Stable Fall Detection Using Head Position Data with Obrid-Sensor
Taito Kumagai, Tomohiro Hayata, Shunki Tanaka, Atsushi Matsubara, Seiji Nishifuji, Shota Nakashima · 2024
The aging society is a contemporary societal issue, creating a high demand for monitoring systems in elderly care settings. Cameras are widely used as monitoring sensors, but their installation in places like bathrooms and toilets raises privacy concerns. Wi-Fi sensing as an alternative system has been proposed, however it requires a lot of prior data on the target persons' state. Our laboratory has proposed an anomaly detection system using the Obrid-Sensor to address this issue. This sensor consists of a cylindrical lens and a line sensor and compresses the brightness information of the space into one dimension. Using this one-dimensional information enables privacy-aware monitoring. Previous research in fall detection proposed an anomaly detection method for a room in a nursing home. The sensor was positioned at a high place and detected differences in waveform shapes between standing and falling states of the target person. Detection becomes difficult when the target person falls forward or backward relative to the sensor, as the waveform shapes are similar. This study proposes a detection method independent of the fall direction aiming to improve detection accuracy in a real environment. The target person's movement transitions were used for detection in order to deal with cases where the waveform shapes were similar. We defined the position of the target person's head as a feature value representing transitions because it changes the most when the target person falls. The LSTM learned only the feature values of the target person in standing and walking states to create a model. Measured values were input into the created model to output predicted values, and the squared error between the measured and predicted values was calculated. We set a threshold for this to detect the target person's fall. Validation experiments showed a true positive rate of 98.1%, indicating how few missed detections of the falling state occurred. The true positive rate for each fall direction achieved 100% even in directions that were difficult in previous research, confirming improved detection accuracy in real environments.