Semantic Segmentation of Urban Street Scenes Using Deep Learning

Abstract

Listen

Translate article

Scene classification is essential conception task used by robotics for understanding the environmental. The outdoor scene like urban street scene is composing of image with depth having greater variety than iconic object image. The semantic segmentation is an important task for autonomous driving and mobile robotics applications because it introduces enormous information need for safe navigation and complex reasoning. This paper introduces a model for classification all pixel’s image and predicates the right object that contains this pixel. This model adapts famous network image classification VGG16 with fully convolution network (FCN-8) and transfer learned representation by fine tuning for doing segmentation. Skip Architecture is added between layers to combine coarse, semantic, and local appearance information to generate accurate segmentation. This model is robust and efficiency because it efficient consumes low memory and faster inference time for testing and training on Camvid dataset. The output module is designed by using a special computer equipped by GPU memory NVIDIA GeForce RTX 2060 6G, and programmed by using python 3.7 programming language. The proposed system reached an accuracy 0.8804 and MIOU 73% on Camvid dataset.