The vision AI platform PerCV.ai (pronounced Perceive AI), could be the secret weapon that enables a company to deploy an AI application when so many others fail. The solution from Irida Labs, a member of the ST Partner Program, is an end-to-end platform that works with STM32N6 devices. It provides all the required infrastructure to build a new application by capturing and managing data, training models, and optimizing algorithms to run on STM32 devices, among other tasks. In a nutshell, it streamlines the creation of vision AI-powered applications at the edge by approaching AI as a whole, rather than the sum of disjointed parts.
Why do so many AI projects fail?
80% failure rate
According to analysts from RAND, “80 percent of AI projects fail”, which, the group explains, is “twice the rate of failure for information technology projects that do not involve AI.” At the top of the list for reasons why AI projects fail, we find a misunderstanding of the problem teams are trying to solve and miscommunication about how to approach it. Machine learning at the edge, while increasingly popular, is still very new. Consequently, developers are still figuring out how to use it efficiently. This is even truer for computer vision applications. Promising proofs of concept often fail to scale beyond the lab, as teams struggle to transition their solutions into robust, production-ready systems that can meet the demands of real-world deployment.
The problems inherent to computer vision
The challenges behind computer vision at the edge are that the inherent constraints of microcontrollers are anathema to video capture and processing. On one hand, IoT and other embedded systems constantly reduce power consumption, shrink their footprint, and lower their memory requirements. On the other hand, capturing a video, extracting its images, and processing the information require a ton of memory, massive computational throughput, and significantly increase power consumption. Additionally, based on current market costs, labeling enough images to get a basic proof-of-concept can cost between 5,000 and 10,000 dollars, according to Kili. And image labeling is just one piece of the puzzle.
How can PerCV.ai and the STM32N6 make a difference?
The vision twin
To solve the problem of designing a computer vision application, PerCV.ai starts by creating a “vision twin”. Just like a regular digital twin, the digital vision twin simulates how the application will run. It examines requirements for optimizing training and inference, including camera placements, field of view, and areas of focus. Too often, developers assemble something and attempt to test it in a lab, which doesn’t accurately reflect real-world use cases. Others conduct a sort of Monte Carlo experiment, which can be outrageously costly and yield poor results. By adopting a vision twin, PerCV.ai uniquely guides engineers, making the rest of the process a lot more intuitive and predictable.
The STM32N6 optimization
Irida Labs’ PerCV.ai is a true end-to-end solution, meaning it helps capture training data, label it, and generate a machine learning algorithm that developers can use in their applications. To solve the computational and efficiency challenges, the company worked closely with ST, using STM32Cube.AI within PerCV.ai to enable the conversion of a neural network with STM32-optimized inference operations, and run end solutions on the STM32N6. By utilizing the NPU of our microcontroller, Irida Labs was able to process more frames per second than on any other MCU and detect significantly more objects simultaneously.
What can PerCV.ai and the STM32N6 do in a real-world scenario?
Automatic number-plate recognition
What does it mean concretely? Irida Labs put together an automatic number-plate recognition software thanks to PerCV.ai, thus showcasing what the platform can help accomplish. An application running on the STM32N6570-DK uses a MIPI CSI image sensor to track up to five license plates simultaneously. Users can see how the software on the STM32N6 behaves through a live view on the display of the Discovery Kit. In a nutshell, thanks to PerCV.ai, the STM32N6 captures a video feed, detects where the license plates are located, and then runs an optical character recognition algorithm. All this is done on the microcontroller, meaning that at no point is cloud computing involved.
And many more applications…
Developers can use the STM32N6 and PerCV.ai to run many more types of applications, such as QR code reading, people monitoring, warehouse surveillance, vehicle tracking, and more. The ability to train a model and run inferences on the same hardware and with the same platform means that creating a machine learning application at the edge is a lot more straightforward and cost-effective, thus vastly improving the chances of its marketability. It also ensures that engineers can meet privacy and regulatory demands. As governments try to protect their citizens’ data, running entirely on the edge helps future-proof their design.
