pose estimation. Specifically, the 2D-3D correspondences are purified before being fed into an Efficient Perspective-n-Point (EPnP) solver, while a nonlinear refinement with rigid distance priors is applied as regularization. Extensive experiments on the dataset and a physical robotic platform demonstrate that the proposed detector achieves 99.2% mAP@0.5 and a 24.6 percentage point improvement in mAP@0.5:0.95 over the baseline YOLOv11. The pose estimation module reduces positioning standard deviations along the X, Y, and Z axes to 4.72 mm, 5.65 mm, and 5.60 mm, respectively, surpassing conventional EPnP by about 60%. In 30 repeated robotic insertion trials, the system attains a 93.3% success rate with approximately 78 ms, fully satisfying real-time and precision requirements for autonomous EV charging.