Download files
Abstract
Considerable research has been performed exploring Activity Recognition (AR) using wearable sensor nodes such as smart phones that incorporate accelerometers, gyroscopes and magnetometers.
This thesis presents an empirical evaluation of AR using on-body sensors. It studies the recognition of 22 Activities of Daily Living using either wearable sensors that have fixed locations on the subject’s body or smart phones carried by the subject. The differences between the two sensor configurations are explored as well as parameters for computationally efficient and accurate AR.
Initially, fundamental classifier settings that impact AR accuracy are explored. These include evaluating the performance of acceleration, rotational velocity and orientation derived features (the three are referred to as sources). In addition, minimum sampling frequencies, window size, window overlap and sensor locations on the body are also explored.
Next, two factors that differentiate AR using wearable sensors from using a mobile phone are studied: the possibility of a mobile phone being carried (1) in an unknown and previously untrained location on the subject’s body, and (2) with any orientation.
Key findings presented within the thesis include: (1) Acceleration derived features perform better than features of the other two sources. (2) Using acceleration features alone performs only marginally worse than using features derived from all three sources. (3) Orientation derived features have the second highest success-rate but require the lowest sampling frequency of features derived from the three sources. (4) Accuracy is affected by location and number of sensors. The wrists yield the highest overall performance for the activities studied. A depreciating returns relationship exists between accuracy and number of sensors used. (5) Of the body locations and activities studied, it is possible to identify the location on which the sensor is carried without knowing the subject’s activity. (6) Reorienting the data from local to global coordinates ameliorates the decrease in success-rate when the sensor’s orientation is transient. However, it incurs a marginal decrease when the orientation is fixed. (7) Reorienting the data using orientations obtained from an IMU results in higher success-rates than using accelerations. (8) Highly confusable activities are found to have similar gross motor movements.