A New Pallet-Positioning Method Based on a Lightweight Component Segmentation Network for AGV Toward Intelligent...

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Automated Guided Vehicles (AGVs) face challenges in warehouses due to varied pallet sizes, hindering operational efficiency.
• Existing semantic segmentation models struggle to balance spatial details and high-level semantic information, leading to redundant computations.

Purpose of the Study:

• To propose a lightweight component segmentation network for precise pallet segmentation and localization.
• To address the limitations of existing models in handling diverse pallet shapes and sizes for automated picking.

Main Methods:

• A novel lightweight component segmentation network with a dual-attention mechanism and an encoder-decoder architecture.
• Integration of a residual structure to reduce network parameters and mitigate gradient issues.
• Utilizing dual-branch input images to extract multi-scale features for enhanced segmentation.

Main Results:

• Achieved precise segmentation of various pallet types using limited annotated images.
• Demonstrated robustness in pallet localization under varying illumination and background noise.
• Improved accuracy by 10.41% and image processing speed by 32.8% compared to traditional models.

Conclusions:

• The proposed network effectively segments and localizes multi-category pallets in complex warehousing environments.
• The method enhances AGV operational efficiency through accurate pallet identification and positioning.
• Validated robustness and performance in real-world warehousing scenarios with diverse conditions.