Multi-Object Target Tracking using a Fleet of Drones: Target Identification and Re-Identification

Recently, drones, or Unmanned Aerial Vehicles (UAVs), have emerged as a transformative technology, finding applications in diverse fields ranging from surveillance to Search and Rescue (SAR) operations. Within these applications, the capabilities of object detection and multi-object tracking (MOT) are of paramount importance, because they allow the drone to recognize and keep trace of the movement of targets without the need for human intervention. While there has been significant research in the domain of multiple-camera multi-object tracking, much of it has been conducted in an offline context. The challenge of online multi-object tracking, especially from moving aerial cameras (bird-eye view), remains relatively unexplored. This thesis, developed in collaboration with Leonardo Labs, seeks to bridge this gap, focusing on object identification and tracking using a fleet of drones. A particular emphasis is laid on the challenges and complications associated with re-identification (Re-ID), culminating in the design of a specialized module to aggregate and match information from individual drones. Simulations are conducted using a PX4-Gazebo simulation environment, with ROS2 as middleware. The preliminary objective of this research was to implement a detection algorithm coupled with a MOT tracking algorithm for single-drone operations, considering cars as specific targets. Efforts were directed towards minimizing the number of incorrect assignments during the tracking of such vehicles. To further refine the re-identification process, a localization module, leveraging a monocular camera, was developed. This module serves to provide spatial information about identified targets, laying the groundwork for subsequent Re-ID tasks. Another contribution consisted in the implementation of a so-called "global module". This module, simulating the functionalities of a ground station or a master drone, enhances the system's capabilities. Indeed, it acts as a centralized hub, collecting target features and positional data from individual drones, and it harnesses this information for the Re-ID task. This approach, even if centralized, offers two benefits: it minimizes communication overhead and facilitates the re-identification of targets that temporarily exit the field of view of a single drone. Therefore, the developed global module paves the way for more advanced and collaborative multi-drone operations in the future.