Design and implementation of an embedded AI system for real-time people tracking using Time-of-Flight sensors and 3D point cloud processing.

Accurate monitoring of space occupancy has become a central challenge for smart buildings. Real-time knowledge of how many people are inside a room enables energy savings, supports safety, and prevents overcrowding in places such as offices, museums, and public events. At the same time, privacy requirements make camera-based systems unsuitable, since they identify individuals rather than providing only anonymous counts. This thesis introduces an embedded people counting system based on an overhead Time-of-Flight (ToF) sensor, designed to operate in real time on resource-constrained devices. The core contribution is the design of a complete pipeline that brings together several original elements: the construction of a dedicated dataset with 3D annotations, the development of a lightweight neural network called ToFNet-light that predicts both the number of people and their centroids, and a tailored tracking logic for single and multiple crossings. By replacing heavy clustering algorithms and directly producing spatial information, the system enables AI on edge, delivering accurate counting directly on the device without external computation. The implementation was optimized for Raspberry Pi with multithreading and ONNX quantization, achieving stable real-time performance at around 30 frames per second. Experiments with people of different ages and characteristics confirmed the robustness of the approach, and showed its potential to be extended to more complex multi-person scenarios. The main outcome of this work is a lightweight, scalable, and privacy-preserving pipeline that combines dataset creation, neural network design, and tracking into a unified framework. This provides a concrete and innovative step toward reliable people counting and intelligent occupancy management in indoor environments.