AmineSam/irail-crowd-counting-yolov8n

iRail Crowd Counting (YOLOv8n) — Head Detection Baseline

This repository contains a YOLOv8n (nano) model fine-tuned for pedestrian head detection in crowded environments (railway platforms + event entrances).
The main objective is crowd counting per frame by detecting heads and counting bounding boxes.

This model is designed as an educational baseline for an iRail/Azure project extension where the iRail API does not provide occupancy, so we estimate a proxy crowd level from images/frames.

Model

• Architecture: YOLOv8n (Ultralytics)
• Task: Object detection
• Class(es): head (single class)
• Training input size: imgsz=832
• Baseline inference settings used for counting eval:
  • conf=0.25
  • iou=0.75
  • max_det=300

Dataset

Trained on RPEE-Heads (Railway Platforms and Event Entrances-Heads) with head bounding boxes.

• 1,886 images
• 109,913 head annotations
• Split:
  • Train: 1,346 images
  • Val: 246 images
  • Test: 294 images
• License: CC BY-SA 4.0 (dataset + derived model share-alike requirements apply)

Paper reference: RPEE-Heads Benchmark: A Dataset and Empirical Comparison of Deep Learning Algorithms for Pedestrian Head Detection in Crowds
Mohamad Abubaker, Zubaida AlSadder, Hamed Abdelhaq, Maik Boltes, Ahmed Alia
DOI: 10.34735/ped.2024.2
Dataset URL: http://ped.fz-juelich.de/da/2024rpee_heads

Baseline Evaluation (Detection)

Ultralytics validation metrics on Val/Test (single class: head):

Validation (246 images / 16,022 instances)

• Precision: 0.910
• Recall: 0.805
• mAP@0.50: 0.881
• mAP@0.50:0.95: 0.522

Test (294 images / 15,285 instances)

• Precision: 0.908
• Recall: 0.803
• mAP@0.50: 0.878
• mAP@0.50:0.95: 0.515

Crowd Counting Evaluation (Counting boxes per image)

Counting is computed as:
predicted_count = number of detected boxes (after NMS at chosen conf and iou).

Baseline counting settings:

• imgsz=832, conf=0.25, iou=0.75

Metrics

• MAE: 4.67
• RMSE: 8
• Bias (pred - gt): 0.097 (slight undercount)

Intended Use

• Educational demo of:
  • fine-tuning YOLOv8 for head detection
  • evaluating detection metrics
  • converting detections to a crowd count proxy
• A building block for a larger iRail/Azure pipeline (occupancy proxy)

Limitations

• Not trained on Belgian station camera viewpoints specifically.
• Counting via “number of boxes” can undercount in very dense crowds.
• Domain shift (camera height, lens distortion, resolution, lighting) may reduce performance.

Author

Amine Samoudi - GitHub: @AmineSam