Dark-SORT: Multi-Person Tracking in Underground Coal Mines Using Adaptive Discrete Weighting

Rui Wang, Jingzhao Li, Zhi XuIEEE Access 2023 · DOI

Abstract

Underground coal mines present a "visual nightmare" for standard multi-object tracking (MOT) systems due to extreme low-light conditions, heavy dust, and blinding point sources. This paper introduces Dark-SORT, a specialized MOT framework designed to overcome these environmental hurdles. By combining a two-step image enhancement process, a novel "discrete-weighted" attention mechanism (ADAM), and an improved motion-plus-appearance tracking model (SOC-SORT), Dark-SORT achieves state-of-the-art performance with a 92.6 MOTA on custom mine datasets while maintaining real-time viability.

Introduction

Multi-Object Tracking (MOT) has advanced significantly, yet most state-of-the-art models are optimized for well-lit, clear environments. In the context of underground coal mines, these models frequently fail due to three primary factors:

  1. Extreme Lighting: Narrow corridors are often pitch black, punctuated by overexposed lamps that create significant glare.
  2. Visual Noise: High concentrations of dust and motion blur degrade feature extraction.
  3. Frequent Occlusions: Narrow spaces and large machinery lead to frequent identity (ID) switches when trackers lose targets.

Dark-SORT addresses these challenges by preprocessing visual data and refining the attention and tracking mechanisms to prioritize information-rich features over environmental noise.

Methodology

The Dark-SORT architecture introduces three core innovations to the standard tracking-by-detection pipeline.

Image Enhancement (IE)

To mitigate the effects of poor lighting, images undergo a two-step cleanup before reaching the detector. First, L-CLAHE (Lightness-based Contrast Limited Adaptive Histogram Equalization) enhances contrast without amplifying noise. Second, a modified 2D Gamma Function (referred to as V-BC in processing) balances illumination by brightening dark regions while dimming overexposed light sources.

The Image Enhancement (IE) pipeline, illustrating the transition from a low-visibility input to a balanced output suitable for detection.
Detailed internal logic of the L-CLAHE and V-BC processes used for contrast enhancement and illumination correction.

Adaptive Discrete-weighted Attention Module (ADAM)

Standard attention modules like CBAM often rely on Global Average Pooling (GAP), which can smooth over critical details in low-contrast images. Dark-SORT utilizes Global Discrete Pooling (GDP), which focuses on the variance of pixels to identify edges and textures. This is implemented via:

  • EDCA (Channel Attention): Prioritizes channels containing high-variance, information-rich data.
  • ADSA (Spatial Attention): Dynamically highlights targets while ignoring overexposed "light blobs."

SOC-SORT (Supplementary-OC SORT)

Building on the OC-SORT motion model, Dark-SORT integrates appearance features (via cosine distance) and Mahalanobis distance into the cost matrix. This hybrid approach allows the tracker to maintain identity even when motion becomes non-linear or visibility is intermittent.

Results and Discussion

The system was evaluated using the custom MINE-TD and MINE-MOT datasets, consisting of 10,000 frames of real-world coal mine footage.

  • Tracking Accuracy: Dark-SORT achieved a MOTA of 92.6 and a HOTA of 67.4, significantly outperforming benchmarks like FairMOT and ByteTrack.
  • Detection Performance: The integration of the ADAM module and Image Enhancement boosted the base YOLOv7 detector's Average Precision (AP) from 80.17% to 92.41%.
  • Efficiency: The framework operates at 23 FPS, making it suitable for near-real-time safety monitoring in industrial settings.

Conclusion

Dark-SORT demonstrates that specialized preprocessing and attention mechanisms can bridge the gap for MOT in extreme environments. While the system effectively handles lighting and noise, the challenge of "Similar Attire"—where workers in identical uniforms are difficult to distinguish—remains an area for future research, potentially involving gait analysis or equipment-specific markers.

Summary updated

BibTeX
@article{wang2023dark,
  author = {Rui Wang and Jingzhao Li and Zhi Xu},
  title = {{Dark-SORT: Multi-Person Tracking in Underground Coal Mines Using Adaptive Discrete Weighting}},
  journal = {IEEE Access},
  year = {2023},
  doi = {10.1109/access.2023.3340914},
}
Figure