DEFESA DE TESE DE DOUTORADO Nº 34

Aluno: David Josué Barrientos Rojas

Título: “NOVA-STAR: Non-Overlapping View Active Sequential Tracking with Adaptive Reinforcement Learning for Coordinated Multi-Camera Systems"

Orientador: Bruno José Torres Fernandes

Coorientador: Pablo Vinicius Alves de Barros

Examinador Externo: Matthias Kerzel (University of Hamburg, Germany)

Examinador Externo: Doreen Jirak (University of Antwerp, Belgium)

Examinador Externo: Yves Mendes Galvão (Livelo Brasil)

Examinador Interno: João Fausto Lorenzato de Oliveira

Data-hora: 24 de setembro de 2025 às 8h

Local: formato híbrido

---

Resumo:

         "Active object tracking across nonoverlapping camera networks is essential for applications such as urban surveillance, industrial monitoring, and indoor security, yet remains less explored in computer vision and reinforcement learning research. Existing methods typically focus on passive surveillance or single-camera active tracking, which overlook challenges specific to sequentially aligned but visually disconnected cameras, including visual discontinuities, accurate inter-camera handovers, and motion reasoning without continuous visual input. The adaptive and predictive capabilities of reinforcement learning make it well-suited to the dynamic nature of active tracking. In this framework, the tracking agent learns from interactions to optimize state estimation, motion prediction, and data association from observed frames and environmental feedback. Building on these strengths, this thesis proposes NOVA-STAR: a reinforcement learning-based approach for coordinated tracking across nonoverlapping multicamera systems. Using Proximal Policy Optimization, a centralized controller with a finite state machine abstraction encodes visibility states and supports predictive reasoning during camera handoffs. Furthermore, a novel simulation environment was created to replicate real-world conditions such as occlusion, spatial discontinuities, and partial observability. Experimental evaluations demonstrated the effectiveness of the proposed method, achieving a Coordinated Tracking Rate of 84.72%, a Handoff Completion Success Rate of 82.32%, a Reacquisition Success Rate of 86.92%, and a Standby Accuracy of 97.50%. Behavioral analyses and Ablation studies highlighted the importance of finite state machine encoding and centralized observation sharing. Additionally, a real-world deployment with one physical and two simulated cameras provided practical validation. Through inductive reasoning based on these results, the system is expected to generalize effectively to multi-camera real-world deployments. Overall, these results indicate that NOVA-STAR offers a reliable framework for active tracking on nonoverlapping sequential cameras, with consistent performance in both simulation and real-world deployment. The approach is well-suited for sequential multi-camera, single-target scenarios but may be challenged by abrupt motion changes, extended occlusions, or untested configurations. By combining coordinated control with predictive reasoning, the method contributes to bridging the gap between computer vision and reinforcement learning, offering a practical basis for future active tracking systems."