Monocular depth estimation for vision-based vehicles based on a self-supervised learning method

Abstract

Unsupervised depth estimation methods that allow training on monocular image data have emerged as a promising tool for monocular vision-based vehicles that are not equipped with a stereo camera or a LIDAR. Predicted depths from single images could be used, for example, to avoid obstacles in autonomous navigation, or to improve in-vehicle change detection. We employ a self-supervised depth estimation network to predict depth in monocular image sequences acquired by a military vehicle and a UGV. We trained the models on the KITTI dataset, and performed a fine-tuning on monocular image data for each vehicle. The results illustrate that the estimated depths are visually plausible for on-road as well as for off-road environments. We also provide an example application by using the predicted depths for computing stixels, a medium-level representation of traffic scenes for self-driving vehicles.