Retinex-guided illumination recovery and progressive feature adaptation for real-world nighttime UAV-based vehicle detection

Nighttime vehicle detection in Unmanned Aerial Vehicle (UAV) imagery is critical for intelligent transportation systems yet faces significant challenges due to low signal-to-noise ratios and pervasive small and medium-scale objects. Existing methods suffer from two critical limitations: (1) conventional low-light enhancement approaches prioritize human perception over downstream detection tasks, and (2) feature fusion frameworks exhibit inadequate cross-level interactions and computational inefficiency for UAV platforms. To address these gaps, we propose the Retinex-guided illumination Differential Transformer Detection network (ReDT-Det), which integrates a nighttime image enhancer with a robust vehicle detection module. Our approach leverages Retinex principles to design an illumination-fused differential transformer block for preliminary image enhancement and illumination recovery, which can effectively improve the quality of nighttime images while preserving critical details. To address the issue of small and medium-scale objects, we introduce a dilation-wise residual cross-stage partial module to enhance the ability to capture fine-grained features. During the feature fusion stage, we propose two key modules: a cross-level feature adaptive adjustment module for the effective integration of multi-scale features and a small object auxiliary feature module specifically designed to enhance the representation of small-scale objects. To validate our method, we curated a comprehensive benchmark dataset for real-world nighttime UAV-based vehicle detection, named NightDrone-Mix. Extensive comparative experiments demonstrate that ReDT-Det outperforms various state-of-the-art image enhancement and detection methods, highlighting its advantages in both accuracy and effectiveness. Additionally, we evaluated ReDT-Det on the DroneVehicle(Night) and ExDark datasets to assess its performance in detecting dark objects, achieving equally promising results.