Improved U-net for Zebra-crossing Image Segmentation

Abstract

A safe crossing system is a prerequisite for improving the mobility of the visually impaired. For the blind, it is important to know exactly where the zebra crossings are. Zebra-crossing detection by machine vision can be a good solution to this problem. In this paper, we propose a model for fast and stable segmentation of crosswalks from captured images. For the blind, it is important to know exactly what area ahead is a zebra crossing. A common feature of all zebra crossings is the periodic appearance of white stripes on a black road. In this paper, we proposed a model for fast and stable segmentation of crosswalks from captured images. The model is improved based on U-net and consists of three steps. First, the input image is subsampled using ResNet-34's convolutional neural network to extract image features. Second, dilated convolution is used to increase the receptive field of feature points without decreasing the feature map resolution. Finally, the abstract features are restored to the original image size through the original up-sampling network of U-net with the complementary information of the skip connection.