Hands‑On Underground Personnel Control in Coal Mines with YOLOv8 and ReID

1. Why Coal‑Mine Scenarios Need "Recognition"

Regulatory pressure on mine safety is increasing, yet traditional methods rely on manual supervision or card‑swipe checkpoints that cannot reveal a worker’s exact location, unauthorized zone entry, or single‑person operations. The goal is to make the underground "see people, recognize people, and control people".

2. Overall Technical Architecture

The solution consists of three layers: a collection layer (explosion‑proof cameras), an edge‑computing layer (boxes running YOLOv8 detection and ReID feature extraction), and a platform layer (business logic, storage, and visualization).

井下摄像头 → 边缘计算盒子 → 识别结果上传 → 平台端业务处理
    │               │               │
 视频采集      YOLOv8检测      身份匹配、轨迹追踪
               ReID特征提取   报警推送、数据存储
               边侧推理      可视化大屏

3. YOLOv8: Solving "Where Is the Person and What Is He Doing"

YOLOv8 was chosen for its speed, accuracy, and lightweight footprint. Two tasks are performed:

Personnel detection and attribute recognition (helmet wearing, reflective vest).

Violation detection (danger‑zone intrusion via electronic fences, single‑person work by combining time‑window rules with business logic).

The model was fine‑tuned on a self‑built dataset of ~20,000 annotated underground images covering diverse lighting and angles. Helmet‑detection accuracy exceeds 95% with a false‑positive rate under 3%; early versions at 88% were improved by adding synthetic mine‑lamp illumination during data augmentation.

4. ReID: Solving "Who Is the Person"

Because faces are often obscured by dust, masks, and harsh lighting, ReID (person re‑identification) based on overall appearance (body shape, clothing color, gait) is used. A lightweight OSNet model (input 256×128, 512‑dimensional feature) runs on the edge box, achieving 78 % mAP and 89 % Rank‑1 on the mine dataset with ~35 ms inference on a Jetson Nano.

Cross‑camera matching uses three strategies:

Time‑window constraints based on camera spacing and walking speed.

Cosine similarity of feature vectors with a threshold of 0.65 (tuned to balance miss and false matches).

Trajectory smoothing to correct jumps and gaps.

5. Edge‑Computing Deployment

Hardware options were compared:

Jetson Orin Nano – 40 TOPS, 15 W, ~¥2,500, requires external explosion‑proof housing.

Rockchip RK3588 – 6 TOPS, 8 W, ~¥800, also needs housing.

Custom explosion‑proof AI box – 21 TOPS, 25 W, ~¥4,500, already certified.

The custom AI box was selected to meet Ex d IIB T4 certification without a lengthy retro‑fit process. Each box supports up to four cameras, with per‑stream inference latency under 200 ms.

Software deployment uses TensorRT: models are exported from PyTorch to ONNX, then converted to TensorRT engines. An automation script can bring a new node online in ten minutes. Video streams stay on‑site; only JSON‑formatted detection results are sent over an industrial Ethernet ring, keeping bandwidth minimal.

6. Cost Breakdown

For a medium‑size mine (≈200 workers, 30 cameras):

Explosion‑proof cameras: ¥3,500 × 30 = ¥105,000

AI edge boxes: ¥4,500 × 8 = ¥36,000

Explosion‑proof switches: ¥8,000 × 4 = ¥32,000

Fiber & cabling: ≈¥50,000

Server (GPU‑enabled): ¥45,000

Software development (YOLOv8 fine‑tuning, ReID training, platform): ≈¥30 M

System integration & testing: ¥50,000

Total hardware cost ≈ ¥268,000; total software cost ≈ ¥350,000; first‑time investment ≈ ¥620,000. Annual O&M cost ≈ ¥95,000 (model updates, hardware warranty, cloud resources).

7. Lessons Learned (Pitfalls)

Data‑annotation quality is the ceiling – poor outsourced labeling caused mis‑learned features; a three‑round in‑house QA was required.

Underground lighting varies dramatically; extensive illumination‑augmentation was essential for robustness.

ReID accuracy drops due to uniform miner uniforms and dust; periodic feature‑library updates and auxiliary cues (lamp IDs, work‑cards) help.

Explosion certification is a hard gate – equipment must be certified before deployment to avoid wasted development.

Conclusion

YOLOv8 + ReID, when engineered with careful data preparation, edge‑box selection, and robust deployment pipelines, can reliably support underground personnel safety in coal mines. The real challenge lies not in model accuracy but in the surrounding engineering – data, deployment, and maintenance – to turn a prototype into a production‑grade system.

Illustration