Simultaneous Real-Time Human Fall Detection and Reidentification Based on Multisensors Data
Matteo Bastico, Verónica Ruiz Bejerano, Alberto Belmonte-Hernández · 2022
Fall detection and reidentification are active areas of research with a wide variety of applications in many fields. Nowadays, wearable devices capable of recording multisensors measurements are being increasingly introduced in people’s daily lives, such as smartphones and smartwatches. Often, these devices are used in healthcare rehabilitation scenarios to monitor patients activities and, eventually, detect important events like falls. In some cases, the equipped Inertial Measurements Unit (IMU) lacks of gyroscope and magnetometer, allowing to acquire only accelerations measures. In this conditions, the detection tasks become more complex and many false detections can arise. Additionally, other methods, such as monitoring cameras, are usually used in the environment to overcome this problem and further track the patients. But, due to the lack of personal identifiers in the camera tracking system, it is not straightforward to associate falls, when detected, to a particular person. In this paper, we propose a complete real-time system to detect falls reducing false positives and simultaneously reidentify the patients using 3D skeleton points, given by a tracking camera, and 3-axis accelerometer data. Firstly, the system performs fall detection by means of a mathematical analysis of the skeleton points evolution along different frames. At the same time, fall detection is computed on the acceleration data using a Finite State Machine approach. If a fall is detected with both mechanisms, reidentification is carried out to associate the skeleton with the wearable device. A dataset of fall sequences has been recorded and is available for testing purposes. The final accuracy of our fall detection and reidentification algorithm is 100% on our dataset.