Omnidrone-Det: Omnidirectional 3D Drone Detection in Flight

The widespread use of drones in applications like delivery, search-and-rescue, and warehouse management has intensified the need for advanced object detection in 3D space, which is crucial for collision avoidance and safe navigation. However, the structural constraints of drones, such as their limited weight and size, restrict the use of sophisticated sensors and processors, posing challenges for effective object detection and tracking within a wide field of view (FOV). Addressing these limitations, this paper introduces a flexible two-stage omnidirectional 3D object detection approach tailored for on-board processors. Specifically, fisheye cameras are utilized to capture omnidirectional images, which are then flattened with a rectification module. Our 3D drone detection system combines a YOLOv7-based drone detector, fine-tuned on a self-collected drone dataset, with a monocular depth estimation model and a supervised depth corrector module. By leveraging fisheye cameras, our system efficiently achieves 3D localization of drones in flight, providing a lightweight yet effective solution for both static and dynamic environments.