Thermal Array Sensor Resolution-Aware Activity Recognition using Convolutional Neural Network
Goodness Oluchi Anyanwu, Cosmas Ifeanyi Nwakanma, Adinda Riztia Putri, Jae‐Min Lee, Dong‐Seong Kim, JeongHan Kim, Gihwan Hwang · 2022
Human activity detection and classification (HADC) has become a growing research issue due to the development of sensor technologies, deep learning models, and the need for the safety of people in smart spaces such as buildings and factories. Various researchers have employed sensors with different resolutions for HADC. However, the impact of sensor resolution on the sensor data quality and the accuracy of the deep learning algorithm in this field has been little discussed. In this work, the impact of three different thermal sensor resolutions was investigated while proposing a convolutional neural network (CNN). The results showed that the proposed CNN displayed a resolution-aware performance, being able to contain the impact of a change in thermal sensor resolutions. Although the CNN model had lower accuracy as the resolution was changed from 32x32 to 4x4H and 4x4, respectively. However, it was able to reduce the type errors and maintain an average accuracy of 82.74%.