A Comparative Study of Deep Learning-based Depth Estimation Approaches: Application to Smart Mobility

Abstract

In autonomous vehicle systems, the quality of scene perception is of great importance for security preoccupation in road environments. In this context, an accurate localization of potential obstacles is one of the most challenging tasks. In recent years, substantial progress has been made in the field of depth estimation for detection purposes with the spread of methods relying on deep learning with monocular or stereo-scopic camera(s). These two families of approaches did show an upstanding yet inconsistent performance in different road scenes circumstances. A deep understanding and comparison of these approaches is required to allow the community an easier assessment, which breeds to more adequate choice for their own systems. In this paper, we propose a comparative study of state-of-the-art deep learning depth estimation methods using monocular and stereoscopic cameras. The evaluation is performed on road environment over the challenging KITTI dataset.