Human Motion Analysis Using Very Few Inertial Measurement Units

Abstract

Realistic character animation and human motion analysis have become major topics of research. In this doctoral research work, three different aspects of human motion analysis and synthesis have been explored. <br /> Firstly, on the level of better management of tens of gigabytes of publicly available human motion capture data sets, a relational database approach has been proposed. We show that organizing motion capture data in a relational database provides several benefits such as centralized access to major freely available mocap data sets, fast search and retrieval of data, annotations based retrieval of contents, entertaining data from non-mocap sensor modalities etc. Moreover, the same idea is also proposed for managing quadruped motion capture data.<br /> Secondly, a new method of full body human motion reconstruction using very sparse configuration of sensors is proposed. In this setup, two sensor are attached to the upper extremities and one sensor is attached to the lower trunk. The lower trunk sensor is used to estimate ground contacts, which are later used in the reconstruction process along with the low dimensional inputs from the sensors attached to the upper extremities. The reconstruction results of the proposed method have been compared with the reconstruction results of the existing approaches and it has been observed that the proposed method generates lower average reconstruction errors.<br /> Thirdly, in the field of human motion analysis, a novel method of estimation of human soft biometrics such as gender, height, and age from the inertial data of a simple human walk is proposed. The proposed method extracts several features from the time and frequency domains for each individual step. A random forest classifier is fed with the extracted features in order to estimate the soft biometrics of a human. The results of classification have shown that it is possible with a higher accuracy to estimate the gender, height, and age of a human from the inertial data of a single step of his/her walk.