An Examination of Learning Data for Behavior Identification and Range Expansion Using Doppler Sensor

In recent years, the rate of older people has been increasing with the aging of population. In such a situation, many elderly people suffer injuries from falls every year. In the case of living alone, it is difficult to find accidents by others, and there is a possibility of sequelae or death due to the delay in discovery. It has been reported that this is the case. In view of this background, we think that it is necessary to quickly identify dangerous situations such as falls and stumbling and daily life situations of elderly people living alone. In this paper, we use Doppler sensors to detect falls in this study. We perform spectral analysis using short-time Fourier transform on the signals obtained from the sensors, and construct a classifier that identifies falling actions by CNN. In addition, we focus on the overall activity recognition in the living space, and aim to expand the practical range of the proposed system, and examine the measurement range of a single sensor and the learning data. Even when installed on the wall (for horizontal detection) or on the ceiling (for vertical detection), dangerous situations such as “Fall” and ”Stumble” were 100% identifiable. In addition, it was revealed that the identification of ”Down” and ”Up” was improved by about 13.3% when installed on the ceiling. Furthermore, by learning the motion data from directly below at 1m, 2m, 3m, and 4m, “Fall” and ”Stumble” within ±4m in the horizontal direction and ±2m in the vertical direction were identifiable with an accuracy of more than 83.3%.This system is considered to be useful for detecting dangerous situations of elderly people living alone, and may improve their safety.