top of page

Monocular Depth Estimation – Less Sensors, Less Manual Data Annotation, Best Safety, Highest Efficiency

  • Writer: janhartmann6
    janhartmann6
  • Apr 27, 2023
  • 2 min read

Monocular Depth Estimation – Less Sensors, Less Manual Data Annotation, Best Safety, Highest Efficiency

How can fully automated locomotion be enabled in a cost-effective and safe manner while minimizing the use for sensor technology and the effort for data labeling? A new and groundbreaking approach is monocular depth estimation, which in combination with a deep learning approach can enable exactly this. Depth estimation of an image or data set, i.e., measuring the distance of the vehicle to objects, is essential for reliable environment perception. Until now, this has relied on the use of multiple cameras (stereo) or redundant sensor technology (camera, lidar, etc.). In addition, there is the complex labeling of data in order to make the AI or the vehicle „familiar“ with its environment.


With the monocular depth estimation in connection with a „self-supervised“ Deep Learning principle it is possible to recognize its environment with only one (cheap) camera and to get depth information. The results are similar to a pseudo lidar. Additionally, the generated point clouds are denser than those of lidars. This promises not only high accuracy but also maximum safety. With simultaneous cost savings (by omitting a lidar laser). Thus, large video data can be generated quickly and easily. These are sufficient to recognize the environment with simple video data and to estimate the distance with the help of Deep Learning methods. . Manual data annotations or the use of reference sensor technology, as mentioned above, become unnecessary.


Thus, highly scalable and inexpensive data acquisition is possible, eclipsing previous methods and applications. Crucially, this approach is both real-time capable and applicable to different camera types, such as fisheye cameras. Furthermore, multitask training can be used to train better semantic segmentation. This „simplification“ is a revolutionary „game changer“ not only in autonomous driving, but in all types of autonomous mobility. Let’s enjoy talking together about the different application possibilities. We look forward to hearing from you.



bottom of page