Building occupancy management solution using the TensorFlow Object Detection API

Introduction

GreenWaves has developed a people counting solution for occupancy management in smart building systems, providing real-time insights into how available space is used by employees and customers. The sensor can be used for tasks such as meeting room or cafeteria usage optimization, desk reservations and usage based cleaning.

People counting with infrared sensors offers best-in-class accuracy with total compliance to privacy-related regulations for indoor environments. GAP processors provide a combination of computing ability for AI and low energy operation that enables this type of application.

As part of our development process, we needed to train an optimized neural network with a single shot detector SSD backend. The TensorFlow Object Detection API comes with a number of prepackaged backbone models, but we wanted to design something more optimized for our detection task. We aimed to

  • reduce memory,
  • reduce complexity, and
  • reduce power consumption

In this document, we will show how we carried this out and how a custom network design can still leverage all the backend SSD creation offered by the TensorFlow Object Detection API. We hope this will allow you

  • become familiar with an object detection API like the one provided by TensorFlow.
  • learn how to modify the API with respect to your custom specifications (i.e., model structure).
  • learn how to employ the API for custom solutions such as occupancy management.
  • learn how to generate optimized code for running your solution on GreenWaves’ GAP processors.
Download full paper
2021-05-31T10:48:52+00:00