PES-3D: A stereo vision-based 3D spherical crop detector with prior-enhanced semantic shape modelling

To address the challenges of occlusion, dense foliage, and visual clutter encountered in real-world agricultural production environments, this study proposes PES-3D (Prior-Enhanced Stereo 3D Detector), a cost-efficient instance-level 3D perception framework designed for spherical crop localisation in harvesting and post-harvest operations. PES-3D integrates stereo vision with category-specific shape priors and a specialised 2D feature extraction backbone to enable effective cross-modal alignment between image and point cloud representations. An enhanced 3D localisation module is further introduced to improve spatial accuracy and stability under complex field conditions. Experiments conducted on a farm-scale dataset collected from commercial crop field demonstrate that PES-3D achieves mAR/mAP scores of 83.33/67.36% at IoU=0.5 and 62.50/38.01% at IoU=0.7, outperforming representative baselines such as FCAF3D and FF3D. Consistent performance gains are observed across scenarios involving ground collection, cluttered backgrounds, and partial occlusions. Additional evaluation on the KITTI benchmark confirms the method’s cross-domain generalisation capability. Preliminary field evaluations suggest that PES-3D provides more accurate crop contour perception for robotic localisation and grasp planning. This improvement is expected to support more stable manipulation and potentially reduce mechanical damage in agricultural harvesting and sorting workflows, although comprehensive system-level validation remains ongoing.