RoadSDNet: A Robust Algorithm for Road Boundary Detection and Segmentation using Mixed Networks and Hough Transform

Abstract

In the present day, Road boundary detection is one of the most focused problems as it is a causative for many road accidents. To ensure the passenger's safety an accurate model that can ensure road segmentation along with detection of the road boundary is inevitable. Road boundary detection in both structured and unstructured roads is a challenging task in machine vision and AI. Classic machine learning algorithms are proposed for this problem, however there exists many difficulties in deploying them in real time. This becomes laborious task which require huge computation in real time. This paper addresses a novel algorithm, RoadSDNet for road boundary detection and segmentation. This algorithm can be easily deployed in real time as it consumes very less computation time giving a significant accuracy compared with the other existing methods. This system can be implemented on AMD Ryzen 250 platform, allowing in easy installation over the vehicles. The hyperbola fitting techniques required for the interpolation of the disguised road is adopted from the Hough Transform and produced as the extended HT Network. This network ensures the smooth polynomial curve in accordance with the road track-line and tangent relationship. The proposed takes input only from the camera but not the other hardware components like LiDAR sensor, Proximity sensor. This can be considered as the novel contribution of the paper. The experiments performed on this model proves proposed method is robust and polent in the huge traffic also and works in the uncertain road conditions too giving noteworthy accuracy and precision.