Why AI Video Analytics Beats Standard Motion Detection
Standard CCTV cameras detect motion (pixel change) and alert on everything: blowing leaves, insects, passing headlights — producing overwhelming false alarms. AI video analytics uses computer vision to identify whether the moving object is a person, vehicle, animal, or something else, alerting only on configured objects.
Frigate NVR: Open-Source AI CCTV Platform
Frigate is an open-source NVR that runs in Docker on a Home Assistant host (Linux) or a separate machine. Key features: RTSP stream ingestion from any IP camera brand (Hikvision, Dahua, Reolink, Amcrest); object detection — person, car, truck, bicycle, dog, cat, bird — using YOLOv8 or TensorFlow Lite models; zone-based alerts that confine detection to defined polygon areas (e.g., alert only for a person crossing the perimeter line, not walking on the street); object tracking across frames to prevent duplicate alerts for the same object; automatic snapshot and video clip capture on detection; and Home Assistant integration via MQTT + Frigate integration for per-zone sensor states.
Google Coral TPU Accelerator
The Google Coral USB Accelerator (~35–40 USD) is an Edge TPU that supplements CPU inference for TensorFlow Lite models. Performance comparison: CPU only (Intel i5): ~5–10 FPS; Coral USB (Pi 5/PC): ~30–50 FPS per stream; Coral PCIe (Coral Dev Board Mini): ~200+ FPS for multiple streams. Frigate uses Coral via TensorFlow Lite delegate — configure in config.yaml: detectors: coral: type: edgetpu device: usb.
YOLO vs TensorFlow Lite Models for Frigate
TFLite Coral Model (EfficientDet-Lite): optimised for Coral TPU — fast, small model size, slightly lower accuracy than YOLOv8.
YOLOv8/YOLOv9 (CPU/GPU): higher accuracy but requires stronger CPU/GPU. Best with Coral PCIe or a dedicated GPU.
For home use on Raspberry Pi 5 + Coral USB, the TFLite Coral model is sufficient and fast enough.
Zone-Based Alert Configuration
For each camera, define zones using pixel-coordinate polygons: Front Door Zone, Perimeter Zone, Car Park Zone. Alert rules: person in Front Door Zone → HA notification + camera snapshot; person in Perimeter Zone after 22:00 → LINE Critical Alert; car in Car Park Zone not on whitelist → alert. Set a minimum confidence score (0.7+ for person) to reduce false detections.
License Plate Recognition (LPR) for Car Parks
Add LPR via Frigate + ALPR integration (OpenALPR or PlateRecognizer API). A camera aimed at the car park entrance captures plates on arrival. If the plate is on the whitelist → open the gate automatically. If not → alert security, log the plate and timestamp.
Comparison: Blue Iris vs Frigate vs Hikvision DeepinMind
Blue Iris: Windows-only, one-time fee ~70 USD, good AI detection but requires a powerful PC. Frigate: open-source and free, Linux/Docker, optimised for Home Assistant, full Coral support. Hikvision DeepinMind: closed camera+NVR ecosystem from Hikvision — excellent results but vendor lock-in.