Co-LLIETrack: Cotrained Low Light Image Enhancement Tracker for UAV Tracking

Tracking Unmanned Aerial Vehicle (UAV) in low-light poses significant challenges due to poor visibility, noise, and insufficient datasets. In recent years, object tracking on optical images with Low-Light Image Enhancer (LLIE) has emerged to address low-light tracking challenges. Existing Low-Light Image Enhancer (LLIE)-Tracker frameworks primarily focus on tracking ground objects in low light from UAV onboard camera. This paper extends the LLIE-Tracker framework for aerial UAV tracking in low-light condition, and also introduces a co-training strategy named Co-LLIETrack to train LLIE and tracker jointly using a novel loss function, Semantic Feature Map with IoU Matching (SFMIoU) loss to further improve the tracking performance.To further address the scarcity of publicly available low-light UAV tracking data due to limitations posed by real-world data constraints, such as non-fly zones, complex regulatory approvals, and adverse weather conditions. We present DeepLowLight-UAV, a purpose-built synthetic dataset tailored for UAV tracking tasks. To create low-light scenarios, we developed a Deep Learning (DL) driven pipeline that synthetically produces low-illumination images. This dataset serves as a robust resource for both training and benchmarking DL based UAV tracking models.Comprehensive evaluations on the AntiUAV-I benchmark reveal that our proposed approach leads to substantial enhancements in tracking performance, delivering an average improvement of 4% in accuracy across several leading State-of-the-Art (SotA) tracking methods.