Visual tracking of person in bus stops
Llorenç Sastre Galmes · UPCommons institutional repository (Universitat Politècnica de Catalunya) · 2025
This project focuses on the visual tracking of individuals at bus stops using only depth data captured by an Intel RealSense D435i camera, with the aim of preserving user privacy by avoiding RGB imagery. Starting from a small initial dataset, the data was expanded to include nine annotated videos with manually verified ground truth for both detection and tracking. A YOLO-based detector was retrained exclusively on depth images to evaluate the feasibility of depth-only detection. Several multi-object tracking algorithms, including SORT, StrongSORT, and OC-SORT, were evaluated and adapted to incorporate depth information into their Kalman filtering and association strategies. Evaluation using the HOTA metric showcases that depth-based detection and tracking can achieve strong performance, making it a viable solution for people monitoring in privacy-sensitive public transport environments.