A computer vision system for real-time construction site safety monitoring. Built for CSCI435 — Computer Vision Algorithms and Systems, University of Wollongong in Dubai, Spring 2026.
SiteWatch ingests live camera or recorded footage from a construction site and continuously flags safety violations: workers without hard hats or safety vests, workers in proximity to moving machinery, and unattended hazards. Violations are overlaid on the video with bounding boxes, surfaced as a critical-alert banner, and logged with timestamps to a CSV for later review.
A 5-second clip from tests/fixtures/demo.mp4 processed through the live pipeline at 640 long edge. Green boxes are PPE-compliant detections (Hardhat, Safety Vest); red are violations (NO-Hardhat, NO-Safety Vest, NO-Mask); blue are workers; the bottom banner shows the highest-priority active alert.
SiteWatch integrates four distinct CV capabilities into a single pipeline:
| Capability | Implementation | Purpose |
|---|---|---|
| Image enhancement | CLAHE on LAB color space, optional bilateral denoise | Normalises harsh outdoor lighting and dust before detection |
| Object detection | YOLOv8n (pre-trained COCO baseline) + YOLOv8m custom fine-tuned on construction safety dataset, merged with IoU-based deduplication | Detects workers, PPE compliance, machinery, vehicles, and safety cones |
| Shape detection | HoughCircles (hard-hat confirmation) + HoughLines/HSV (safety cones) | Cross-validates YOLO detections with classical CV |
| Motion / change detection | MOG2 background subtraction + binary morphological operations | Flags moving machinery; raises alerts when workers approach machinery zones |
The motion module additionally satisfies the brief's "binary morphological operations" capability through its cv2.morphologyEx open-then-close cleanup of the foreground mask.
Camera / upload → Streamlit frontend → CV pipeline → Decision fusion → Visual overlay + CSV log
↓
[enhance → detect → shapes → motion]
See the report for the full architecture diagram.
- Python 3.10 or newer
- A working webcam (for live mode) — built-in laptop camera is fine
- ~2 GB free disk space for models and dependencies
git clone https://github.com/<your-username>/sitewatch.git
cd sitewatch
python -m venv .venv
source .venv/bin/activate # macOS / Linux
.venv\Scripts\activate # Windows PowerShell
pip install -r requirements.txtrequirements.txt pulls the CPU-only PyTorch wheel by default, which works everywhere but limits the live tab to ~3–5 FPS. If you have an NVIDIA GPU with CUDA drivers installed, replace the torch install with the matching CUDA wheel for a ~10× speed-up:
pip uninstall -y torch torchvision torchaudio
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu124(Adjust cu124 to match your installed CUDA version — see PyTorch's install matrix. The reported benchmarks below were measured on the CUDA 12.4 build.)
streamlit run app.pyThe app opens in your default browser at http://localhost:8501. If encountering errors, make sure to run the streamlit within the .venv you created by activating it following the steps mentioned before.
- Live monitoring tab: grant camera access; the system processes your webcam feed in real time.
- Review footage tab: upload a
.jpg,.png,.mp4, or.avifile for offline analysis.
The custom fine-tuned YOLOv8m weights (models/sitewatch_best.pt) are bundled with the repository, so PPE classes (Hardhat, NO-Hardhat, Safety Vest, etc.) work immediately. The COCO YOLOv8n baseline (models/yolov8n.pt) is auto-downloaded on first detector init. To re-train the custom model from scratch, see "Training the custom model" below.
The custom YOLO is fine-tuned on the Construction Site Safety Image Dataset from Roboflow (2,801 images, 25 classes including the 10 safety classes plus machinery / vehicle subtypes). Training the YOLOv8m configuration takes ~60–90 minutes on a free Google Colab T4 GPU.
