Ambient AI (A2I)
Infrastructure & Architecture

Ambient AI (A2I) started from the emergence of DNNs accelerating chips, which make AI everywhere. The number of Ambient AI devices are exploding; thus, their management issues arise.

The A2I Infrastructure includes an A2I Middleware that manages the AI Models Metadata and handles dynamic reconfiguration of edge AI network for collaborative inferences of large-scale distributed devices.

A2I devices like Google Coral Board and Nvidia Jetson Nano capture rich behavioral data from dynamically reconfigurable audio-visual sensors and enable local actions even when disconnected from the network.


Related Projects
Deep Learning Engineering Life Cycle Management
Multi-Agent System Problems on Edge AI Network


Specification

A2I Infrastructure

Ambient AI (A2I) started from the emergence of Deep Neural Networks accelerating chips, which make AI everywhere. The number of A2I devices are exploding; thus, their management issues arise.

The A2I Infrastructure includes an A2I Middleware that manages the AI Models Metadata and handles dynamic reconfiguration of edge AI network for collaborative inferences of distributed devices.

Our A2I Infrastructure also integrates with popular Deep Learning frameworks such as Tensorflow, Pytorch, Google Cloud, Amazon Cloud.

A2I Architecture

We propose an A2I architecture that helps manage and build real-world applications based on edge AI devices. It includes Four layers:

  • The Storage and Computing Infrastructure Layer: keeps all information and integrates with Deep Learning computing platforms such as Tensorflow, Pytorch, Google Cloud.

  • The Ambient AI Life-Cycle Management Layer: manages the Ambient AI life-cycle, from the data, models, devices to applications, pipelines, and versions management.

  • The Ambient AI Services Layer: provides useful services for building Ambient AI applications such as model searching, pipeline suggestion, pipeline scaling, and federated learning for device-locally model training.

  • Ambient AI Applications Layer: supports real-world applications in many fields such as Health-care, Smart Factory, Smart Cities, and Sports & Entertainment.


Our Testing Ambient AI Devices

Google Coral Board with Camera

Nvidia Jetson Nano Device

Our Current Results and Demos

Face-Masked Detection

Two Step Mask Detection using Google Coral:

  • Step 1: Face Detection -> captures all faces

  • Step 2: Mask Classification -> mask or no-mask

Sufficient for the real-time inference

2D Human Pose Estimation

2D Pose Estimation on Google Coral:

  • Use a PoseNet model from Google that can detect 17 pose key-points of a person in an image.

  • Run in real-time with a live Camera.

3D Human Pose Estimation

Real-time 3D Human Pose Estimation on Google Coral:

  • Step 1: Use a 2D Human Pose Estimation model to predict 2D body key-points.

  • Step 2: Learn 3D poses from the estimated 2D joints.