Visual Sorting of Express Parcels Based on Multi-Task Deep Learning.

Area of Science:

• Robotics and Automation
• Computer Vision
• Artificial Intelligence

Background:

• Intelligent logistics sorting systems face challenges in accurately and rapidly sorting express parcels in complex, disorderly stacked scenes.
• Existing methods struggle to achieve both speed and precision for real-time parcel handling.

Purpose of the Study:

• To propose a novel robot sorting method leveraging multi-task deep learning for enhanced detection and efficient sorting of express parcels.
• To improve the real-time performance and accuracy of automated logistics sorting systems.

Main Methods:

• Development of a lightweight object detection network with scale variability and joint weights for model sparsification and channel identification.
• Implementation of pruning strategies to reduce model size and increase detection speed without compromising accuracy.
• Design of an optimal sorting position and pose estimation network using an end-to-end structure for real-time joint training of position and pose information.

Main Results:

• The proposed lightweight object detection network demonstrates improved real-time performance through model sparsification and pruning.
• The multi-task learning approach for position and pose estimation significantly enhances sorting accuracy.
• Robotic sorting experiments validated the method's high accuracy and real-time capabilities.

Conclusions:

• The multi-task deep learning robot sorting method effectively addresses the limitations of current systems for handling disorderly stacked express parcels.
• The integrated approach of lightweight detection and joint pose/position estimation offers a promising solution for intelligent logistics.