IoT Roadshow, Denver – Colorado Engineering Inc.: Neural networking and NVIDIA bring machine vision to IoT

CEI actually evolved from Catalina Research, and being located in Colorado Springs near the U.S. Air Force Academy, you can imagine a long history of DSP and signal processing design services for military applications. And if you did, you’d be right. But in the context of the Internet of Things, it never ceases to amaze how much left field technology and expertise is used to serve the needs of a growing market, and not in the ways you’d expect.

CEI’s major partners include NVIDIA and Altera, with the former being key to their IoT system developments. At their facility, Greg Deemer, the company’s Director of Advanced Systems, introduced me to the TK1-SOM, a system on module based on NVIDIA’s Tegra K1 system on chip (SoC).

The TK1-SOM is essentially a compact, refined version of NVIDIA’s Jetson platform that enables developers to leverage the Tegra K1 SoC’s 326 GFLOPS at only 5-10 W idle power consumption, all in an embedded form factor (2” x 2.3”). This makes the platform ideal for applications such as radar, driver assistance, and machine vision, as well as facial recognition.

But that alone does not constitute an IoT play, at least in my opinion. What does, however, is the use of a histogragram of oriented gradients (HoG) coupled with a deep neural network to identify particular objects in the field of view of a security system. The TK1-SOM, which can be outfitted with a number of I/O expansion modules including , cellular, and a number of application-specific boards from Infineon and Devices can also run the NVIDIA cuDNN library, which is part of the NVIDIA Deep Learning SDK. By combining the power of the Tegra K1 for with these libraries and some connectivity that allows the DNN libraries to be updated (perhaps remotely) over time, you have, what I would consider, the foundation for extremely advanced IoT system development.

To learn more, visit

Return to slideshow.

Topics covered in this article