Temporal models for online detection of weather events from roadside cameras

Having clear and fast information is decisive to ensure road safety. Nowadays, countless applications go in this direction: from innovative vehicle infotainment to intelligent traffic lights and control systems. Bad weather conditions are popularly acknowledged as a reason for diminished visibility and challenging brakings, thus resulting in higher risks for drivers safety. On-time information gathering and quick driver alerting are arguably among the lightest solutions to deploy for safer mobility. To this extent, many road stretches worldwide are already provided with roadside surveillance cameras, which are used in this scope as the source of real-time data. The work of this thesis falls in a wider collaboration with WaterView s.r.l., and it aims to develop a neural network-based model able to exploit these preexisting devices and detect the presence of water on the road pavement. This thesis provides a new solution for the pipeline of the project, namely selecting and integrating one of the temporal models present in the deep learning scenario. In this setting, LSTM, Dense LSTM, and Temporal Convolutional networks are the architectures assessed. This assessment led to finding that a stack of LSTM cells was the most suitable among these architectures. Afterwards, the selected model was successfully integrated with a semisupervised approach - which was chosen as the eligible strategy in one of the former iterations of the project. Finally, a second version of the model was developed, which accepted frames with a higher resolution. The temporal model improved the quality of the predictions compared to the former semisupervised trained model at higher and standard image sizes. Results showed that both these models reach top performances in the context of this project.