A double-layer automatic orientation correction method for human activity recognition

Human activity monitoring systems using inertial sensors have found wide applications in the field of health and wellness by providing valuable information for diagnostics and rehabilitation processes to doctors and clinicians. As the scales of studies increase, sensor orientation placement errors have become one of the most commonly seen difficulties for such systems. Assuming patients to wear sensors at the correct orientation is unrealistic and will result in a large amount of data loss or distortion. In order to tackle this problem, we propose a double layer classification model. The first layer, not assuming correct sensor orientation, uses orientation-invariant accelerometer magnitude to construct a highly conservative walking detection model. The detected walking beacons from this layer are used to compare to the training template to obtain the true sensor orientation. Then proper rotation matrix can be applied to the whole day data, and fed into the second layer of a finer classifier where orientation-variant features are used. In order to show validity of this method, we hired 7 healthy subjects and 2 stroke patients in the rehab process to wear the sensors for two days and at least 6 hours each day. Ground truth are labeled manually with a Matlab GUI tool. Precision and recall for walking detection in each day are reported and discussed.