Segmenting Snow Scene from CCTV using Deep Learning Approach

Recently, data from many sensors has been used in a disaster monitoring of things, such as river wa- ter levels, rainfall levels, and snowfall levels. These types of numeric data can be straightforwardly used in a further analysis. In contrast, data from CCTV cameras (i.e. images and/or videos) cannot be easily interpreted for users in an automatic way. In a tra- ditional way, it is only provided to users for a visual- ization without any meaningful interpretation. Users must rely on their own expertise and experience to interpret such visual information. Thus, this paper proposes the CNN-based method to automatically in- terpret images captured from CCTV cameras, by us- ing snow scene segmentation as a case example. The CNN models are trained to work on 3 classes: snow, non-snow and non-ground. The non-ground class is explicitly learned, in order to avoid a confusion of the models in differentiating snow pixels from non- ground pixels, e.g. sky regions. The VGG-19 with pre-trained weights is retrained using manually la- beled snow, non-snow and non-ground samples. The learned models achieve up to 85% sensitivity and 97% specificity of the snow area segmentation.