Published February 5, 2021 | Version v1
Preprint Open

Enabling energy efficient machine learning on a Ultra-Low-Power vision sensor for IoT

  • 1. Fondazione Bruno Kessler (FBK)

Description

The Internet of Things (IoT) and smart city paradigm includes ubiquitous technology to extract context information in order to return useful services to users and citizens. An essential role in this scenario is often played by computer vision applications, requiring the acquisition of images from specific devices. The need for high-end cameras often penalizes this process since they are power-hungry and ask for high computational resources to be processed. Thus, the availability of novel low-power vision sensors, implementing advanced features like in-hardware motion detection, is crucial for computer vision in the IoT domain. Unfortunately, to be highly energy-efficient, these sensors might worsen the perception performance (e.g., resolution, frame rate, color). Therefore, domain-specific pipelines are usually delivered in order to exploit the full potential of these cameras. This paper presents the development, analysis, and embedded implementation of a realtime detection, classification and tracking pipeline able to exploit the full potential of background filtering Smart Vision Sensors (SVS). The power consumption obtained for the inference - which requires 8ms - is 7.5 mW.

Files

2102.01340.pdf

Files (467.8 kB)

Name Size Download all
md5:5ea6356e8b22c56d6919899241a7196f
467.8 kB Preview Download

Additional details

Related works

Is published in
Conference paper: 10.48550/arXiv.2102.01340 (DOI)

Funding

MARVEL – Multimodal Extreme Scale Data Analytics for Smart Cities Environments 957337
European Commission