1 Smarter Cameras at the Edge

Artificial Intelligence (AI) on embedded devices at the network edge is growing rapidly as complex data processing and analytics becomes essential in making cities, factories, automobiles, and homes smarter and more efficient. There is a wealth of information in imagery, which humans rely on heavily. Computer vision (CV) and machine learning (ML) extract meaning, for example, where a person is, from the information-dense image. CV and ML are invaluable in improving use cases like defect detection for machine vision, visual odometry and mapping for robots, lane detection for automobiles, and many more. Human-centric applications like identification, biometrics, fall detection, and behavior recognition further push the need for smarter cameras in building access and public security applications.

Cloud services have dominated vision analytics and machine learning inference in recent years as embedded devices lacked the processing power to operate on streams of camera data, such as a security camera producing 1080p at 30+ frames per second (FPS). In 2021, the number of home security cameras was estimated by Yole at 45.1 million, with an expectation to grow by 2.5x by 2026 [1], the recurring costs of processing in the cloud grows linearly with the number of devices, inhibiting scalability. By processing at the edge, this recurring cost is minimized or removed altogether. Edge AI also avoids additional network latency in time-sensitive applications and reduces privacy concerns in human-centric or consumer electronic applications.

SoCs capable of running deep learning and other complex analytics locally are gaining popularity. Market projections estimate 25% of home security cameras will employ Edge AI by 2026 [1] and that this subset of the market will grow with 88% CAGR. Embedded processors are now equipped to handle these applications, but must carefully manage the power vs. performance tradeoff. Battery powered and hand-held devices must stay within a few Watts of power to avoid too quickly draining the battery or becoming too hot to handle. However, they must also retain enough performance to a sufficient analytics processing rate, for example, 15 fps for a home security camera. The right choice for an embedded vision processor depends strongly on the AI performance, image preprocessing, video encode/decode, and power consumption requirements.