Brandon is interested in using inertial sensing for state estimation. In particular, he is working on ways to utilize machine learning to acquire supplementary information from inertial sensors that cannot be easily modelled by other techniques.
Currently, he is studying methods to improve foot-mounted inertial tracking of first responders through learning. In emergency situations, real-time localization of first responders (such as firefighters) is imperative in order to ensure their safety. As Global Navigation Satellite Systems (e.g., GPS) provide unreliable signals indoors, alternative methods of localization are required.
We use foot-mounted inertial navigation as an alternative to GPS localization, estimating a user’s trajectory by measuring their linear acceleration and angular velocity over time. Since low-cost IMUs typically produce very noisy measurements, position error is unbounded. To reduce error growth, we detect when the foot is in midstance: the very short period in which the foot is flat and stationary relative to the ground. During midstance, the IMU velocity is known to be zero, and a non-zero velocity reading can be attributed to error. Using an extended Kalman filter, we fuse these zero-velocity measurements with our state estimate to significantly reduce error growth.
A drawback to this technique is that current midstance detection methods are not robust to varying motions. If a user begins to run, crawl, or climb stairs, standard step detection can fail. Brandon’s current work aims to achieve more accurate and robust midstance detection by leveraging learning methods.