Let’s take a look at the tremendous penetration of intelligent vision in so many and various applications. A few years ago, computer vision algorithms were implemented in applications directly linked with imaging, like computational photography for smartphones and cameras. We can mention today a bunch of segments like automotive, human machine interface or machine vision where computer vision is now the backbone of applications which have been created, thanks to the capabilities of the imaging technology.
CEVA has launched the 5[SUP]th[/SUP] generation architecture for imaging and computer vision, the CEVA-XM6 DSP, and offer a comprehensive vision platform built around the DSP. In a previous blog, we have explained how to build machine learning device implementing the Convolutional Deep Neural Network (CDNN) from CEVA. But let’s take a look at CEVA-XM6 platform, which is much more than a DSP core as it includes the CDNN toolkit comprised of hardware accelerators, neural network, software framework, software libraries, and a set of algorithms.
Automotive driver assistance systems (ADAS) is the most prominent example illustrating the penetration of computer vision completely shaking an automotive segment in which the electronic innovation was in a quiet mode. Now you will find DSP-based imaging in applications like traffic sign detection, free space, pedestrian detection, lane departure, forward collision warning and probably more. Why did it took so long for these types of application to be adopted in automotive? The answer is as usual linked with cost, performance (per dollar) and power consumption.
If we take a look at the CEVA-XM6 DSP architecture we can list the (four) scalar processors SPU0 to SPU3 and the three 512-bit vector processing units VPU-0 to VPU-2, all of which 128 single-cycle 16×16-bit MACs, bringing the total MAC count to 640. In fact, the most important enhancement may be the neural-network hardware accelerator (HWA) that offers 512 additional single-cycle 16×16 bits MACs, connecting to the DSP core’s processing cluster through an AXI4 interface. This HWA is one of the User-defined Coprocessors located in the bottom right box labelled TCE. Taking the example of the CDNN based machine learning, the convolutional layers consuming most neural processor cycles are implemented in the HWA, freeing the DSP core and providing a boost to the machine learning function. When the CEVA-XM6 DSP solution is implemented in a 16-nm chip, it offers unbeatable performance going with very decent power consumption and low footprint.
This performance/power efficiency, coupled with a reasonable chip price, is making CDNN based machine learning an affordable technology to be implemented in mass-market application today. A few years ago, such technology was only demonstrated in a lab, not implemented into a piece of silicon available at mass-market price.
The development of computer vision in Human Machine Interface applications is also opening new possibilities and new markets. The CEVA-XM6 DSP can be integrated to support gesture recognition, emotion sensing, eye tracking, face recognition or face detection. These applications are often linked with the need to provide more security in a world becoming more interconnected, not only thanks to faster communication but also due to higher flow of human moving across the planet. No doubt that these new markets will need more efficient algorithms and higher performance to develop the computer vision based applications increasing safety and security.
Deep learning and augmented reality are two segments, directly linked with machine vision, literally exploding and expected to generate innovation in the industry as well as in our future day to day life. Both are very demanding in term of raw performance and algorithm efficiency. Because the CEVA-XM6 platform is coming with imaging & vision SW libraries and CDNN network generator, it will help the developers to fasten their system time-to-market. CDNN support a variety of popular CNN technologies, including AlexNet and GoogleNet and CEVA offers software libraries for OpenCV, OpenVX and OpenCL support.
If you are interested by CDNN and deep learning solutions for ADAS applications, you should attend to this on demand webinar and download this product note:
Learn how to use deep learning solutions for ADAS applications; How to run AdasWorks Free space detection neural network, while utilizing CEVA’s low power vision DSP combined with CEVA Deep Neural Network SW toolkit.
CDNN product note