Build Your Own Fatigue‑Detection System with Tencent VisionSeed and STM32

Driving while drowsy is a major safety risk; modern cars use driver‑monitoring systems such as active braking, lane‑departure warnings, and driver‑monitoring systems (DMS) to mitigate this. This tutorial shows how to build a low‑cost, DIY fatigue‑detection device using Tencent Youtu VisionSeed AI vision module and an inexpensive STM32 microcontroller.

AI Vision Module Overview

VisionSeed is a compact AI accelerator board that includes a dedicated AI chip capable of 512 multiply‑add operations per cycle, hardware support for convolution, pooling, and fully connected layers, and two camera modules. It can run face detection, 90‑point facial landmark localization, pose estimation, and face recognition directly on the board, outputting results via UART.

Performance Considerations

Running face detection on a Raspberry Pi 3B with OpenCV Haar cascades takes over 160 ms per frame. Using a MobileNet‑SSD model accelerated by NCNN increases latency to about 260 ms, and adding a facial‑landmark model reduces the overall frame rate to roughly 3 fps, which is insufficient for real‑time monitoring.

VisionSeed Features

VisionSeed provides the following algorithms out of the box:

Face detection

90‑point facial landmark localization

Pose angle calculation

Face recognition

Custom models can also be compiled and uploaded to the AI chip.

Connecting VisionSeed to a PC

When VisionSeed is connected via USB, it creates a video device (Video0) and a serial device (ttyACM0). Opening ttyACM0 with the provided configuration tool streams real‑time analysis data, such as facial landmarks visualized on a virtual avatar. The “left‑eye open/close” curve spikes when the user blinks and stays low when the eye is closed.

Hardware Integration with STM32

To generate an audible alarm when the driver’s eyes are closed, the UART output from VisionSeed is fed into an STM32F103 microcontroller. The STM32 reads the serial data, determines whether the eye‑closure duration exceeds a safety threshold, and triggers an MP3 playback module to emit a warning sound. The circuit uses a 30‑pin FPC connector to replace VisionSeed’s original interface board, a micro‑USB connector for power, an ST‑Link header for programming, and a 3.5 mm audio jack for the speaker.

Firmware Development

VisionSeed supplies open‑source SDKs for C++, Python, and JavaScript that parse the UART stream into face bounding boxes and 90 landmark coordinates. Using only a few dozen lines of code, the STM32 firmware extracts the 18 points around both eyes, computes the eye‑closure ratio, and decides whether to activate the alarm.

Testing and Results

After assembling the hardware and flashing the firmware via ST‑Link, the module is mounted on the steering wheel. When the driver closes their eyes, the system instantly emits a loud beep, effectively waking the driver and preventing drowsy driving.

Conclusion

This project demonstrates that advanced AI‑driven driver‑monitoring capabilities, previously limited to high‑end vehicles, can be replicated with affordable components. By leveraging VisionSeed’s on‑board AI algorithms and a low‑cost STM32, hobbyists can build a reliable fatigue‑detection system for personal use.