1 Introduction

Smart camera applications in retail sectors are increasing in popularity given the rich information content in imagery as well as the improving capability of processing that data at the edge. Use-cases like checkout scanners, barcode images, asset and people tracking, theft detection, and more are helping to automate, simplify, and streamline customers’ experiences.

Although vision is an intuitive concept for humans, computer vision is challenging. Imagery is information-dense and visual patterns can occur in many shapes, sizes, and contexts. Conventional computer vision using filters, transforms, and specialized algorithms are effective; However, there algorithms are often difficult to accelerate in their entirety and can require careful tuning to the environment and specifics of the task. Machine learning and deep learning approaches to image processing are much more generalized and often boast higher accuracy. While these learning-based algorithms (AI) are often more computationally complex, they are also easier to accelerate given their strong reliance on matrix math.

Processing imagery with deep learning and AI in the cloud be done easily, yet comes at more significant and recurring cost than processing locally. Reactive applications like checkouts would be slow and frustrating for customers due to network latency. Security applications can cause privacy concerns. Furthermore, as the solutions scale, the associated cloud costs will scale similarly. Processing imagery locally on the device that captures video solves these issues, but requires a processor that matches requirements for cost, power, and performance.