In this paper, we describe a real-time vision machine having a stereo camera as input generating visual information on two different levels of abstraction. The system provides visual low-level and mid-level information in terms of dense stereo and optical flow, egomotion, indicating areas with independently moving objects as well as a condensed geometric description of the scene. The system operates at more than 20 Hz using a hybrid architecture consisting of one dual-GPU card and one quad-core CPU. The different processing stages of visual information have rather different characteristics that in some cases make fine-grained parallelization on a GPU less applicable. However, for most of the stages that are not efficiently implementable on a GPU, a coarse parallelization on multiple CPU-cores is applicable. We show that with such hybrid parallelism, we can achieve a speed up of approximately a factor 90 and a reduction of latency of a factor 26 compared to processing on a single CPU-core. Since the vision machine provides generic visual information it can be used in many contexts. Currently it is used in a driver assistance context as well as in two robotic applications.

No abstract

No abstract