MineSDS: A Unified Framework for Small Object Detection and Drivable Area Segmentation for Open-Pit Mining Scenario

Area of Science:

• Computer Vision
• Robotics
• Autonomous Systems

Background:

• Mining environments present unique challenges for autonomous systems, including dense small objects and poorly defined drivable areas.
• Existing perception systems struggle with the unstructured and complex nature of open-pit mines, impacting safety and efficiency.
• Multi-task learning in convolutional networks shows promise for enhancing perception capabilities in autonomous driving.

Purpose of the Study:

• To develop an end-to-end framework for simultaneous small object detection and drivable area segmentation in mining scenarios.
• To improve the accuracy and robustness of perception systems for autonomous vehicles operating in open-pit mines.
• To address limitations in detecting small objects and segmenting fuzzy road boundaries in unstructured environments.

Main Methods:

• Utilized a convolutional network backbone for feature extraction in a multi-task learning approach.
• Introduced a lightweight attention module to enhance small object detection by focusing on spatial and channel dimensions.
• Employed a convolutional block attention module in the segmentation subnetwork to improve road boundary feature mapping.
• Implemented weighted summation in the loss function to boost perception accuracy for both tasks.

Main Results:

• Achieved an 87.8% mean Average Precision (mAP) for object detection on the Minescape dataset, surpassing state-of-the-art by 9.3%.
• Improved drivable area segmentation by 1% in Mean Intersection over Union (MIoU) compared to existing algorithms.
• Demonstrated competitive performance and enhanced perception accuracy for both detection and segmentation tasks.

Conclusions:

• The proposed end-to-end framework effectively addresses the challenges of small object detection and drivable area segmentation in mining.
• The integration of attention modules and weighted loss functions significantly improves perception accuracy in complex mining environments.
• The framework offers a robust solution for enhancing the safety and efficiency of autonomous operations in open-pit mines.