ATZ live - Automated Driving

The RST was represented also this year again on the ATZ live - Automated Driving conference. A novelty is this year's Science Pitch. Scientists are offered here the possibility of presenting their own research to the entire auditorium in the form of a short pitch. Afterwards, the ideas can be discussed in detail at the corresponding poster.
Our contribution dealt with depth estimation using Deep Learning. For this purpose, the methods of depth completion and depth prediction were examined. In depth completion, a mechanism is learned to interpolate the sparse lidar data to produce a dense depth image. The monocular depth prediction is based on the Structure from Motion (SfM) principle and a neural network is trained unsupervised via video. A so-called U-Net architecture serves as a basis. This architecture was extended by a lidar padding mechanism and a smoothness propagation. Thereby the depth estimation results could be significantly improved.