A Real-Time and Lightweight Method for Tiny Airborne Object Detection

Yanyi Lyu, Zhunga Liu, Huandong Li, Dongxiu Guo, Yimin Fu; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2023, pp. 3016-3025

With wide applications of unmanned aerial vehicles (UAVs), the detection of airborne objects has become crucial to ensure the flight safety of UAVs and prevent their illegal use. Although object detection has achieved great success in past years, it is still a challenging problem to detect tiny airborne objects. To solve this problem, we propose a simple and effective Tiny Airborne object Detection (TAD) method. It locates potential objects using inconsistent motion cues between airborne objects and backgrounds instead of the low-quality representation of tiny objects. This enables TAD to sensitively detect tiny objects with limited appearance information. Specifically, we first establish correspondences of pixels between adjacent frames based on the local similarity of spatial feature vectors to achieve motion modeling. Next, the local similarity of motion patterns is computed to explicitly describe the motion consistency of each position with its surrounding pixels. Then, a simple network is used to output the heatmap that reflects the probability of object presence. A higher probability of containing an object will be assigned to positions with a greater difference in motion from their surrounding pixels. Finally, an independent network branch is employed to regress center offsets and scale information of objects, which are used to correct the error in the estimated object position from the heatmap and obtain the final bounding box, respectively. Experiments on three challenging datasets demonstrate that the proposed method can achieve advanced performance. Notably, TAD is highly lightweight, and the detection speed is significantly better than existing methods.