Open-source, real-time video analytics engine optimized for edge devices
Turn any camera into a smart sensor — object detection, tracking, line crossing, and zone counting at the edge.
Locus Vision is a self-hosted video analytics platform built for edge devices. It performs real-time object detection and tracking using YOLO + ONNX Runtime, with support for INT8/FP16 quantized models that cut model size by 70% while maintaining accuracy. Everything stays local — no cloud, no subscriptions, no data leaving your network.
demo.mp4
- 🎯 Detection & Tracking — YOLO models (v5–v11) via ONNX Runtime + ByteTrack multi-object tracking across concurrent streams
- 🔲 Spatial Analytics — Polygon zones for occupancy counting, directional lines for crossing detection, dwell time, capacity alerts
- 📹 Live Streams — MJPEG video with overlaid detections, real-time heatmaps, and event feeds
- 🧠 Adaptive Models — Automatically detects hardware (Hailo, CUDA, CoreML, CPU ARM) and picks the optimal model format. Download models with one click from Settings.
- 📊 Rich Analytics — Peak hours analysis, hourly aggregations, CSV/JSON export, Prometheus metrics endpoint
- 🎞️ Batch Processing — Upload videos for offline analysis with a crash-resilient job queue
- 📷 Camera Flexibility — RTSP streams, ONVIF auto-discovery, USB webcams, V4L2 hardware decoding
- 🔒 Access Control — JWT authentication, role-based access (admin/viewer), rate-limited login
- 📈 System Insights — Per-camera FPS breakdown, CPU/memory/storage monitoring, data archival with Parquet
- Retail & Foot Traffic — Count visitors entering/exiting zones, measure dwell time across store areas
- Warehouse & Logistics — Monitor loading docks, track forklift movement, detect restricted zone breaches
- Parking & Access Control — Track occupancy in real time, log vehicle entry/exit with line crossing
- Wildlife & Environmental — Observe animal activity patterns, generate spatial heatmaps from trail cameras
| Layer | Technology |
|---|---|
| Frontend | SvelteKit 2, Svelte 5, TypeScript |
| Styling | Tailwind CSS 4, shadcn-svelte |
| Backend | FastAPI, Python 3.11+ |
| Database | SQLite (async), DuckDB (analytics) |
| AI / Vision | ONNX Runtime, YOLO (v5–v11), ByteTrack, OpenCV, Hailo-8L |
| Auth | JWT + Argon2id |
git clone https://github.com/kongesque/locus-vision.git
cd locusvision
docker compose up --buildOpen localhost:3000 (app) or localhost:8000/api/docs (API docs).
# Clone & install
git clone https://github.com/kongesque/locus-vision.git
cd locusvision && pnpm install
# Setup backend
cd backend && python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt && cd ..
# Run
pnpm devOpen localhost:5173 (app) or localhost:8000/api/docs (API docs).
Models are managed from Settings > Models in the UI. The system auto-detects your hardware (Hailo, CUDA, CoreML, or CPU) and shows which model formats are compatible.
Downloading models:
- Click "Download" next to any model to fetch it from GitHub Releases
- The best precision for your hardware is automatically selected (INT8 on Pi, FP32 on Mac/desktop)
- No additional dependencies required on your machine
Current models: YOLO11 (n/s/m/l/x), YOLOv8 (n/s), YOLOv6n, YOLOv5 variants, YOLOX. See MODELS.md for the full list and available formats per model.
For developers: To add new models to the release, export them with the CLI tool (requires ultralytics on your dev machine only):
source backend/.venv/bin/activate
pip install ultralytics
python backend/scripts/export_model.py yolo11n --int8
# Then upload to the GitHub release and update model_catalog.json with the download URLAll numbers are inference-only (ONNX Runtime, 640×640 input, 15s timed run after 3s warmup). End-to-end stream FPS is lower due to video decode, tracking, and analytics overhead.
| Model | Size | FPS | Latency (median) | Latency (p99) |
|---|---|---|---|---|
| YOLO11n | 10 MB | 200 | 5.0 ms | 6.6 ms |
| YOLO11s | 36 MB | 118 | 8.5 ms | 9.9 ms |
| YOLO11m | 77 MB | 62 | 16.1 ms | 19.4 ms |
FP32, CoreML (Apple Neural Engine), ONNX Runtime 1.24.2. 15s timed run after 3s warmup, 2s cooldown between models. FPS derived from median latency.
| Model | Precision | Size | FPS | Latency (median) | Latency (p99) |
|---|---|---|---|---|---|
| YOLO11n | FP32 | 10 MB | 5 | 188.9 ms | 275.5 ms |
| YOLO11n | INT8 | 3 MB | 14 | 69.9 ms | 93.7 ms |
ONNX Runtime 1.24.4. Static INT8 quantization with QDQ nodes. 15s timed run after 3s warmup, 2s cooldown between models. FPS derived from median latency.
| Model | Size | FPS | HW Latency |
|---|---|---|---|
| YOLOv6n | 14 MB | 355 | 7.0 ms |
| YOLOv5n-seg | 7 MB | 65 | 14.2 ms |
| YOLOv8s | 35 MB | 59 | 13.1 ms |
| YOLOv5s-personface | 26 MB | 64 | 13.7 ms |
| YOLOX-s | 21 MB | 61 | 14.8 ms |
| YOLOv8s-pose | 23 MB | 51 | 18.9 ms |
HailoRT 4.23.0, HEF format.
hailortcli benchmark, streaming mode (includes host-side DMA transfer). 15s per model. End-to-end app FPS will be lower due to host-side preprocessing (letterbox resize, normalization) and postprocessing (NMS, tracking, zone analytics). YOLO11 HEF models not yet available — compiled from pre-built Hailo model zoo.
To reproduce:
# CPU benchmark
cd backend && source .venv/bin/activate
python scripts/benchmark_inference.py
# Hailo benchmark
hailortcli benchmark /usr/share/hailo-models/<model>.hefContributions welcome — bug reports, feature requests, and pull requests all help. See open issues for where to start.
Made with ❤️ by kongesque