ESP32-S3 Can Detect People Locally for $12, No Cloud or WiFi Needed

A developer tutorial published on DEV Community demonstrates how to run person-detection AI entirely on an ESP32-S3 microcontroller, costing as little as $12, without sending any data to the cloud. The approach uses TinyML with a quantized MobileNetV1 or MobileNetV2 model trained on the Visual Wake Words dataset, which classifies frames as person or no-person. Running on hardware like the XIAO ESP32S3 Sense or Freenove S3 WROOM CAM, the system processes QVGA frames in 80 to 120 milliseconds using the chip's int8 hardware acceleration. Unlike cloud-based alternatives, which introduce latency, recurring costs, and privacy concerns by transmitting footage to remote servers, this method keeps all image data confined to the device's PSRAM. The tutorial argues that for simple presence detection, local inference is not only cheaper but more practical and privacy-preserving than any cloud solution.