A vision-based zebra crossing detection method for people with visual impairments

Abstract

Safe navigation for visually impaired is challenging without assistive technology. This paper proposes a pedestrian crossing detection approach to help visually impaired people. We introduce the use of multiple convolutional neural networks (CNNs) by utilizing wavelet transform subbands as inputs in which networks are trained to detect zebra crossing. In our method, the original image is decomposed into wavelet subbands, and the input images are constructed from image approximation based on the coefficients of three subbands. In the multiple networks approach, the segmentation results of the networks were integrated to create the final segmentation map. The results presented in this study prove that our method fully outperforms the SegNet networks and other state-of-the-art results using the Synthia database.