Capturing 'Cough Sounds' Inside Masks... Development of Cough Recognition Camera

[Asia Economy Reporter Junho Hwang] An AI camera that recognizes coughing sounds and accurately identifies the location of the person coughing has been developed. It is expected to enable the screening of suspected patients through coughing, one of the main symptoms of the novel coronavirus infection (COVID-19), alongside fever.

A research team led by Professor Yonghwa Park of the Department of Mechanical Engineering at the Korea Advanced Institute of Science and Technology (KAIST), in collaboration with SM Instrument, announced on the 3rd that they have developed a 'cough recognition camera' that recognizes coughing sounds in real time and visually displays the location of the person coughing.

Accurately Identifying the Location of the Person Coughing

The research team developed a cough recognition camera applying a deep learning-based cough recognition model that recognizes coughing sounds in real time. This camera recognizes coughing sounds and locates the person coughing. It also features cough count and real-time tracking functions.

The cough recognition model was created by repeatedly applying machine learning to a convolutional neural network, an artificial neural network used for analyzing visual images. It takes features of a 1-second audio signal as input and outputs a binary signal of 1 (cough) or 0 (other), with the learning rate set to decrease if it stagnates over a certain period to optimize learning.

For machine learning, publicly available voice datasets actively used for research by Google and YouTube were utilized. The AudioSet was used for training and evaluation datasets, while the rest was used for data augmentation to help the cough recognition model learn various background noises.

For data augmentation, background noise was mixed into the 'AudioSet' at ratios of 15% to 75%, and volume was adjusted between 0.25 to 1.0 times to adapt to various distances. The training and evaluation datasets were composed by splitting the augmented dataset at a 9:1 ratio, and a separate test dataset was recorded in an office environment.

Test Accuracy of 87.4%

As a result of utilizing these data, the research team achieved a test accuracy of 87.4% and projected that accuracy would further improve with learning in real-time usage environments in the future.

The research team expects the cough recognition camera to be used as medical equipment capable of detecting epidemics in crowded public places or continuously monitoring patients’ conditions in hospitals.

Professor Yonghwa Park said, "In the ongoing spread of COVID-19, using the cough recognition camera in public places and densely populated facilities will greatly aid in epidemic prevention and early detection." He added, "Especially when applied in hospital rooms, it can record patients’ conditions 24 hours a day for treatment purposes, reducing the workload of medical staff and enabling more precise monitoring of patient status."