MBFILNet: A Multi-Branch Detection Network for Autonomous Mining Trucks in Dusty Environments

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Autonomous mining trucks require accurate object detection for safety and reliability.
• Dusty open-pit environments pose significant challenges to current object detection systems.
• Existing methods struggle with performance degradation due to dust interference.

Purpose of the Study:

• To develop an advanced object detection network resilient to dust interference for autonomous mining trucks.
• To improve the accuracy and reliability of object detection in challenging dusty environments.
• To introduce a novel network architecture that enhances feature extraction and spatial dependency modeling.

Main Methods:

• Proposed a multi-branch feature interaction and location detection network (MBFILNet).
• Incorporated multi-branch feature interaction with differential operation (MBFI-DO) for channel-wise feature extraction.
• Utilized depthwise separable convolution-enhanced non-local attention (DSC-NLA) for long-range spatial dependencies.
• Constructed and augmented a custom Dusty Open-pit Mining (DOM) dataset using CycleGAN.

Main Results:

• MBFILNet achieved a mean Average Precision (mAP) of 72.0% on the DOM dataset.
• Demonstrated a 1.3% mAP increase compared to the Featenhancer model.
• Showcased a 2% mAP increase over YOLOv8, indicating superior performance in dusty conditions.

Conclusions:

• MBFILNet effectively improves object detection accuracy in dusty open-pit mining environments.
• The proposed network architecture addresses the limitations of existing methods in adverse weather conditions.
• The study validates the efficacy of MBFILNet for enhancing the safety and reliability of autonomous mining operations.