1 of 19

AeroSense: Sensing Aerosol Emissions from Indoor Human Activities

Bhawana Chhaglani, Camellia Zakaria, Richard Peltier, Jeremy Gummeson, Prashant Shenoy

​

UbiComp 2024, October 8, 2024

1

2 of 19

Aerosol Sensing: Motivation

  • Increasing risk of indoor airborne transmission during flu and pandemic.

​

  • Bacteria or viruses are transmitted through small respiratory droplets. 

​

2

​

​

Sensing aerosols concentrations is essential to determine risk of transmission.�

​

3 of 19

Aerosol Sensing: Challenges

3

​

We develop a system to predict aerosol emissions that is accurate, low-cost and easy to deploy.

​

Low Risk

Aerosols:

5 p/cc

  • Existing works use CO2 and occupancy sensors, which are not accurate indicators of aerosols.
  • Aerosol sensor is bulky, expensive, complex for everyday use.

4 of 19

Proposed Approach

  • Instead of directly sensing aerosols, we sense activities that contribute to aerosol emissions by using audio sensing.

4

Aerosols

Aerosols

Activity

(Speaking,

Coughing, etc.)

5 of 19

Audio Sensing for Aerosols

5

Aerosol and Activity Type

Aerosol and Activity Level

  • Aerosol emissions depend on the type and level of human respiratory activity.

6 of 19

AeroSense Design

6

1. Activity Type Recognition

2. Activity Level Detection

3. Aerosol Prediction

7 of 19

1. Activity Recognition

  • We extract non-reconstructible features from raw audio to ensure user privacy
    • Time domain, spectral, linguistic, and handcrafted features

​

7

Mask reduce aerosol emission by 60%

Trained classifiers on open source datasets (AudioSet, MASC) and our data

8 of 19

2. Activity Level Detection

  • To detect the activity intensity, we estimate the distance between the microphone and activity source.

​

8

Received dB level

dB level at the source = ??

Average error = 7.74%

9 of 19

3. Aerosol Prediction

  • We combine activity type, activity level, mask presence, active sources, and voice liveness to predict aerosol.
  • We train a regression model to predict aerosol in cleanroom setup, followed by aerosol aggregation.

​

​

9

Low Risk

Aerosols:

5 p/cc

10 of 19

AeroSense Prototype

  • Our prototype consists of two microphone arrays connected to a Raspberry Pi.
  • We conduct user studies (~100s of hours) in real-world settings: office room, classroom, and conference room.

​

10

�Github Link: https://github.com/umassos/AeroSense

11 of 19

Results: Audio-based Aerosol Prediction

11

​

We achieve low MSE and high r2 of 2.34 and 0.73 respectively for predicting aerosol concentration on clean room dataset.

​

​

12 of 19

Results: Activity Type and Intensity

  • Activity detection models achieve over 86.33% recognition accuracy and 75% mask detection accuracy.
  • High fidelity in aerosol prediction for different speech loudness levels.

​

​

​

12

13 of 19

Conclusions and Future Work

​

​

13

Contact me: bchhaglani@cs.umass.edu

  • Augment smart home assistants with aerosol prediction capabilities.

​

  • Integrate AeroSense with BMS to modulate ventilation and provide alerts advising occupants.

​

  • The broader goal of this work is to understand the causality between aerosol and acoustic features.
  • Developed AeroSense, a novel audio-sensing approach to predict the rate of aerosol generated from human activities.

​

  • Demonstrated efficacy of AeroSense through controlled and in-the-wild experiments.

14 of 19

Supplementary Slides

14

15 of 19

Audio Features

15

16 of 19

Mask Detection

  • We observe the aerosol reduction by 60% due to mask in our experiments

​

​

​

​

  • Since face masks attenuate high-frequency speech, we design audio features to detect the presence of mask. �

17 of 19

Electronic Vs Human Speaker

  • Voice Liveness Detection: We use features like energy balance ratio to distinguish between human and electronic speaker

18 of 19

Evaluation

19 of 19

19