Unmanned forklift pallet positioning algorithm based on an improved human pose estimation model

Pallet positioning often faces challenges in accuracy and efficiency within complex scenarios involving stacking and occlusion. This study presents a high‐precision pallet positioning method based on an improved YOLOv11s‐pose architecture, integrating transfer learning and optimized star operations with a dual‐domain edge feature enhancement module for efficient multiscale feature extraction and fusion. By incorporating a region‐focused topology attention mechanism into the cross‐stage partial module with kernel size 2 (C3k2), the detection accuracy of the 12 key points defining the E‐shaped cross‐section of the pallet is significantly enhanced. Additionally, the efficient perspective‐n‐point algorithm, which combines visual weights with topological constraint optimization, accurately computes the pallet's pose. Experimental results demonstrate that the server‐side model achieves an object detection accuracy of 95.1%, key point localization accuracy of 94.2%, and a processing speed of 105.6 frames per second (FPS). When deployed on the RK3568 development board, the model maintains consistent detection accuracy, achieving an actual FPS of 44.1, an inclination angle error below 2.9°, and a horizontal error under 19 mm. This method, integrating pose estimation with transfer learning, significantly enhances the accuracy and stability of pallet positioning in complex environments, showing strong potential for applications in industrial automation and logistics.