Tracciamento delle Persone nell'Era dell'Intelligenza Artificiale - Dai Vincoli Geometrici ai Modelli Data-Driven

Abstract: In recent years, computer vision has empowered machines to "see" and recognize objects, yet tracking multiple objects over time, especially in video sequences, remains a significant challenge. Pedestrian tracking in crowded environments is particularly complex due to frequent occlusions, shape variations, and appearance changes. Traditional tracking methods, which rely on hand-crafted heuristics, have provided valuable insights into motion prediction and identity maintenance across frames but often falter in dynamic, real-world scenarios. Recent advances in deep learning have shown the potential to overcome these limitations by leveraging large datasets and learning more generalized models. In other fields, fully end-to-end approaches have outperformed heuristic-driven methods, offering more accurate representations. However, deep learning models continue to face difficulties in tracking, particularly in balancing detection with tracking tasks, often struggling to generalize across different scenarios. This Ph.D. thesis provides a comprehensive exploration of pedestrian tracking methodologies, from traditional heuristic-based approaches to modern deep learning advancements. It introduces novel geometric techniques to address the limitations of current systems. Additionally, the thesis proposes a modular framework for fully end-to-end, data-driven trackers, allowing for the dynamic selection of specialized modules based on scene characteristics. This framework enhances adaptability to unseen domains, offering operators the flexibility to tailor tracking systems to specific camera configurations and environments.

Citation:

Mancusi, Gianluca "Tracciamento delle Persone nell'Era dell'Intelligenza Artificiale - Dai Vincoli Geometrici ai Modelli Data-Driven" 2025

 not available

Paper download:

• Author version: