New-Tech Europe Magazine | Oct 2017 | Digital Edition

think in RGB (and the pixels are Red, Green and Blue) just like the human eye, but this method is not suitable for accurately calculating an image. Thus, firstly RGB has to be transferred into HIS (Hue, Saturation and Intensity). Rectifying the image to compensate for distortion in the lenses is the next necessary step. Following this, stereo matching can be performed between the two cameras. These steps are executed within an FPGA that is seconding the x86 core processor. All the following calculations are application-specific and best executed on the integrated, highly flexible programmable x86 processor platform which has to fulfill quite challenging tasks to understand and interpret the content of a picture. To understand how complex these tasks are, it is necessary to understand that interpreting picture content is extremely complex for software programmers and that, until recently, the human visual cortex has been superior to computer technology. These days, however, technological advancements are, quite literally, changing the game: An excellent example of computer technology improvement is Google’s AlphaGo computer which managed to beat the world’s best Go player. And this was achieved by executing neural network algorithms. Is this really so revolutionary? Haven’t we seen neural network algorithms in the recent past? Indeed we have. Neural networks are not new. They are just one of many AI (Artificial Intelligence) methods. Although exactly this kind of network was considered very promising in the nineties it is even more promising today as all the basic technologies now have far more computing power. Progress simply came to a halt due to the limited performance. To be even more precise, the barrier came down partly due to a lack of compute power and partly due to problems with networks with too many hidden layers. Recent methods use even more layers

Figure 2: Unibap’s mission-critical stereo Intelligent Vision System (IVS) with 70 mm baseline features advanced heterogeneous processing. Extensive error correction is enabled on the electronics and particularly on the integrated AMD G-Series SoC and Microsemi SmartFusion2 FPGA. in building the neural networks and today the term deep-learning means a neural network with many more layers than were used previously. Plus, the heterogeneous system architecture of modern SoCs allows deep-learning algorithms to be used efficiently (e.g. with the Deep Learning Framework Caffe from Berkley). x86 technology is also interesting for intelligent stereoscopic machine vision systems due to its optimized streaming and vector instructions developed over a long period of time and very extensive and mature software ecosystem, vision system algorithms and driver base. Plus, new initiatives like Shared Virtual Memory (SVM) and the Heterogeneous System Architecture (HSA) now offer an additional important companion technology to x86 systems by increasing the raw throughput capacities needed for intelligent machine vision. HSA enables efficient use of all resources With the introduction of latest generation AMD SoCs, a hardware ecosystem is now in place which accelerates artificial intelligence

Figure 1: HSA provides a unified view of fundamental computing elements, allowing a programmer to write applications that seamlessly integrate CPUs with GPUs while benefiting from the best attributes of each. sensors and the compute units. These systems not only provide high speed and high resolution to compete with our human vision, they also provide accurate spatial information on where landmarks or objects are located. To achieve this, stereoscopic vision is the natural choice. Industrial applications for this type of stereoscopic vision system can be found, for example, in item-picking from unsorted bins. Mounted on a robot arm, a vision system can carry out ‘visual servoing’ with 50 fps and identify the most suitable item to pick at the same time the gripper of the robot arm is approaching the bin. This makes scanning - which can take a couple of seconds – and reprogramming the robot arm superfluous. Autonomous cars are another obvious application for vision technologies, as well as a whole range of domestic robot applications. The artificial visual cortex So how does this process work in detail? The first stages of information handling are strictly localized to each pixel, and are therefore executed in a FPGA. Common to all machine vision is the fact that color cameras

58 l New-Tech Magazine Europe