Scenes Segmentation in Self-driving Car Navigation System Using Neural Network Models with Attention

The article describes the design process of a software module for the road signs recognition used for the self-driving car, developed at the Ulyanovsk State Technical University (UlSTU) at the Faculty of Information Systems and Technologies in cooperation with the Faculty of Mechanical Engineering. One of the main tasks to be solved when creating technical vision systems based on neural networks, including for self-driving cars, is to create a training dataset sufficient to train network models. At the same time, in the task of semantic segmentation of the scene, the preparation of a large train set may require considerable effort for manual labeling. The article describes a convolutional network model with a soft attention mechanism, which is trained for the classification task with the possibility of extracting an attention mask from the internal network state, which can be used for semantic image segmentation. This approach can significantly reduce the cost of data labeling.