ESP32-CAM People Recognition System

I’m building a smart-camera solution that runs entirely on an ESP32-CAM. The unit must detect people in real time and, whenever a face is present, perform face recognition, basic pose estimation, and age/gender inference. All processing needs to happen on-device, with an annotated video stream served over Wi-Fi and JSON events sent via MQTT so the data can be consumed by my wider IoT stack.

I expect the vision pipeline to use lightweight frameworks—TensorFlow Lite for Microcontrollers, a tiny-YOLO or MobileNet variant, plus OpenCV where feasible—while still sustaining at least 10 fps. The face gallery will start small, but I need clear instructions so I can retrain or extend the model later.

Deliverables
• Source code and compiled binary ready to flash on ESP32-CAM
• Trained model files and the notebook/script used to create them
• Any wiring diagram if extra peripherals are required
• Step-by-step setup and usage guide
• Short demo (clip or live link) proving detection accuracy and the MQTT payload format

The project is complete when the camera reliably tags people with the requested attributes and streams both video and metadata as specified.