TAD: Tracking-Aided Detection Siamese Network for Visual Drone Surveillance

Abstract

There is an urgent demand for detecting and tracking intruding drones, where visual-based methods are commonly used. Nevertheless, prior work on visual-based drone surveillance is not capable of achieving high precision and high frame rate at the same time. In this paper, we propose a tracking-aided detection Siamese network for drone surveillance in urban environments, which can maintain a high precision and high speed simultaneously. The proposed framework consists of a detection module and a tracking module, where the tracking module reuses part of the backbone of the detection module to pursue lower computational costs. To ensure high precision, the detection module is used to provide the drone’s initial location and the corrected results for the tracking module if needed. The tracking module fully exploits the temporal information between the frames and aids the detection module in drone localizing, thus achieving high speed. Furthermore, we propose a rule-based heuristic scheduling algorithm to schedule alternate runs of the tracking and detection branches to achieve a good balance between precision and speed. Experiments on the Drone-vs-Bird dataset demonstrated that the speed of the proposed framework is faster than detection-only algorithms, and the accuracy is better than tracking-only algorithms in the meantime.