vivekvar/helmet-v4

Helmet v4 — single-stage YOLO helmet detector

Predecessor to vivekvar/helmet-v5. Single YOLO model trained directly on helmet/no-helmet crops from Andhra Pradesh RTGS CCTV.

Files

• best.pt — best YOLO checkpoint from training
• train_helmet_v4.py — training script
• training_log.json, train.log — metrics
• crops/ — training crop dataset (with labels.db SQLite)

Why superseded

v4 was a single-stage detector — the model had to simultaneously find the rider and classify helmet presence. This led to:

• Wrong-body-part boxing on multi-bike frames
• Helmets misclassified as no-helmet on profile / back views
• Accuracy plateau at F1 ≈ 0.53

v5 splits this into detect (YOLO fine-tuned) → localize driver head (pose) → classify (EfficientNet-B0), hitting F1 = 0.864 on head crops and mAP50 = 0.979 on detection.