About this project

A camera board propped on the floor counts sit-ups: frames are compressed and sent out to measure the shoulder-hip-knee angle, with lying above 110° and sitting below 55°, so one down-and-up is one repetition and the band between the thresholds absorbs jitter. Recognition stays off the device because a pose model needs over seven seconds per frame while a sit-up lasts two to four, but counting runs locally so a dropped link loses nothing. It can call a Python server on a home machine, an OpenAI-compatible vision API, or run in a demo mode. The firmware builds inside the ESP-Mosaico workspace with ESP-IDF and produces a 1.9MB image.