Stereo Correspondence

The central challenge in generating 3D data using passive stereo is solving the stereo correspondence problem. To measure 3D scenes using 2D images, either 2 or more cameras are usually recorded at the same time (stereo) or a moving camera is recorded in chronological order (structure-from motion). In order to calculate the depth information, it is necessary to know the corresponding 3D points in the captured images. For this research question, which is called a correspondence problem, there is a large number of solutions in the literature (Middlebury Dataset). However, the majority of these methods focus on optimization with respect to a limited test dataset.

Various methods of stereo correspondence analysis are investigated and developed at the Chair of Digital Signal Processing and Circuit Technology. The goal of the development is the solution of the practical challenges:

• Measurement of high observation volumes (range, viewing angle, high image resolution)
• Generation of dense depth maps with unstructured surface texture
• Unusual lighting conditions
• Real-time implementation on embedded systems

The figures compare a simple method of correspondence analysis (block matching) with a modern method of generating dense depth maps