A Dataset Annotation System for Snowy Weather Road Surface Classification

Abstract

In this paper, we introduce an artificial intelligence-based annotation system for a dataset of snow-covered road images. We operate on a large dataset consisting of CCTV images and time and weather metadata. The dataset is fed to a series of data processing techniques to automatically assign each image one of four snow-cover categories aligned with snow removal operations. The processing pipeline includes feature learning using convolutional autoencoders and graph clustering using the Louvain community detection algorithm. The resulting dataset comprises over 41000 images automatically annotated in different weather and time settings. We train and test multiple deep learning models to validate the annotated dataset to classify snow-covered road images. We customize the models to consider the class distribution within the dataset. We achieve precision and recall scores of 97% using an EfficientNet model trained on separate day and night image datasets and using a class-weighted loss function.