Dynamic Lane Segmentation for Autonomous Vehicles Using Neural Networks

Advanced Driver Assistance Systems (ADAS) are being researched widely throughout the world, given their enormous potential to reduce the number of road accidents and ensure the security of the passengers inside the vehicles. One of the critical components for the functioning of ADAS is awareness of the environment around them which includes; but is not limited to; localization of the road, determining the position of the objects and obstacles in front of the vehicle, and the direction in which the car is heading. This paper proposes a computer-vision-based approach capable of dynamically adjusting its speed based on the curvature and continuity of the road and the relative position of the obstacles in its path. Hough Transform is used for lane detection when the vehicle encounters a straight and continuous highway. Simultaneously, road segmentation using FCN is performed to detect possible curvature and discontinuities of the road, and instance segmentation is used to generate boundary boxes around any obstacles on the road. The approach was tested on existing datasets about road data for autonomous cars. The results show that it can be incorporated into the current ADAS cars to improve efficiency. Furthermore, the potential for future development of the model was also discussed to overcome some of the limitations encountered.