Monocular Depth Estimation using Deep Neural Networks

Depth estimation is to use an RGB image from single or multiple perspectives to estimate the distance of each pixel in the image relative to the shooting point. Currently, monocular depth estimation has achieved good performance in the field of autonomous driving. Due to its portability of self-supervised monocular depth estimation technology, it is considered to have huge prospects in the field of unmanned aerial vehicle(UAV). But it is difficult for drones to match such good effects of autonomous driving because the motion of the drone is less constrained than the car camera and the images taken from drones have a more complex structure. This paper proposes a self-supervised monocular depth estimation (SS-MDE) model suitable for UAV by improving the accuracy of depth estimation. Recent works tend to choose high-resolution images for depth prediction with the development of GPU computing performance. However, the prediction accuracy does not increase with the increase of the resolution in the input images. This paper analyzes the reason is that inaccurate depth estimation makes bilinear interpolation errors reduce the gap between high-resolution and low-resolution images. Therefore, we refer to the U-Net depth estimation network structure to design a dense skip connection between the encoders and decoders of DepthNet, which makes the decoders better fuse image features to improve the accuracy of depth estimation. We train and test the SS-MED on the MidAir dataset[1], which is a dataset collected on the synthetic trajectory of UAV. The experiment shows that the proposed model not only performs state-of-the-art results on the MidAir but also performs well on the video taken by the real UAV scenes.