Edge CV Safety Monitor

Architecture

The pipeline is designed for field reliability: the RTSP reader runs in a dedicated thread with auto-reconnect, inference is decoupled from the display loop, and telemetry is broadcast over UDP so any monitoring node on the subnet (or across a VPN tunnel) can receive it without a persistent connection.

Deployment is managed by a single Bash script (deploy.sh) that handles dependency install, network interface config, static IP assignment, firewall rules, and systemd service registration — the same pattern used for field rollouts at customer sites.

Code sample

# Full install on a fresh Jetson or Linux edge device
sudo ./deploy.sh install \
    --source rtsp://192.168.1.22:554/stream1 \
    --api-key rf_abc123

# Verify camera connectivity before deployment
sudo ./deploy.sh check-camera rtsp://192.168.1.22:554/stream1

# Set static IP (no DHCP at customer site)
sudo ./deploy.sh set-ip eth0 192.168.1.100 192.168.1.1

# Tail live inference log
./deploy.sh logs

def _reader_loop(self):
    while self._running:
        cap = cv2.VideoCapture(self.source)
        if not cap.isOpened():
            log.warning(f"Cannot open {self.source}, retry in {self.reconnect_delay}s")
            time.sleep(self.reconnect_delay)
            continue
        while self._running:
            ok, frame = cap.read()
            if not ok:
                log.warning("Stream dropped — reconnecting")
                break
            with self._lock:
                self._frame = frame   # thread-safe frame swap
        cap.release()
        time.sleep(self.reconnect_delay)

Skills demonstrated

Key outcomes

• Sub-35ms inference latency on Jetson Orin using Roboflow TensorRT-optimised models
• Zero-downtime RTSP reconnect — threaded reader recovers from dropped streams without stopping the inference loop
• Single-command site deployment — deploy.sh configures networking, firewall, and systemd service for a new edge node in under 5 minutes
• Hardware-agnostic — runs on Jetson (CUDA), x86 CPU, or ARM without code changes; device is auto-detected
• Telemetry visible from any subnet node — UDP broadcast works across VPN tunnels used in customer deployments

Source