Fast 3D Sensing Through FPGA-BasedImage Processing скачать в хорошем качестве

Fast 3D Sensing Through FPGA-BasedImage Processing

This presentation was recorded as part of the 2020 UKIVA machine vision conference: https://www.machinevisionconference.c... The three-dimensional acquisition of objects, surfaces and structures has gained immensely in importance in recent years. New camera and sensor technologies have made this possible, while existing approaches have been further developed and have also contributed to the triumphant advance of 3D image acquisition. The 3D methods known today are manifold. In 2012, for example, Fraunhofer IDG listed 12 (non-contact) active methods for geometry acquisition and another 3 passive methods. Depending on the application - of course - all have their advantages and disadvantages. In this context, "active" means the use of an external light source, while "passive" methods do not. This fact already classifies the potential applications; other criteria such as measurement accuracy or speed and minimum or maximum measurement distance limit the applications for the respective applicable technology. Stereo vision, in simple terms the simulation of human depth perception with the help of (of course 2) cameras, is a passive method and truly nothing new. Allegedly Euclid has already dealt with the topic. For obvious reasons, however, he has probably not yet used the semi-global matching (SGM) algorithm - an established standard for stereo image processing for years. However, the demands on computing power are extremely high, which has hindered the widespread use of this method up until now. It was therefore necessary to develop a solution that combined the advantages of stereo vision with high measuring speed. As already mentioned, even today's highly clocked CPUs do not provide the performance to process stereo image data of a few dozen fps and thus to allow meaningful applications, because massively parallel operations, as required by the SGM algorithm, are almost impossible to realize. But how about GPUs, ASICs and FPGAs? Todays graphics processors offer the performance to implement the SGM algorithm massively in parallel. However, this is offset by the relatively high power consumption of several hundred watts as well as the space requirement, especially for mobile applications. Just think of robots, where small sensors are usually in demand. ASICs would actually be the ideal hardware choice for this algorithm. Unfortunately, this is hampered by high development costs and, in addition, a certain inflexibility with regard to adapting the configuration to the task at hand. Changing this afterwards often proves difficult, if not impossible - an annoying obstacle especially for machine vision system, which frequently face changing requirements. FPGAs, on the other hand, are freely programmable and can therefore be adapted to a wide variety of applications. Their power consumption is low, as is their footprint. This was the reason for Nerian Vision GmbH from Stuttgart to develop a hardware platform based on a (Xilinx) FPGA and make it available to the market. The result is SceneScan - an embedded processing system that connects to two standard industrial cameras, which are mounted at slightly different viewing positions. Similar to the human visual system, SceneScan is able to infer the depth of the observed scene by comparing the images of both cameras. This enables SceneScan to provide a depth estimate for each observed camera pixel. The resulting depth map is transferred through a gigabit ethernet link to a computer or an embedded system for onward processing. More information at https://nerian.com