Vehicle depth estimation for autonomous driving

In the latest years, it has been done a lot of research regarding the depth estimation of the objects in an image. In autonomous driving, one of the most important tasks is to estimate the distance to the surrounding cars. This paper analyzes the state of the art regarding the depth estimation from single and stereo sources and the datasets that were made for these tasks and proposes a new depth dataset, recorded with an Intel RealSense depth camera. This dataset is used together with another dataset made previously in our university in order to compare one of the most used neural networks for depth prediction from a single camera, for both full image depth estimation and vehicle-only depth estimation. The results are computed regarding the Root Mean Square Error (RMSE) and take in account the time of the day (day, dusk or night), the inference time and the dimension of the cars. © 2021, Politechnica University of Bucharest. All rights reserved.