Visitor counter with machine learning
Vänskä, Janne (2025)
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-2025120532997
https://urn.fi/URN:NBN:fi:amk-2025120532997
Tiivistelmä
The aim of this thesis was to design and implement a low-power, autonomous visitor counting system that operates reliably in remote outdoor environments without continuous network connectivity. The goal was to explore how TinyML based person detection could be integrated into a microcontroller-based device to enable local decision-making, movement direction estimation, and event logging while maintaining minimal energy consumption. The work was motivated by the need for improved visitor monitoring in areas such as national parks, where traditional counting technologies are ineffective and machine vision-based alternatives are rarely optimized for off-grid conditions.
The study implemented a dual-microcontroller architecture consisting of an Arduino Nano 33 BLE for sensing and image capture and an ESP32 microcontroller for running a TensorFlow Lite Micro person-detection model. A Python-based PC environment was developed to support dataset generation, preprocessing, inference testing, and debugging before deployment onto the embedded hardware. The system incorporated structured frame transmission, Region of Interest adjustments, and lightweight CNN-based detection using a quantized TensorFlow Lite model. Additional features included motion-triggered wakeup, directional estimation based on x-axis movement, and CSV logging with timestamped detection events.
The resulting prototype achieved stable image capture and inference speeds of approximately one to two frames per second, demonstrating that low-cost microcontrollers can perform onboard person detection and movement interpretation effectively under constrained resources. The findings confirm that TinyML is a viable approach for distributed sensing in remote environments and can operate without cloud services while maintaining energy efficiency. The thesis concludes that with further improvements such as deep-sleep power management, solar-powered operation, and environment-specific model training the system can be developed into a robust, field-deployable visitor counter suitable for long-term autonomous use.
The study implemented a dual-microcontroller architecture consisting of an Arduino Nano 33 BLE for sensing and image capture and an ESP32 microcontroller for running a TensorFlow Lite Micro person-detection model. A Python-based PC environment was developed to support dataset generation, preprocessing, inference testing, and debugging before deployment onto the embedded hardware. The system incorporated structured frame transmission, Region of Interest adjustments, and lightweight CNN-based detection using a quantized TensorFlow Lite model. Additional features included motion-triggered wakeup, directional estimation based on x-axis movement, and CSV logging with timestamped detection events.
The resulting prototype achieved stable image capture and inference speeds of approximately one to two frames per second, demonstrating that low-cost microcontrollers can perform onboard person detection and movement interpretation effectively under constrained resources. The findings confirm that TinyML is a viable approach for distributed sensing in remote environments and can operate without cloud services while maintaining energy efficiency. The thesis concludes that with further improvements such as deep-sleep power management, solar-powered operation, and environment-specific model training the system can be developed into a robust, field-deployable visitor counter suitable for long-term autonomous use.