- Open
training/train_yolo.ipynbin Google Colab - Enable T4 GPU (Runtime → Change runtime type → T4)
- Replace
YOUR_ROBOFLOW_API_KEYwith your free Roboflow account key - Run all cells — the notebook downloads the dataset, fine-tunes YOLOv8m for up to 200 epochs (with
patience=30early stopping), validates, and producesbest.pt - Download
best.ptand place it atmodels/sitewatch_best.pt - Restart the Streamlit app — custom detections will now activate
Validation results from our training run are summarised in the Performance section below; the full per-class breakdown and training curves are in the project report.
sitewatch/
├── app.py # Streamlit entry point
├── cv_pipeline/ # The four CV capabilities
│ ├── enhancement.py # CLAHE + denoise
│ ├── detection.py # YOLO wrapper (COCO + custom)
│ ├── shapes.py # Hough-based shape validation
│ ├── motion.py # MOG2 + morphology + proximity
│ └── fusion.py # Decision fusion + prioritised alerts
├── utils/ # Drawing, benchmarking
├── training/train_yolo.ipynb # Colab fine-tuning notebook
├── models/ # YOLO weights (COCO + custom)
├── tests/ # Pytest suite
├── logs/ # Auto-generated violation logs
└── benchmarks/ # Auto-generated performance reports
Benchmarks on a Windows 11 laptop with an NVIDIA GeForce RTX 4060 Laptop GPU (8 GB VRAM), CUDA 12.4 PyTorch build, measured on the bundled tests/fixtures/demo.mp4 (4K source, 30 sec, 900 frames). The custom backbone is yolov8m fine-tuned on the Construction Site Safety dataset.
| Metric | Live tab (640 long edge) | Review tab (1280 long edge) |
|---|---|---|
| Average FPS (whole pipeline) | 11.68 | 5.40 |
| Total latency p50 / p95 | 64.5 / 77.6 ms | 131.7 / 221.0 ms |
| Detection latency p50 / p95 | 48.0 / 60.6 ms | 67.3 / 115.6 ms |
Custom YOLOv8m validation results (114 val images, 733 instances):
| Metric | Value |
|---|---|
| mAP@0.5 | 0.563 |
| mAP@0.5–0.95 | 0.400 |
| Precision / Recall | 0.755 / 0.482 |
| Hardhat mAP@0.5 | 0.741 |
| NO-Hardhat mAP@0.5 | 0.578 |
| Safety Vest mAP@0.5 | 0.696 |
| NO-Safety Vest mAP@0.5 | 0.630 |
| Person mAP@0.5 | 0.766 |
| Safety Cone mAP@0.5 | 0.884 |
| machinery mAP@0.5 | 0.761 |
To reproduce: python -m utils.benchmark tests/fixtures/demo.mp4 --max-long 640. The --max-long flag mirrors the review-tab _resize_long_edge in app.py; pass 0 to benchmark at native resolution.
pytestThe test suite covers each pipeline module with smoke tests against a fixture image.
The current prototype operates on a single camera feed and produces immediate, rule-based alerts. Aggregating observations across multiple cameras and over time would enable a temporal risk model that estimates incident likelihood per zone in real time, allowing supervisors to intervene before a near-miss occurs. Other directions include integration with site access systems to gate entry on PPE compliance, multi-camera worker tracking, and edge deployment on Raspberry Pi or Jetson Nano hardware. See the report for a fuller discussion.
- Dataset: Construction Site Safety Image Dataset by Roboflow Universe Projects, used under the dataset's open license.
- YOLOv8: Ultralytics YOLOv8 (AGPL-3.0)
- OpenCV: opencv.org
- Streamlit + streamlit-webrtc: streamlit.io, whitphx/streamlit-webrtc
| Name | Student ID | Role |
|---|---|---|
| Hani Moustafa | 8960215 | Lead Developer & Project Manager |
| [Teammate 2] | [ID] | [Role] |
| [Teammate 3] | [ID] | [Role] |
| [Teammate 4] | [ID] | [Role] |
CSCI435 — Spring 2026 — Dr. Patrick Mukala
Original source code in this repository is released under the MIT License — see LICENSE for the full text. Bundled and referenced components (Ultralytics YOLOv8 under AGPL-3.0, the Roboflow Construction Site Safety dataset, the COCO dataset) remain subject to their own licenses; the LICENSE file enumerates them. Submitted as coursework for CSCI435 at UOWD.