Convolutional Neural Network based classification of road conditions from on-road positioned cameras

The road condition has a major impact on car accidents and related casualties, specifically the presence of water on the road. Therefore, estimating the water's existence is an essential application in order to make the road safer. Nowadays, The Convolutional Neural Network (CNN) is introduced as one of the main solutions for image classification problems. Due to the fast improvement in deep learning and artificial intelligence, CNN found its position as one of the various classes of neural networks, which is often applied to analyze and process the image dataset. In this presented work, the CNN capabilities are used to implement a real-time condition classifier which seems an appropriate way to reduce the number of car accidents as well as aiding drivers to be aware of the surrounding environment while driving. The main goal of this thesis is to implement a classifier to evaluate the road condition with respect to the presence of water. This project is performed in collaboration with Politecnico di Torino University and Waterview srl. The system is designed to process frames captured by surveillance Full-HD cameras. Hence, in order to classify the road condition, firstly the semantic annotation was performed, and then a CNN model is trained to distinguish samples that belong to different classes. The proposed model is able to detect wet frames with significant accuracy, especially when the images with half of the resolution are used as the input data. In addition, it observed that the focus of the neural network is mostly on the road's part of the images called the region of interest, instead of concentrating on the unwanted areas of the image such as sky or road-sides.