Research on dimension measurement algorithm for parcel boxes in high-speed sorting system.

Area of Science:

• Computer Vision
• Machine Learning
• Logistics Technology

Background:

• Traditional manual sorting is inefficient and costly, struggling with the demands of the modern logistics industry.
• Existing parcel box detection algorithms face challenges in balancing accuracy, efficiency, and deployment costs.

Purpose of the Study:

• To propose a novel 3D localization algorithm for rectangular packaging boxes using deep learning.
• To design a lightweight and efficient parcel box detection model, the Efficient Object detection Network (EODNet).

Main Methods:

• Developed a deep learning-based 3D localization algorithm for rectangular packaging boxes.
• Designed the Efficient Object detection Network (EODNet) incorporating a linear attention mechanism in the backbone.
• Implemented a high-low layer feature fusion structure and C2f-GhostCondConv in the model's neck for selective feature fusion.

Main Results:

• The EODNet model demonstrated efficient feature selection with low computational cost.
• Achieved selective fusion of features with a small parameter count and computational load.
• Verified model effectiveness and universality on packing box and public datasets.
• Attained high accuracy in parcel box size prediction with an average error below 3.7%.

Conclusions:

• The proposed deep learning algorithm and EODNet effectively address the limitations of traditional sorting and existing detection methods.
• The model achieves high accuracy and speed (under 10ms), making it suitable for real-world logistics applications.
• The lightweight design and efficient mechanisms contribute to a balance between performance and deployment cost.