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

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Pallet positioning in industrial automation and logistics is hindered by challenges in accuracy and efficiency, particularly with stacking and occlusion.
• Existing methods often struggle with complex environments, necessitating advanced solutions for precise pose estimation.

Purpose of the Study:

• To develop a high-precision pallet positioning method that overcomes accuracy and efficiency limitations in complex scenarios.
• To enhance the detection and localization accuracy of pallet key points and accurately compute the pallet's pose.

Main Methods:

• An improved YOLOv11s-pose architecture was utilized, incorporating transfer learning and optimized star operations.
• A dual-domain edge feature enhancement module and a region-focused topology attention mechanism within the C3k2 module were employed for multiscale feature extraction and fusion.
• An efficient perspective-n-point algorithm, combining visual weights and topological constraint optimization, was used for pose computation.

Main Results:

• The server-side model achieved 95.1% object detection accuracy, 94.2% key point localization accuracy, and 105.6 FPS.
• On the RK3568 development board, the model maintained detection accuracy with 44.1 FPS, inclination angle error < 2.9°, and horizontal error < 19 mm.

Conclusions:

• The proposed method significantly enhances the accuracy and stability of pallet positioning in complex environments.
• The integration of pose estimation with transfer learning demonstrates strong potential for industrial automation and logistics applications.