UC Santa Cruz

Wayfinding systems using inertial data recorded from a smartphone carried by the walker have great potential for increasing mobility independence of blind pedestrians. Pedestrian dead-reckoning (PDR) algorithms for localization require estimation of the step length of the walker. Prior work has shown that step length can be reliably predicted by processing the inertial data recorded by the smartphone with a simple machine learning algorithm. However, this prior work only considered sighted walkers, whose gait may be different from that of blind walkers using a long cane or a dog guide. In this work, we show that a step length estimation network trained on data from sighted walkers performs poorly when tested on blind walkers, and that retraining with data from blind walkers can dramatically increase the accuracy of step length prediction.