Artificial intelligence (AI) and machine learning (ML) are rapidly becoming the core behind many applications and products. One of the reasons for this is that AI and ML allow designers to train systems instead of coding every possible event. In doing so, this allows for rather complicated tasks to be more easily implemented (though the method in which a computer arrives at a conclusion is more obscure).
For the past decade, since the widespread introduction of commercial AI, there has been a reliance on data-processing centers to perform the machine-learning algorithms due to their complexity. Unfortunately, this comes with a few issues. First, any device that uses AI via a data center needs to have an internet connection, and this can increase security risks. Second, potentially confidential and private data, such as conversations, may need to be transmitted to the data center, which can raise privacy concerns. Finally, internet connections often come with some form of latency, thereby reducing the responsiveness of such systems.
To overcome these issues, a new form of AI processing called “edge computing” is quickly becoming popular, whereby devices can run simple AI algorithms locally without the need of a data center. Google’s Coral range of products include edge-computing–enabled devices that can run local AI systems on TensorFlow cores. But what are these products, and in what applications may they find themselves useful?
Coral Dev Board
The Coral Dev Board is a powerful single-board machine based on the i.MX 8M SoC. It integrates a tensor processing unit (TPU) that can perform up to 4 trillion operations per second (TOPS) and consumes only 0.5 W per TOPS. The board also integrates a GC7000 lite GPU, 1 GB of LPDDR4, and 8 GB of eMMC flash memory.
This particular dev board has an integrated Wi-Fi controller, which removes the need for wired Ethernet connections and allows for remote operation. The computing section of the board can be removed from the motherboard, which allows for direct integration into existing products, and the removable module can also be ordered in bulk for scaling prototypes to full-fledged products. The miniature size of the dev board makes it ideal for remote locations, while the 40-pin GPIO header allows for connecting to both external circuits.
Coral USB Accelerator
Not all projects can directly integrate the Coral Dev Board, especially those that rely on legacy hardware. In these instances, in which only an AI co-processor is required, the Coral USB Accelerator becomes an invaluable add-on. The Coral USB Accelerator integrates a TPU that can perform up to 4 TOPS while consuming only 0.5 W per TOPS. When connected to a Debian-based host, efficient AI processing is enabled. This can free up precious process resources for other tasks that would otherwise be stuck on complex AI neural networks, and the small size of the USB accelerator makes it ideal for portable, slim-line applications. The TensorFlow processor enables the USB accelerator to execute state-of-the-art mobile-vision modules, such as Mobile Net V2, at 100 fps and supports AutoML Vision Edge, which can provide high-accuracy custom image classification for projects.
Coral 5-MP Camera Module
Many AI and ML projects are based around vision and image processing, and if the Coral Dev Board is doing just this, then the 5-MP Camera Module is the perfect candidate. The 5-megapixel module uses the Omnivision OV5645 SoC sensor for image capturing and features auto-focusing, auto-exposure, white balance, band filter, illumination, and back-level calibration. These features are incredibly important in an AI and ML environment, as changes in exposure or backlight can dramatically affect the results from neural networks.
Automatic controls, such as those stated above, can help make images of objects more consistent and therefore improve the accuracy of the AI/ML network. On top of that, automatic controls also remove the need for any image processing by the CPU and can free up resources that would otherwise try to make adjustments to the camera. The miniature size of the camera (25 × 25 mm) makes it ideal for many applications, including mobile, RC, drones, and other remote locations.
Coral Environmental Sensor Board
Many IoT and AI-/ML-based projects rely on some form of measurement-taking from sensors. The Coral Environmental Sensor Board is an excellent add-on for projects using the Coral Dev Board, as it integrates a 128 × 32 OLED display, a HDC2010 humidity sensor, an OPT3002 ambient light sensor, and a BMP280 barometric sensor. The Environmental Sensor Board also includes a Microchip ECC608 crypto-chip with Google Keys, which enables the creation of secure IoT projects and works with Google’s IoT Core. The expandable groove connectors enable the Environmental Sensor Board to connect to other sensors, enabling easy project expansion, and the board also includes other connections for UART, I2C, and PWM.
Google AIY Vision Kit
The Google AIY Vision Kit is a Raspberry Pi-based system that incorporates a Raspberry Pi Camera Board, an Intel Movidius Myriad Vision Processing Unit (VPU), and TensorFlow capabilities. The Google AIY Vision Kit allows for the quick creation of visual machine-learning systems that can recognize facial emotions, object recognition, and even object tracking. The system requires no soldering at all and is supplied with a cardboard enclosure that allows any engineer to rapidly prototype visual-based AI/ML systems.
The AIY Vision Kit is programmable via Python, and the Python API Library removes complex code and provides an intuitive API for designers who want to start with visual machine learning.
