L. Reese

Intel Movidius Empowers Mobile, Cloud-Free Artificial Intelligence


Artificial Intelligence (AI), long relegated to the realm of science fiction and more recently to high-powered computing machinery is slowly finding its way to lower-end embedded hardware. For about $100 USD, it is now possible to acquire the all necessary hardware and software to build a customized, vision-based AI solution. Last month Google released their AIY Vision Kit that is powered by the Intel® Movidius MA2450 Vision Processing Unit (VPU). This is the same embedded hardware that has powered the Intel’s Neural Compute Stick USB platform, Google’s Project Tango, and more recent generations of DJI-branded drones. Put simply, VPUs are customized microprocessor hardware built to handle the specialized machine learning algorithms that enable onboard machine vision processing. These algorithms include convolutional neural networks (CNN) and scale-invariant feature transform (SIFT). VPUs are necessary for efficiently handling neural network computer vision algorithms just as Graphical Processing Units (GPU) were developed apart from Central Processing Units (CPU) to provide better handling of graphics intensive tasks.

AIY Vision Kit’s do-it-yourself assembly (Source: Google)

The onboard processing is a key distinction. Instead of relying on cloud-based processing of locally captured images, the $45 AIY Vision Kit’s VisionBonnet (the plug-in board that contains the MA2450 and associated circuitry) with its neural network algorithms can process imagery without the need for Internet connectivity. Sometimes you do not want an IoT device to require connectivity for processing data. No requirement for constant connectivity means some smarts are in the IoT device itself, which removes potentially unacceptable performance delays and security risks associated with the images traversing the Internet. This concept is sometimes referred to as “edge computing” or “fog computing .” Going to the cloud for processing resources is not always desirable. Pushing computation down to as close as the sensor node as possible is preferable in many cases, especially for applications such as autonomous vehicles. In this way, latency is improved and connectivity is reserved for things like downloading an improved training model.

 The current iteration of the AIY Vision Kit is built specifically for the Raspberry Pi Zero W Linux-based single board computer. The kit supports two deep machine learning frameworks including Google’s own TensorFlow and Caffe from Berkley’s AI Research Lab. It can handle 30 frames per second of image processing. From a practical perspective, the AIY Vision Kit offers starry-eyed startups unprecedented capability in an extremely affordable and hackable package. The VisionBonnet comes with three pre-built neural network models. The first is a model that can both detect faces and determine the emotion emanating from that face. The next relies on MobileNet models (a suite of open source computer vision models built for TensorFlow running on resource constrained embedded devices) to recognize thousands of different objects. Lastly, there is a model that can differentiate between humans, dogs, and cats. Google has released the TensorFlow source code for the models and a compiler for the intrepid innovators who wish to tweak the models or develop their own. From the product prototyping and development perspective, the architecture of the AIY Vision Kit allows for very powerful yet straightforward interfacing. On the hardware side, the Raspberry Pi Zero W has four general purpose input/output (GPIO) that are available for interacting with external sensors and actuators. In addition, from a software point of view, the VisionBonnet can be interacted with using the increasingly popular Python programming language.

 Over time, we have seen technology evolve rapidly due in part to falling prices in conjunction with increasing capability. Part of that Moore’s Law story is also making affordable tech accessible to those who work in low-end, low-cost embedded hardware such as Arduino and Raspberry-Pi. Affordable equates to accessible. Accessibility leads to opportunity. Rock on.




Contact Information

Green Shoe Garage

Hollywood, Maryland,

tele: (301) 880-4538
support att greenshoegarage dot com


Test Post

test post