Leveraging airport CCTV footage through video understanding techniques for visibility prediction

Accurate atmospheric visibility prediction is critical for enhancing airport safety, especially under adverse weather conditions. However, existing visibility prediction methods predominantly rely on single-image analysis, using either traditional image processing techniques or deep learning models, which often fail to fully capture the dynamic and temporal characteristics inherent in video data. In this study, we explore various video understanding models for visibility prediction, achieving promising results and pioneering the use of video understanding techniques in this domain. Unlike traditional static image-based methods, our proposed three-stream network model integrates spatial information from individual frames, motion dynamics through optical flow, and key reference points extracted via SIFT (scale-invariant feature transform) descriptors. This enables the model to capture both short-term and long-term environmental changes. Experimental results show that our three-stream network significantly outperforms single-frame and image-based models in predicting $RVR\text{\_}1A$ (average runway visibility range of one minute). Our model achieves an $R^2$ mean of 0.896 and an accuracy mean of 0.860 on the test set, substantially outperforming traditional methods. These results not only demonstrate the superior performance of our approach in real-world scenarios but also highlight the potential of video understanding techniques for airport safety monitoring applications.