A Sensor Data Fusion Algorithm for Human Motion Estimation

Abstract

In the elderly population, falls are one of the major causes of injuries. Fall detection algorithms for wearable devices in the literature were found to focus on differentiating Activities of Daily Living (ADL) from falls, rather than on early fall detection and prevention before impact. This thesis work was aimed at providing accurate estimates of human motion parameters like translational velocities of the center of mass, so as to aid early fall detection algorithms in the future. An algorithm which fuses visual and inertial data obtained from a sensor setup consisting of a pair of cameras and an Inertial Measurement Unit (IMU) was developed. An Extended Kalman Filter (EKF) framework was used for sensor data fusion. A neural network was trained to map motions of interesting points or features obtained from image processing, to actual motion of the sensor setup attached to the hip of a person. This neural network provided the correction term to the EKF in the form of 3 dimensional(3D) translational velocities which is an important motion parameter for the detection of falls. First, this algorithm was trained and tested for hand-held sensor setup motions. Acceptable velocity tracking was observed for slow motions. Then the algorithm was trained and tested on motions of a test subject. The accuracy of estimation of 3D translation velocities and in particular the vertical component of the 3D velocity was studied for two distinct sets of activities or motions namely the walking motion and the sitting down/standing up motion. It is shown that the combination of the EKF and the neural network is capable of reacting and tracking the velocities for the sitting down/standing up motion.