Multi-stage fusion of dual attention mask R-CNN and geometric filtering for fast and accurate localization of occluded apples

In unstructured orchard environments, factors such as complex lighting, fruit occlusion, and fruit clustering significantly reduce the accuracy of apple detection and 3D localization in robotic harvesting systems. To enhance the perception and reconstruction of occluded fruits, this paper proposes a multi-stage fusion framework for high-precision and robust processing, from image enhancement to 3D reconstruction. First, an adaptive image enhancement algorithm based on the HSV color space is employed to effectively alleviate image degradation caused by uneven lighting. Then, an improved Mask R-CNN with a dual-attention mechanism (SE-CBAM) is introduced to achieve scale-adaptive fruit segmentation under occlusion conditions. Next, a hierarchical point cloud purification strategy combining depth clustering and geometric feature analysis is applied to remove branch and leaf interference. Finally, a two-step RANSAC-LM spherical fitting algorithm is used to quickly and accurately recover the 3D shape of the fruit from incomplete point clouds. Experimental results show that the proposed method achieves Mask IoUs of 0.89, 0.82, and 0.65 under mild, moderate, and severe occlusion, respectively, with the lowest 3D localization error of 0.42 cm and an overall processing frame rate of 20 FPS. In real orchard environments, the harvesting success rate under occlusion conditions reaches up to 82.7 %, significantly outperforming traditional point cloud centroid and 2D positioning methods. This study provides an efficient, robust, and real-time deployable visual solution for fruit localization and robotic harvesting in complex orchard environments.