[4D3-E-2] Machine learning: living environment

[4D3-E-2-01] Privacy-Preserving Resident Monitoring System with Ultra Low-Resolution Imaging and the Examination of Its Ease of Installation

〇Takumi Kimura¹, Shogo Murakami¹, Ikuko Egushi Yairi¹ (1. Sophia University)

Keywords:Low resolution sensor, Human behavior recognition, Resident monitoring

Monitoring systems using infrared array sensors allow monitoring of residents while protecting their privacy. However, since such a sensor is vulnerable to subtle movements, accuracy of posture classification is low, and limits the locations and methods available for installation. This study proposes a posture classification method with higher accuracy. Over 93% accuracy was achieved in posture classification by RGB conversion of infrared array sensor images and successfully decreased loss due to displacement by DCNN. Additionally, this research considers methods to create artificially simulated data for postural-behavioral study. To check the validity of this method, postures of 3 subjects were examined using a classifier with studied simulation data. Finally, simulation environments with different sensor altitudes and angles were created to examine the ease of installation for the proposed method. As a result, the experiments showed that accuracy was highest at approximately 90% when the sensor was located 50cm below the height of the target and when the tilt angle was within ±2°.