Obstacle Segmentation with Encoder-Decoder Architectures in Low Structured Environments for the Navigation of Visually Impaired People.

Orientation and mobility of visually impaired people usually requires intensive training with mobility aids (e.g. white canes). Assistance systems capture information from the environment, process sensor data and provide the results to the impaired user. The paper presents an approach for efficient segmentation of obstacles in low-structured outdoor environments using encoder-decoder deep learning architectures and depth images. Therefore, an efficient method for generating training data using the v-disparity method is presented. Based on an extensive dataset of RGB and depth images and the corresponding binary label images, different state-of-the-art encoder-decoder architectures are evaluated on a mobile computing unit with respect to accuracy and efficiency. Besides pure depth-based architectures, RGB-D fused architectures are evaluated, too. The quantitative results show some limitations, but an additional qualitative evaluation proves the applicability of the approach to support the navigation of VIP by mapping the position of surrounding obstacles. Thus, an efficient combination of classical image processing, the integration of knowledge about the physical nature of the environment and deep learning can be made. Clinical Relevance- The approach supports the navigation of visually impaired people, which enables a more self-sufficient life related to higher quality of life.