Semantic Segmentation using Convolutional Neural Network for Road Recognition

Abstract

Increasing demand on autonomous vehicles and their capabilities led the computing techniques to cope up with the new goal. Although there were many electric and electronic approaches including SONAR and LIDAR to solve the problems, cameras are now capable of taking part of the autonomous industry. Camera as a sensing medium for autonomous driving has a number of advantages over other methods. First of all, it contains spatial information if deployed in a stereoscopic structure, as other medium does, as well as texture and color information for semantically recognizing any oncoming objects. For an example, time-of-flight based sensing medium cannot distinguish indistinct changes on flat surfaces. In this paper, I propose a framework to recognize road surface that an autonomous vehicle can proceed on, with various machine learning techniques and problem-domain driven approaches to quickly achieve a trained machine. Convolutional encoder-decoder network was tuned for the task, transfer learning was applied to improve generic performance, and Contrast Limited Adaptive Histogram Equalization, a classical contrast enhancement method, was put before segmentation phase for coping with input variance. The results show overall improvements across a number of conditions that causes challenges.