Orientation Invariance Methods for Inertial Gait

Intelligent devices such as smart phones, smart watches, virtual reality (VR) headsets and personal exercise devices have become integral elements of accessories used by many people. The ability of these devices to verify or identify the user could be applied for enhanced security and user experience customization among other things. Almost all these devices have built-in inertial sensors such as accelerometer and gyroscope. These inertial sensors respond to the movements made by the user while performing day to day activities like walking, getting up and sitting down. The response depends on the activity being performed and thus can be used for activity recognition. The response also captures the user\u27s unique way of doing the activity and can be used as a behavioral biometric for verification or identification. The acceleration (accelerometer) and rate of rotation (gyroscope) are recorded in the device coordinate frame. But to determine the user\u27s motion, these need to be converted to a coordinate frame relative to the user. In most situations the orientation of the device relative to the user can neither be controlled nor determined reliably. The solution to this problem requires methods to remove the dependence on device orientation while comparing the signals collected at different times. In a vast of majority of research to date, the performance of authentication algorithms using inertial sensors have been evaluated on small datasets with few tens of subjects, collected under controlled placement of the sensors. Very often stand alone inertial sensors have been used to collect the data. Stand alone sensors afford better calibration, while the sensors built into smart devices offer little or no means of calibration. Due to these limitations of the datasets used, it is difficult to extend the results from these research to realistic performance with a large number subjects and natural placement of off-the-shelf smart devices. This dissertation describes the Kabsch algorithm which is used to achieve orientation invariance of the recorded inertial data, enabling better authentication independent of device orientation. It also presents the Vector Cross Product (VCP) method developed to achieve orientation invariance. Details of a realistic inertial dataset (USF-PDA dataset) collected with commercial smart phones placed in natural positions and orientations using 101 subjects are given. The data includes sessions from different days on a subset of 56 subjects. This would enable realistic evaluation of authentication algorithms. This dataset has been made publicly available. The performance of five methods that address the orientation dependence of signals are compared to a baseline that performs no compensation for orientation of the device. The methods as a part of a overall gait authentication algorithm are evaluated on the USF-PDA dataset mentioned above and another large dataset with more than 400 subjects. The results show that the orientation compensation methods are able to improve the authentication performance on data with uncontrolled orientation to be close to performance on data collected with controlled orientation. The Kabsch method shows the highest improvement