Deep Learning Approaches for Visibility Classification Using Traffic Cameras

Fog is a meteorological phenomenon that reduces visibility and poses safety threats to road, maritime and aeronautical traffic. Visibility observations are commonly obtained manually or by sensors that are expensive and usually limited to critical locations such as airports. Fog is difficult to accurately forecast since several factors play a role in its formation and dissipation. Furthermore, fog appears and disappears rapidly and it can be spatially extremely localized. Thus, it is essential to have more visibility observations to issue warnings and for assimilation into models to improve the fog forecast.

To increase the amount of fog observations we use traffic monitoring cameras as visibility sensors. About 5000 such cameras are already installed and operational along the Dutch motorways. We have collected images from 320 cameras every 10 minutes for the last 2 years. These cameras differ in type and are freely controllable by traffic operators, thus changing scenery (e.g., pan, tilt, zoom) at any time. Dynamic sceneries induced us to apply deep learning for fog detection in images starting with two classes. This approach has provided good results for day and night conditions, but not for twilight conditions. Results have been further improved in quality and robustness by post-processing techniques (e.g., spatio-temporal locality of fog). To satisfy the need for fog detection in dusk and dawn (i.e., twilight) conditions as well as more precise fog detection using multiple classes, we revised our approach. We decided to use different models for different twilight conditions and to apply simple data augmentation. Dawn and dusk have proven challenging, especially the latter, due to the scarcity of foggy images in those timeframes. For more precision we applied 4 visibility classes and trained a new deep learning model. The results were of insufficient quality for operational use, mainly due to the higher complication of the problem in addition to the dynamic scenery. Retraining the models using much more images of different fog conditions might solve current limitations, however this is not feasible with the current image collection system. Therefore, our future efforts are to explore complex data augmentation using artificially created (realistic) images to enrich the image dataset.