Navigational Path Detection for the Visually Impaired using Fully Convolutional Networks

In this paper a novel approach for navigational path detection problem for the visually impaired was presented. A deep learning model based on state-of-the-art fully convolution neural networks have been proposed that can accurately semantically segment any navigational areas on pixel-wise level in different scenes without any prior assumptions about the environment of the scene such as textures or specific appearance cues. The proposed approach have been evaluated on two different publicly available dataset and have achieved a pixel accuracy of 91% over the testing images dataset. Furthermore, the performance of the proposed approach have been compared against other commonly used approach for the problem of predicting navigational areas in input RGB images, and the proposed approach outperformed it with more than 14%, 11% and 10% on the mean intersection over union, mean accuracy and pixel accuracy evaluation metrics respectively.