Deploying Visual AI on Edge Devices: Lessons From the Real World

Developing an AI-based edge device is complex, involving hardware, algorithms and software. It requires making many technology-selection decisions and balancing numerous trade-offs. In this talk, we explore several of the most important decisions and challenges typically encountered in creating an edge visual AI solution. On the hardware side, we’ll explore choosing a processor, including evaluating performance, cost and power, and understanding the capabilities of on-chip processing engines such as CPUs, AI accelerators and ISPs. We’ll also examine the process of selecting, training, and evaluating deep neural network topologies, and we’ll illustrate how Astrocyte, Teledyne’s deep learning tool, helps simplify this process. Next, we’ll discuss key considerations for creating an optimized, efficient hardware/software pipeline implementing the trained DNNs on the edge device. We’ll use an AI camera solution developed for traffic tolling to show how these concepts play out in real applications.