Robotic surveillance with re-identification for long-term tracking
conference paper
With the increased autonomy of mobile robots, it has become possible to execute automatic exploration and surveillance with multi-agent systems in open world environments. However, ensuring effective situational awareness without overwhelming a supervising operator requires concise, consistent tracking of a priori unknown individuals over extended periods. The goal is a continuous shared mapping between detections and unique persons, so that full individual trajectories can be reconstructed over time, even when certain persons are temporarily not in the view in any of the cameras. Reliably differentiating between unique individuals that are detected in a monitored area is a difficult and yet unsolved problem also referred to as open-world re-identification. A closely
related challenge is to differentiate the unique detected persons on the go, referred to as online long-term tracking. Previous tracking works have used re-identification features to solve short term tracking issues. However, such approaches in general fail to associate tracks over longer periods. We therefore investigate a re-identification based persistent tracking method that enables the long-term association of tracks over multiple cameras. The proposal aims at clustering different tracks on the basis of feature vector similarities where resulting track clusters represent the unique persons seen during deployment. Our method is evaluated in out-of-domain scenarios to get a realistic idea of the performance in unseen environments, paving the way to practical improvements in situational awareness for multi-robot surveillance.
related challenge is to differentiate the unique detected persons on the go, referred to as online long-term tracking. Previous tracking works have used re-identification features to solve short term tracking issues. However, such approaches in general fail to associate tracks over longer periods. We therefore investigate a re-identification based persistent tracking method that enables the long-term association of tracks over multiple cameras. The proposal aims at clustering different tracks on the basis of feature vector similarities where resulting track clusters represent the unique persons seen during deployment. Our method is evaluated in out-of-domain scenarios to get a realistic idea of the performance in unseen environments, paving the way to practical improvements in situational awareness for multi-robot surveillance.
Topics
TNO Identifier
1020274
Publisher
SPIE
Source title
Autonomous Systems for Security and Defence II Proceedings of SPIE 15–16 September 2025 Madrid, Spain
Editor(s)
Kampmeijer, L.
Masini, B.
Milosevic, Z.
Masini, B.
Milosevic, Z.
Files
To receive the publication files, please send an e-mail request to TNO Repository.