Stereo Vision

Spatial Perception via stereo vision exploits the fact that objects in the left and right camera are imaged at slightly different positions. As shown on the right, this displacement called "disparity" becomes smaller with increasing distance of the scene point. Typically, these disparities are estimated independently for every image point using local methods.

The Daimler researchers use a modern stereo algorithm that exploits couplings between neighboring image points , hence performs an optimization step to determine depth. This algorithm called „Semi-Global Matching“ (SGM) determines depth to almost every image point in the scene precisely[8b]. On international benchmarks, SGM is among the top-scoring methods. The algorithm was optimized and further developed in order to maintain a high performance level also at night time and in adverse weather conditions [7]. Above animation shows the 3D reconstruction quality with a base line of only 20cm. In 2008, the first real-time implementation of SGM on a low-power and inexpensive FPGA (field programmable gate array) was completed, computing stereo images 25 times a second [6]. Due to this implementation , any object up to 50m distance can be detected and measured. Compare this to the human stereo vision capability that ends around 10-12m.Further literature: On a internationally accepted ranking of modern stereo algorithms can be found including further literature references.