Attention-Based Feature Fusion Empowered Encoder-Decoder Framework for Nighttime Traffic Perception From High-Altitude Surveillance System

Traffic perception is essential for intelligent transportation systems (ITS) to improve road safety and facilitate traffic management. High-altitude video surveillance systems, known for their wide coverage, provide an effective solution for continuous traffic monitoring and are increasingly deployed worldwide. However, the application of high-altitude video surveillance faces significant challenges, such as the small size and high density of vehicles, and low ambient brightness at night or low-light conditions. In this study, we propose a novel lightweight model specifically for reliable traffic perception in low-light conditions, utilizing an encoder-decoder architecture. The proposed framework begins with an advanced image enhancement algorithm designed to amplify the semantic content of low-brightness nighttime images. Subsequently, we employ a lightweight, deep two-channel neural network to extract image features both pre- and post-enhancement. To effectively manage the noise introduced during image enhancement, we introduce a feature fusion module based on an attention mechanism, which adaptively merges the features extracted by the two-channel network. The final step involves a lightweight decoder that synergizes high-level semantic information with low-level features. Extensive evaluation on our nighttime dataset shows that the proposed method can reduce error by 40% compared to the methods without image enhancement and by 28% compared to those relying solely on enhancement, outperforming all baseline methods. Notably, the lightweight module embedded in our model empowers the framework to perform real-time traffic perception at night, providing an efficient solution for traffic monitoring.