Developed in partnership with the computing specialist Cloudtop, the reference design can read slow-moving license plates at a distance of 3-5 metres with high accuracy. Helped by the capabilities of XMOS’ xcore.ai silicon, Cloudtop’s machine learning model, which was originally designed to work with high resolution video frames, has been adapted to work in a low power, low cost scenario without sacrificing accuracy.
Parking garages that utilise ALPR have traditionally integrated hardware that is far beyond the spec required for slow-moving, close-range plate recognition. High-resolution cameras, operating on complex machine learning models that depend on cloud connectivity for image processing, have made the implementation of ALPR prohibitively expensive in many cases.
XMOS’ reference design is able to provide the required power and intelligence on-device, dramatically lowering both power consumption and the Bill of Materials (BOM) in comparison to standard ALPR solutions. In removing the need for high-cost hardware and virtually eliminating the need for cloud connectivity, these types of devices can now become a more realistic component of the ALPR infrastructure across the smart city.
“For smart parking, cloud connectivity and huge processing power is simply overkill,” commented Aneet Chopra, VP Product, Marketing & Business Development, XMOS. “It makes ALPR networks far more expensive than they need to be, makes maintenance more complex, and comes rife with privacy concerns inherent to the cloud.
“The reference design we’ve developed eliminates those issues simply by streamlining the process. If you can deliver the intelligence and power you need on-device, you avoid sending all raw data to cloud, or excessively expensive or powerful hardware. That’s only going to help us drive progress in ALPR in the long run.”
“Simplicity and affordability are two priorities in the ALPR space, not only to drive sales but to encourage innovation” commented, Prof. Zhang, Co-founder of Cloudtop. “Making devices cheaper, simpler and more reliable will be hugely important for the smart city, and downscaling machine learning models so that they can run on mass-producible silicon like xcore.ai affords developers the funding and design flexibility to experiment.”
XMOS and Cloudtop will showcase the solution at tinyML Summit in San Francisco, between 28-30th March.
