Steps towards autonomous driving: deep semantic segmentation among weather conditions

Nowadays the machine learning algorithms and more in general the artificial intelligence have become crucial in the scientific development, there are a lot of new research fields involved in this subject and this topic is highly correlated to every environment. One particular application on which this thesis is focused is autonomous driving. A very important field involved in this topic is computer vision. In this case the goal is to become capable to understand every part of the image, being able to distinguish between the different areas of the picture (i.e. which part of the view is the street, which part is a sidewalk, or if there is a traffic sign). This branch of computer vision based on neural networks is called semantic segmentation and it’s very useful but extremely hard to train due to the lack of data. In fact, to train a simple network for image recognition, we must provide a set of training images labeled as the subject while in this case a label for each pixel of the image is needed, to allow the network to become more and more precise. This type of labeling is hard and extremely time expensive, so a possible solution are synthetic datasets, generated by simulators that can automatically assign to every pixel a label, being very precise and low time consuming, but with the disadvantage to have less realistic images. Up to now, there aren’t a lot of this datasets for the training of networks for autonomous driving, and it is one of the main reasons for this work. In this thesis, the objective is to describe the newly created dataset IDDA - ItalDesign Dataset, that has been thought to address this problem. In fact, it contains a lot of different conditions well-separed in order to perform these types of trainings. It is based on the CARLA simulator, and with it we generated over one million images, with different conditions of weather, vehicle point of view and town. After that, we tried a state-of-the-art network (ADVENT) with our dataset, in order to try to understand how it performs on the domain shift generated by the weather conditions change.