Embedded vision technology is everywhere, from robots to smart cars

Embedded vision technology is very hot these days.

It runs in everything from smartphones to factory production lines, where high-resolution images can be used to detect product flaws and other irregularities within milliseconds.

Today, engineers are also developing applications for smart cities and smart cars. Tiny cameras can process precise images of objects near and far, recognizing everything from barcodes to license plate recognition.

The use of IoT in factories, medical and retail environments is also increasing year by year. Many require image capture to be processed locally and connected to the cloud for further processing, data analysis or storage. Industrial applications of embedded vision are similar to smartphones, but can be customized for specific application needs and are ruggedized and designed for a longer service life.

“Embedded vision robotics has a lot of value for applications that require fast response times,” said Brian Geisel, CEO of Geisel Software. “From space to coal mining, robotics is often used in remote or dangerous locations.”

Advanced driver assistance systems are also incorporating embedded vision systems. “As hardware gets smaller, faster and cheaper, embedded vision will become mainstream,” Geisel said. Improvements in algorithms will allow developers to use less powerful devices to perform tasks.

“We’re going to have more ability to take advantage of embedded vision because we’re able to scale down the necessary computation,” Geisel added. “There are a lot of places where you can’t transmit a lot of data, so we’re going to see an explosion of new applications because we can do more computer vision at the edge.”

Embedded vision for industrial cameras “is moving from the cloud to the edge and will become mainstream over the next decade,” said Tim Coggins, head of sales, Americas, embedded imaging modules at Basler AG. “Early adopters have strong and compelling business cases, and now that the technology is mature, they don’t have to wait.”

Engineering students might be interested in Basler's web-based Vision Campus tutorials, which explain camera technology, interfaces and standards, vision systems, components, and applications in simple terms. Some of these come with text and diagrams, while others are explained in quick videos by professional presenter Thies Moeller. Tutorials on everything are available here, including sessions such as "What is the best way to implement a modern CMOS camera?" and "At what point does software come into play during image processing?"

A recent video presents five tips for designing embedded vision systems to avoid common pitfalls. It is better to develop the system from a single source rather than developing key components separately. With separate development, components may not interact in a high-performance manner, leading to costly delays, Moeller explains.

Increasingly, IoT designs involving image capture require non-recurring engineering (NRE), and the one-time cost of researching, designing, developing and testing a new approach is high. “Many early adopters have high requirements for volume or strategic applications, and in these cases, NRE is not a barrier. They can justify it with the cost,” he said.

In the absence of a standard plug-and-play solution in the market, custom embedded vision products can take time to get to market, and costs can vary depending on the complexity of the application. “The main challenge for developers is the need to bring together a variety of hardware and software variables to adopt common connectivity capabilities,” Coggins said.

Adam Taylor, founder of Adiuvo Engineering, points to OpenCV for image processing as a good example of standard embedded vision software. First developed by Intel in 2000, OpenCV is a library of programming functions for real-time computer vision that is cross-platform and open source under the Apache 2 license. Developers use it to process images, capture video and analyze video for object or facial detection or other purposes.

“Standards are the way to scale and define the biggest gains to accelerate development and simplify engineering,” Taylor said. “Embedded vision should be plug-and-play, allowing companies to focus on their differentiating activities rather than just trying to get images from more cameras.”

Basler is driving standards in the embedded vision industry. “The embedded ecosystem already exists and is growing with many talented companies and individuals who can provide education and assistance, answer questions or develop system-level solutions,” said Coggins.

Basler provides complete system design as well as mass production, but so do many of Basler's partners. Basler's partners include Nvidia Jetson. For example, Basler released an embedded vision development kit last June, which expanded Basler's support for Jetson products to provide AI at the edge of robotics, logistics, smart retail and smart cities. There are hundreds of thousands of AI edge developers in the Nvidia ecosystem.

The kit comes with the Basler dart BCON for MIPI cameras with 1.3 MP lenses and an adapter board developed for the Jetson Nano module.

The key benefit of embedded vision technology is its ability to support low cost, high performance, and real-time operation through edge-to-cloud connectivity, thus providing scalability for the entire computer vision market.

“Companies that have turned to embedded vision technology are very happy, and the early adopters are doing so for good reasons,” Coggins said. “Most of our customers want industrial rugged solutions with long-life cycle reliability, and they can’t get that from the consumer market.”

Grand View Research estimates that the overall global computer vision market was worth $10.6 billion in 2019, with 70% coming from hardware such as high-resolution cameras. It is expected to grow by 7% annually through 2026. The 2019 total does not specifically break down embedded vision systems, but Grand View says that vertical industry accounted for half of all revenue in 2019.

The researchers ranked Intel, Omron, Sony and Texas Instruments as the most prominent players in the computer vision field.