Autonomous Driving Through Road Segmentation Based on Computer Vision Techniques

Autonomous Driving refers to self-driving vehicles without the need of intervention from the human driver. Safety enhancement, energy optimization, comfort, maintenance and cost are the key benefits of Autonomous Driving. Other benefits include - productivity, reduced congestion/traffic, prevention of car crashes, reducing carbon footprint and ease of parking in congested cities as driverless vehicles could drop passengers off and move on. Autonomous Driving can be achieved via processing visual images/videos at runtime and then converting them to vehicle control signals. This study will help in detecting objects on roads (such as moving vehicles, pedestrians, other static objects on road and road segmentation) using deep convolutional neural network (CNN), which in turn help in aiding driverless future. Semantic segmentation can help recognize objects and their location. Semantic segmentation refers to labelling images with pixel-by-pixel classification that in turn helps to perceive the surrounding environment. Human driver also perceives the driving environment in a similar manner. The purpose of this study is to do an overview and check the feasibility of semantic segmentation using deep learning algorithms in the field of Autonomous Driving specifically in road segmentation task. This study will also be comparative study on CNNs models in terms of accuracy, precision, mean IOU and processing time. The scope of this study will include examination and comparison of two popular algorithms - Fully Convolutional Network (FCNs) and Semantic Segmentation model (SegNet). In detail, this study will mainly focus on road segmentation task in Autonomous Driving with the help of CNN models and will conclude which model is best suited for road segmentation task under different weather conditions, in terms of precision, accuracy, mean IoU and processing time.