What This Solution Does
Industrial sites need eyes on restricted zones around the clock — but cloud-based AI surveillance means video leaves the premises, latency spikes, and a dropped network connection blinds the whole system. This solution runs AI person detection and rule-checking entirely on a reComputer Industrial (Jetson Orin) at the edge: video never leaves the local network, alerts fire in under a second, and it keeps working when the internet doesn't.
Core Benefits
| Benefit | Details |
|---|
| Data stays on-site | Video frames are processed locally on the reComputer — nothing is uploaded, no cloud account, no data egress |
| Sub-second alerts | The Jetson GPU analyzes every frame at 15-30 FPS, intrusions are flagged within ~1 second |
| Flexible safety rules | Draw restricted zones, virtual fence lines, and dwell-time triggers directly in the browser — no code |
| Works offline | Once deployed, the system runs without internet — perfect for remote sites, factories, mines |
| Standard cameras | Plug in any RTSP IP camera (Hikvision, Dahua, ONVIF generic) — no proprietary hardware lock-in |
| Multi-camera support | Monitor multiple areas simultaneously with independent processing pipelines sharing one detection model |
| Event persistence | SQLite database stores all events with date filtering — restart without losing history |
Use Cases
| Scenario | What It Catches |
|---|
| Factory floor safety zones | Workers entering machinery exclusion areas during operation |
| Warehouse restricted areas | After-hours intrusion into chemical / high-value storage zones |
| Construction site perimeters | Unauthorized people crossing fence lines or entering excavation pits |
| Mining hazard zones | Personnel entering blast areas or unstable ground |
| Substation / utility yards | Trespassing in high-voltage equipment areas |
What You Need
Hardware
| Device | Purpose | Required |
|---|
| reComputer Industrial (Jetson Orin NX / Nano) | Runs the detection model with Jetson GPU acceleration | Yes |
| RTSP IP camera | Provides the live video feed (Hikvision, Dahua, or any ONVIF-compatible) | Yes |
| Local network switch / PoE switch | Connects the Jetson and the camera | Yes |
Software
- Jetson with JetPack 6.x (L4T 36.x) flashed
- Docker + NVIDIA Container Runtime (pre-installed on Seeed reComputer Industrial)
- SSH access from your computer to the Jetson
Detection Models
| Model | Framework | Precision | Performance |
|---|
| YOLO26n | Ultralytics (NMS-free) | FP16 TensorRT | ~268 QPS, ~3.7ms latency |
| YOLOv8n | Ultralytics | FP16 TensorRT | Real-time inference |
| YOLOv5n | Ultralytics | FP16 TensorRT | Lightweight option |
The first run automatically builds a Jetson GPU-optimized TensorRT engine from the ONNX model and caches it for subsequent restarts.
Notes Before You Deploy
- The first deployment pulls a ~3-5 GB Docker image — keep an eye on disk space
- The first run builds a Jetson GPU-optimized detection engine (1-2 minutes) and caches it for subsequent restarts
- Test your camera RTSP URL with VLC or
ffprobe before deploying — most issues come from a wrong URL
- One Jetson handles multiple cameras in this configuration — each camera has an independent processing pipeline
Key Features
Multi-Camera Management
- Simultaneous RTSP/USB input — connect multiple cameras with independent processing pipelines
- Camera auto-discovery — automatically scan for RTSP cameras on the subnet
- Shared detection model — all cameras share one TensorRT engine for efficient GPU utilization
- Per-camera configuration — each camera has independent zones, lines, and rules
TensorRT FP16 Acceleration
- YOLO26n support — latest Ultralytics model with NMS-free end-to-end inference
- GPU hardware decoding — GStreamer NVDEC for RTSP streams with near-zero CPU overhead
- Optimized for Jetson — tested on reComputer Industrial J4012 (Orin NX 16GB) at 30+ FPS
Interactive Web Dashboard
- Real-time annotated video — bounding boxes, zones, and fence lines overlaid on live feed
- Browser-based zone drawing — draw restricted zones and fence lines directly on the video
- WebSocket streaming — low-latency video delivery to the browser
- Adaptive grid layout — automatically adjusts to multi-camera setups
- HDMI fullscreen mode — press F key to toggle fullscreen display
Event Persistence & Tracking
- SQLite database — events stored persistently, survive restarts
- Centroid tracking — track detected persons across frames
- Date filtering — query events by date for historical review
- Auto-cleanup — automatically removes expired event data
