Pose Estimation Algorithm for Photovoltaic Junction Box Grasping Guided by Pixel-Level Segmentation Based on Deep Learning

In light of the accelerating global transition towards renewable energy, the significance of clean and efficient photovoltaic (PV) power generation is growing. The PV junction box, a crucial element in PV systems, encounters substantial hurdles in accomplishing exact identification and positioning during automated material handling operations, especially in unstructured settings characterized by stacking and occlusions. Traditional strategies for point cloud processing are frequently confined by their incapacity to achieve accurate segmentation, while point cloud registration methodologies are prone to encountering difficulties in delivering accurate pose information when faced with deformable items. To address these limitations, this paper introduces an original strategy for estimating the grasping pose of PV junction boxes, which is equipped with the capability to realize high-precision segmentation and pose estimation amidst complex occlusion circumstances. Experimental results confirm that, in comparison to conventional techniques, our proposed framework not only substantially boosts the accuracy of identification and localization endeavors but also drastically minimizes processing time, leading to a considerable augmentation in the success rate of robotic maneuvers targeted at grasping PV junction boxes.