Towards Efficient Detection and Tracking for Tiny Airborne Object

The detection and tracking of airborne objects is the key technology to ensure flight safety and prevent the illegal use of Unmanned Aerial Vehicles (UAVs). The airborne objects typically appear small due to the considerable distance between the observer and the object. Additionally, the erratic movement of the onboard camera makes it difficult to track airborne objects steadily. In order to solve these problems, a lightweight method is proposed in this paper to effectively detect and track tiny airborne objects with limited appearance information. First, a parameter-free approach is designed with parallel implementation to model motion patterns fast. The local similarity measure is employed to determine the pixel-level correspondence between adjacent frames. Then, an efficient Tiny Airborne object Detection (TAD) method is developed based on local similarity calculation. In TAD, objects are located using inconsistent motion cues between airborne objects and backgrounds instead of the low-quality representation of tiny objects. Finally, we extend TAD to the Tiny Airborne object Tracking (TAT). The TAT method works on a more relaxed matching condition to reduce the mismatch caused by low tolerance of small objects for bounding box perturbation. Experiments on three challenging datasets demonstrate that the proposed method, including TAD and TAT, can achieve advanced performance. Moreover, the method is easy to implement, and the detection speed is significantly better than existing methods.