ESCVehicle: A Drone-Based Visible–Infrared Vehicle Benchmark With Extensive Scene Coverage

UAV-based vehicle detection aims to efficiently identify and distinguish vehicle targets from aerial remote sensing imagery. It has been widely applied in areas such as traffic management and emergency response. However, existing datasets are limited by scene diversity and insufficient environmental complexity, making it challenging to effectively train and comprehensively evaluate multimodal algorithms under real-world complex scenarios. To address this limitation, we introduce ESCVehicle, a novel visible-infrared vehicle detection dataset captured by drones with extensive scene coverage and annotations. The dataset comprises 10,727 pairs of aligned visible and infrared images, over 360,000 finely annotated rotated bounding boxes, and spans seven vehicle categories, nine representative scene types, and typical day-night periods. Data collection was conducted under both natural and adverse weather conditions, offering rich and comprehensive data support for vehicle detection and classification tasks. In addition, to tackle the challenges of robust detection in complex environments, we propose a novel Cross-Modal Complex Scene Vehicle Detection framework (C2-VeD). The framework incorporates an Adaptive Feature Enhancement Convolution (AFEConv), which focuses on distinguishing target features from complex background patterns. By employing feature selection and channel reorganization mechanisms, AFEConv enhances the representation of target-relevant features while suppressing background interference. Furthermore, a Cross-Modal Context-Aware Fusion (CM-CAF) module is introduced to model the spatial dependencies of local features through cross-modal fusion and spatial context awareness, thereby reinforcing the complementarity between modalities and improving detection accuracy and robustness. Experimental results on the ESCVehicle dataset demonstrate that the proposed framework achieves superior performance under various complex scene conditions. The dataset and code will be published at https://github.com/sjm2001-rslab/ESCVehicle.