Date of Award
Doctor of Philosophy (PhD)
In this research we develop the next generation methods for inertial navigation, which seek to estimate trajectories of natural human motions with a smartphone that equips a low-cost Inertial Measurement Unit (IMU) sensor. A robust inertial navigation method has been a dream for academic researchers and industrial engineers for its ideal properties, e.g. low- energy consumption and high flexibility. However, the double integration from accelerations to translations are extremely vulnerable to even a tiny amount of sensor biases.
We utilize the recent advance in Machine Learning algorithms and develop data-driven methods for robust inertial navigation in the wild. Our key insight is that human motions consist of a few major modes, which can be leveraged to constraint the double integration. We create a large-scale high quality dataset with our novel two-device data collection protocol and train various machine learning models to capture inherent human motions. Extensive evaluations are performed and demonstrate that proposed methods are able to estimate accurate trajectories under natural motions in the wild, e.g. checking messages while walking forward, answering a phone while stepping sideward or sitting in a sofa while browsing the web with the phone.
The practical implications of proposed methods are also profound. We show an example of such applications, where motion trajectories from our inertial navigation algorithms are used to automatically construct WiFi radio maps, which are essential for providing an indoor positioning service. Other applications include AR/VR, fitness and health monitoring.
Yasutaka Furukawa Tao Ju
Tao Ju, Ayan Chakrabarti, Yasutaka Furukawa, Roman Garnett,
Available for download on Monday, May 15, 2119