A Quaternion-Based Motion Tracking and Gesture Recognition System Using Wireless Inertial Sensors

Abstract

This work examines the development of a unified motion tracking and gesture recognition system that functions through worn inertial sensors. The system is comprised of a total of ten wireless sensors and uses their quaternion output to map the player's motions to an onscreen character in real-time. To demonstrate the capabilities of the system, a simple virtual reality game was created. A hierarchical skeletal model was implemented that allows players to navigate the virtual world without the need of a handheld controller. In addition to motion tracking, the system was also tested for its potential for gesture recognition. A sensor on the right forearm was used to test six different gestures, each with 500 training samples. Despite the widespread use of Hidden Markov Models for recognition, our modified Markov Chain algorithm obtained higher average accuracies at 95%, as well as faster computation times. This makes it an ideal candidate for use in real time applications. Combining motion tracking and dynamic gesture recognition into a single unified system is unique in the literature and comes at a time when virtual reality and wearable computing are emerging in the marketplace.iii

Quaternion-Based Human Gesture Recognition System Using Multiple Body-Worn Intertial Sensors

“…The previous research on this topic validated the accuracy of this type of gesture classification system using a single sensor as a baseline.It used Hidden Markov Model and Modified Markov Chain [1] [7] for classification. Following up with that, we are posing two main research questions.…”

“…Each sensor outputs its current orientation in the form of a quaternion, about which we will discuss in the next section. Upon startup, the sensor generates its own global reference frame such that the positive z-axis is oriented vertically upwards and the x and y axes are determined by the device's orientation (figure 5) [7]. After that, the MPU on each sensor takes readings from the accelerometer and the gyroscope and calculates and estimate of the current orientation of the sensor relative to those initial axes.…”

“…As mentioned in section 1.4, we used six gestures: Jab, Uppercut, Throw, Lift, Block and Sway ( figure 6). We kept four gestures from the previous study [7] and changed two -Jets and Asgard to Lift and Sway respectively. The reason behind this change was the difference of context between this study and the previous one.…”

“…The reason behind this change was the difference of context between this study and the previous one. While our previous study selected gestures within the context of its robot-based action game [7], the rationale behind our selection was to use gestures that would be simpler for anyone to understand and perform. On top of this, we also wanted to have a common set of gestures with the previous study for a meaningful comparison of results.…”

“…The sensors themselves also suffer from a drift issue which was covered in detail in the previous study [7]. Sensor drift causes the values of a continuous stream of data to deviate from their initial value even if the sensor were kept still and untouched.…”