PIE: Low-light Image Enhancement for Dark Tracking of UAV Based on Paired Images
Chuang Luo, Hongyuan Zheng, Xiangping Zhai, Jing Zhu · 2024
The development of target tracking based on unmanned aerial vehicles (UAVs) has made great progress. These trackers demonstrate commendable performance under optimal lighting conditions. However, when trackers track at night, their tracking performance decreases significantly. In such scenarios, even the state-of-the-art (SOTA) trackers exhibit a discernible reduction in robustness. To improve the UAV’s ability to track at night, this paper introduces a novel low-light image enhancement approach, namely PIE. The approach leverages pairs of low and normal light images as inputs, executing image decomposition and subsequent reconstruction guided by a loss function. The integration of this algorithm seamlessly occurs within the pre-processing phase of the tracker. The devised methodology comprises a lightweight optimization module, a data-driven decomposition module, and an adaptive illumination adjustment module. These components collectively advance the effective decomposition and illumination adjustment of low-illumination images. Through a meticulous evaluation utilizing the widely acknowledged benchmark dataset, UAVDark135, our approach demonstrates pronounced advantages in UAV night tracking scenarios.