Road Boundary Detection using 3D-to-2D Transformation of LIDAR Data and Conditional Generative Adversarial Networks

The Detection of road boundaries in an urban driving scenario has been an active area of research as perception of the surrounding environment is critical for vehicle's navigation and maneuverings. Most of the approaches in the literature deal with perception in the day light conditions only. This paper proposes an end-to-end method to extract road boundaries in both daylight and night driving conditions using conditional generative adversarial networks (CGANs). This is achieved by using input data only from 3D LIDAR which is not dependent on external light sources. The point cloud data obtained from a LIDAR was pre-processed by selecting the region of interest and deleting noise points resulting in data size reduction. From the pre-processed point cloud, top view images are generated, thus converting data into 2D image space. These images are fed as input to train a CGAN to segment road boundaries. Experimental results demonstrate that in different road scenarios the proposed method has achieved good performance with Structural Similarity index (SSIM) above 96%.