TrackVision AI

Multi-object tracking for uploaded videos, using YOLOv8 and ByteTrack in a Flask web app.

Example frame from the tracking workflow. Upload a video to produce your own annotated MP4.

Features

• Detects 80 COCO object classes with YOLOv8n and assigns persistent track IDs with ByteTrack.
• Accepts MP4, AVI, MOV, MKV, and WebM uploads up to 100 MB, and checks each file's contents before it is processed.
• Returns an annotated MP4 with a label, confidence score, and track ID on each object.
• Shows measured values for each video: unique track IDs, detections, frames processed, and processing time.
• Provides a JSON API with a health check, an OpenAPI 3.0 spec, and consistent JSON errors.
• Runs on CPU, loads the model once at startup, and downloads the weights on first start.

Quick start

git clone https://github.com/Shashwat-19/Object-Detection.git && cd Object-Detection
python3.11 -m venv .venv && source .venv/bin/activate && pip install -r requirements.txt
export PORT=8000 && gunicorn app:app --bind 127.0.0.1:$PORT --timeout 600 --workers 1 --worker-class gthread --threads 4

The first start downloads the model weights, yolov8n.pt (about 6.5 MB), into the project folder. Wait for Model loaded successfully in the log, then open http://127.0.0.1:8000.

On macOS, the pinned +cpu PyTorch builds are not available. Follow step 3 in the quickstart instead.

Example API call

Upload the sample clip. The server redirects to the result page, and the file parameter names the output:

curl -s -o /dev/null -D - -F video=@samples/sample.mp4 http://127.0.0.1:8000/upload

HTTP/1.1 302 FOUND
Content-Type: text/html; charset=utf-8
Location: /result/?file=track_sample_8a78d9517b_29dc62e9a8.mp4

Fetch the measured values for that video:

curl -s http://127.0.0.1:8000/static/outputs/track_sample_8a78d9517b_29dc62e9a8.json

{"frames_processed": 40, "total_detections": 133, "unique_track_ids": 15, "processing_seconds": 0.57}

Errors return JSON when the request asks for it (X-Requested-With: XMLHttpRequest or Accept: application/json):

{"error":"The file is not a supported video. Its contents do not match the file extension."}

Documentation

• Quickstart: from a fresh clone to a tracked video.
• API reference: every route, with examples and status codes.
• Technical report: architecture, model settings, measured performance, and known limitations.
• OpenAPI spec: the machine-readable contract.

Tech stack

Area	Tools
Detection and tracking	Ultralytics YOLOv8 (8.4.49), ByteTrack
Web app	Flask 3.1, Gunicorn 23
Computer vision	OpenCV, NumPy
Runtime	PyTorch 2.11 (CPU), Python 3.11
Front end	Bootstrap 5.3, Bootstrap Icons, vanilla JavaScript
Tests	pytest

Testing

pip install -r requirements-dev.txt
pytest

The tests use a fake model, so they do not download weights or run inference.

License

Released under the MIT License.

Author

Shashwat · github.com/Shashwat-19