THERMAL-RGB FUSION WITH LIGHTWEIGHT CNNS FOR NIGHT-TIME DRONE SURVEILLANCE AND REAL-TIME ADAPTIVE SENSOR SELECTION

This paper presents a lightweight and adaptive object detection framework tailored for night-time drone surveillance using thermal-RGB fusion. The proposed system integrates a dual-branch CNN backbone to extract modality-specific features, followed by a mid-level fusion strategy that combines thermal and RGB representations into a unified feature map. A real-time adaptive sensor selection mechanism dynamically adjusts the contribution of each modality based on scene context, enhancing robustness under varying illumination conditions. The architecture employs YOLOv5-Nano as a compact feature extraction backbone, supported by quantization and pruning for real-time deployment. Extensive experiments conducted on publicly available datasets, demonstrate superior accuracy and efficiency compared to single-modality and non-adaptive baselines. Ablation studies validate the contribution of each component. The results highlight the system’s capability to achieve high detection performance with low computational overhead, offering a practical and scalable solution for intelligent aerial surveillance in complex night-time environments