🪖 Helmet Violation Detection — NVIDIA AI City Challenge 2023 · Track 5

Detecting motorcycle helmet rule violations in real-time surveillance video — a 7-class object detection system built for the CVPR 20

📌 Overview

Motorcycle accidents are a leading cause of traffic fatalities, especially in developing countries where helmet compliance is inconsistently enforced. This project addresses Track 5 of the NVIDIA AI City Challenge 2023, which tasks participants with automatically detecting helmet violations for motorcyclists in surveillance footage.

The system identifies up to 7 object classes per frame — the motorcycle, driver, and up to two passengers — and classifies each rider as wearing or not wearing a helmet.

🏁 Challenge Details

🏷️ Detection Classes

🧠 Approach

Detection Pipeline

Raw Video Frames
      │
      ▼
 Frame Extraction (10 FPS)
      │
      ▼
 Object Detection (YOLOv8 / custom backbone)
      │
      ├──► Motorcycle BBoxes
      └──► Rider BBoxes + Helmet Classification
                │
                ▼
        Post-Processing (NMS, score thresholding)
                │
                ▼
        Submission File (.txt per video)

Key Techniques

• Multi-scale training — trained at multiple input resolutions to handle both distant and close-up riders
• Data augmentation — mosaic, mixup, random flip, HSV shifts, and copy-paste augmentation to improve generalization across varied lighting and occlusion conditions
• Pseudo-labeling — generated soft labels on unlabeled frames to expand effective training data
• Test-Time Augmentation (TTA) — horizontal flip and multi-scale inference averaged at prediction time
• Weighted Box Fusion (WBF) — ensemble-level bounding box merging across multiple model checkpoints to improve precision

📁 Repository Structure

aio_pending_track5/
├── data/
│   ├── raw/                    # Original dataset videos
│   ├── frames/                 # Extracted frames (10 FPS)
│   ├── annotations/            # YOLO-format labels
│   └── splits/                 # train / val split configs
├── configs/
│   ├── model.yaml              # Model architecture config
│   └── hyp.yaml                # Hyperparameter config
├── src/
│   ├── extract_frames.py       # Video → frame extraction
│   ├── train.py                # Training entry point
│   ├── detect.py               # Inference on test videos
│   ├── postprocess.py          # NMS, WBF, score filtering
│   └── utils/
│       ├── augmentation.py     # Custom augmentation helpers
│       └── submission.py       # Format output for AIC submission
├── notebooks/
│   ├── EDA.ipynb               # Dataset exploration
│   └── Evaluation.ipynb        # mAP analysis and error inspection
├── requirements.txt
└── README.md

⚙️ Setup & Usage

1. Install Dependencies

git clone https://github.com/tuanquang95/aio_pending_track5.git
cd aio_pending_track5
pip install -r requirements.txt

2. Prepare Data

Download the Track 5 dataset from the AI City Challenge portal and extract it:

# Extract video frames at 10 FPS
python src/extract_frames.py \
    --video-dir data/raw/videos/ \
    --output-dir data/frames/ \
    --fps 10

3. Train

python src/train.py \
    --data configs/model.yaml \
    --img-size 1280 \
    --batch-size 16 \
    --epochs 100 \
    --weights yolov8x.pt \
    --device 0

4. Inference & Generate Submission

python src/detect.py \
    --source data/raw/test_videos/ \
    --weights runs/train/exp/weights/best.pt \
    --conf 0.25 \
    --iou-thres 0.45 \
    --img-size 1280 \
    --output submissions/result.txt

📊 Results

Exact mAP scores on the official test leaderboard are tied to the challenge evaluation server. Validation set metrics are tracked in notebooks/Evaluation.ipynb.

🔍 Challenges & Observations

• Class imbalance — DHelmet instances vastly outnumber P2NoHelmet, requiring careful sampling strategies during training
• Small object detection — distant motorcycles with tiny rider bounding boxes were difficult to classify reliably; high-resolution inputs (1280+) were critical
• Occlusion — riders stacked on motorcycles heavily overlap each other, making passenger 1 and 2 classification particularly challenging
• Lighting variance — the dataset spans daytime, nighttime, and mixed-lighting scenes, requiring strong HSV augmentation

🛠️ Tech Stack

📚 References

• NVIDIA AI City Challenge 2023 — Official Site
• Track 5 Data & Evaluation Details
• Ultralytics YOLOv8 — github.com/ultralytics/ultralytics
• Solovyev et al., Weighted Boxes Fusion, 2021 — ensemble box merging for object detection
• Tsai et al., Video Analytics for Detecting Motorcyclist Helmet Rule Violations, CVPRW 2023

Built for the NVIDIA AI City Challenge 2023 CVPR Workshop · Track 5