Fall Detection Using Time-Series Data with Obrid-Sensor

Abstract

The elderly people dying alone have become a significant concern. Monitoring systems using cameras have been widely used conventionally, although it has the possibility of invading privacy. Therefore, a system based on WiFi sensing has recently been proposed, but data on various states of the target is necessary for fall detection. We proposed a fall detection system in an ideal environment using an Obrid-Sensor. This sensor consists of a cylindrical lens and a line sensor. The sensor compresses the brightness feature of the subject space into one dimension for capturing, thus enabling privacy-conscious monitoring. In this study, we proposed positioning the sensor at a high place to detect falling in practical environments. A waveform approximating a Gaussian function was created from the brightness distribution waveform obtained by the background subtraction method. The difference between the approximate waveform and the measured waveform was calculated by Pearson's correlation coefficient and used as the degree of approximation. LSTM predicts the degree of approximation after a time interval from the current time, and the error from the actual measurement is used for fall detection. The results of the validation experiment showed that the true positive rate was 90.9%. Furthermore, we calculated the true positive rate for each fall position. It was 100% at specific fall positions. These results confirm the effectiveness of highly accurate fall detection from a small feature.