A Comparative Study on Semantic Segmentation Algorithms for Autonomous Driving Vehicles

Abstract

Abstract: Semantic segmentation refers to the process of classifying each pixel in an image for better understanding and analysis of image. This is how machines look at the real world and identifies different objects. These are the times when autonomous vehicle industry is blooming and establishing new heights. There are so many research studies going on around semantic segmentation that are advancing to break boundaries in the world of computer vision. Despite of so much of progress made in the field in the recent years, autonomous vehicles needs more improved and efficient models to ride on the roads. In this research paper we compare the currently proved popular choices of models for semantic segmentation with respect to autonomous vehicles on different parameters to create an in-depth analysis on which all models and their variations improve and affect the quality of real time segmentation of the real world. The popular model architecture choices that were assessed and compared in this research were Fully Convolutional Network (FCN), U- Net and DeepLab. The data used for analysis was taken from Lyft Perception Challenge on Lyft Perception Challenge on Udacity.