Cooperative unmanned aerial vehicles with privacy preserving deep vision for real-time object identification and tracking

Abstract

Human tracking is an important challenge in a wide variety of applications, including but not limited to, surveillance, military operations, and disaster relief services. Unmanned Aerial Vehicles (UAVs) allow the surveying of dangerous or impassable areas from a safe distance. They also provide a machine-based capability, which may not only solve resource constraint issues, but can also improve effectiveness and efficiency in the tracking task. The effectiveness of tracking is directly related to the angle of view and degree of freedom of the camera system. In this paper, we introduce a decentralized, distributed deep learning algorithm for Real-Time Privacy-preserving Target Tracking Re-Identification (RPTT-ReID) used by cooperative UAVs in complex and adversarial environments involving motion, crowded scenes, and varied camera angles. The efficiency of RPTT-ReID makes it amenable to edge computing applications. The proposed algorithmic approach resolves shortfalls with current tracking algorithms, specifically challenges in maintaining tracking when subjects cross paths, switch identity, or are occluded in a frame of view. We demonstrate the power of our approach both in single and multi-UAV scenarios to track movable targets by extracting the facial embedding information in crowds, in order to ensure the privacy of individuals captured by the UAVs without compromising the capability for target re-identification. We validate RPTT-ReID on a challenging video dataset of crowded scenes. Our experimental evaluation shows that the proposed approach is capable of tracking and re-identifying people in crowds despite blended trajectories with minimum and maximum accuracy of 79.91 ± 0.2% and 93.27 ± 0.1% respectively. The proposed approach is 18% faster than previous methods for tracking in crowded urban environments.