Sentinel Vision: YOLO Video Detector

This project includes a native desktop application for Windows. It runs video inference locally with a modern CustomTkinter interface and does not require a browser or a local web server.

This application supports exactly three processing modes:

• Model 1: uses models/best1.pt | best_aug_cv_2.pt only.
• Model 2: uses models/best2.pt | best_cv_5 only.
• Ensemble: runs best1.pt and best2.pt on the same frame and combines their predictions with Weighted Box Fusion.

The ensemble is logic, not a third model file.

Dataset classes

The supplied data.yaml defines:

• 0 = mask
• 1 = spectacle
• 2 = helmet
• 3 = balaclava

These classes are allowed to overlap in the initial configuration. No global cross-class suppression is applied because, for example, a helmet and spectacle can legitimately appear in the same region.

Run the desktop application

1. Put the trained weights in models/:

models/best1.pt
models/best2.pt

2. Create a virtual environment and install packages:

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

3. Start the desktop application:

python app.py

Select a video, choose Model 1, Model 2, or Ensemble, and press Analyze video. The result is previewed inside the application and saved in outputs/.

Run the legacy web application

The original browser interface remains available when needed:

python app.py --web

Then open http://127.0.0.1:5000.

On a Raspberry Pi, access it from another device using the Pi's LAN IP, for example:

http://192.168.1.50:5000

Raspberry Pi tuning

config.py contains the main tuning controls.

Start with:

FRAME_SKIP = 1

If inference is too slow on the Pi, try 2, 3, or 5. Skipped frames reuse the last detection result, so the output video remains annotated while the detector runs less frequently.

The current statistics count detection occurrences across analyzed frames, not unique physical objects. Unique-object counting requires an object tracker and should be added separately.