Enabling embedded vision on generic processors and microcontrollers is critical for the wide adoption of this technology. Current state-of-the-art object detection algorithms require Cortex-A class application processors or specialized neural-network accelerators. In this presentation, we explain the design challenges and technical insights of implementing a single-shot-detector object detection algorithm on a very resource-constrained microcontroller. We describe the application challenge and constraints which led to the architecture design. We’ll share insights into the design approach and methods employed as well as the technical challenges encountered. The presentation will wrap up with results achieved and lessons learned.