SPRADB8 may   2023 AM62A3 , AM62A7

 

  1.   1
  2.   Abstract
  3.   Trademarks
  4. 1Introduction
  5. 2AM62A Processor
  6. 3Deep Learning Benchmarks
  7. 4Retail Checkout Scanner Application
  8. 5Core Loading
  9. 6Part Selection
  10. 7Power Usage
  11. 8Summary
  12. 9References

4 Retail Checkout Scanner Application

A reference application was developed for the AM62A to showcase its capabilities for an automated checkout system using an object detection neural network. A custom model was trained to recognize a dozen different food items, and a Linux Python3 application was written around this model using TI’s gstreamer plugins to leverage hardware acceleration where possible. Application source code [2] and in-depth description on how the demo works are available on Github. For further guidance on building an application like this, see the associated application note [3]. The block diagram in Figure 4-1 depicts the application flow within gstreamer and how various plugins execute on remote cores within the SoC.

GUID-20230517-SS0I-PHBK-PX9D-P91BFRVLPLGD-low.svg Figure 4-1 Retail Checkout Application Flow With Resolutions and Pixel Formats. (30fps is the maximum achievable; the application is bottlenecked by the application code such that FPS is closer to 15)

This document analyzes this application and uses the core load both to guide selection of a suitable AM62A variant as well as provide a power usage estimate using the Power Estimation Tool [4]. This analysis can be followed for other applications designed and benchmarked on the Starter Kit EVM, which uses the superset variant AM62A74 (2 TOPS acceleration, 4 Arm® Cortex A53 cores).