A simple python program that monitors an IPC stream for known faces and controls a wicket gate. It does realtime object detection with Google Coral through rest API achieving >20 FPS. Having detected person or car arriving, it zooms in the IPC and does face recognition with Deepstack with ~4 FPS. When a known face is recognized with confidence of 80% or higher, the program opens wicket gate (Dahua VTO2000). It also communicates with Home Assistant with MQTT.
The program reads RTSP stream of 6MP Dahua IPC with cv2 and processes frame by frame. It slices frames (numpy arrays) to relevant areas to boost speed. It controls camera's optical zoom and unlocks the wicket with Dahua rest API.
20+ FPS (max IPC frame rate) is achieved on 8GB RPi4 with Google Coral USB accelerator accessible through https://github.com/sickidolderivative/tensorflow-lite-rest-server.
4 FPS is achieved on Deepstack running on 4GB Jetson nano: https://forum.deepstack.cc/t/deepstack-release-on-jetson/
The program communicates with Home Assistant with Paho MQTT. It send sliced frames to feed HA MQTT camera and provides status updates on detections and recognitions.
The program is specific to my home setup, but perhaps someone finds parts of it useful.