Zebra crossing segmentation based on depthwise separable convolutions

Abstract

The research of zebra crossing recognition plays an extremely important role in vehicle detection and blind guidance system. From the experiment, we find a method of zebra crossing detection and recognition based on depthwise separable convolutions. In the coding part, we use depthwise separable convolutions to extract the layer from the trunk model, which has been convolved many times and has certain characteristics. In the decoding part, the PSPNet model is improved, the coding layers with different downsampling generations are used for comparison, and all the extraction layers are stacked to get the prediction results, thus solving the problem of small targets disappearing. Through experimental analysis and data training, the mean intersection over union of zebra crossing segmentation reaches 90.7%, and the average recognition speed is less than 0.1s.