Techniques making human-machine interaction easier and more natural are increasingly sought in today’s modern society. Hand gestures are easy to interpret, several shapes can be drawn by the user. Gesture recognition solutions are already available on the market, but these all share the disadvantage of limited mobility.
To create a simple, ubiquitous and unobtrusive device, the recognition process should run locally on the wearable sensor module. Wearable devices are mostly collecting and sending sensor data via a wireless protocol and these are unable to run gesture recognition algorithms because of their large computational requirements. Typical sensor modules are used for precise movement reconstruction with acceleration, angular rate and magnetic sensors. Instead of movement reconstruction, gesture recognition methods are searching for regularities in the sensor signals that can be linked to some gestures. My aim is to design a device achieving real-time processing and gesture recognition after the data collection.
With the goal of compressing sensor data information, a high efficient peak detection method will be presented. Analysis of gestures, based on simple geometric forms (straight, slant, curve lines) will be discussed in detail. This new approach has an ability for non-user specific gesture recognition. Implementations of the algorithm have been made, first with a microcontroller system for quick prototyping then in a regular watch-sized device. Efficiency and multi-level testing will be demonstrated. Appearing problems, solutions and further development options which could make industrial usage available will also be presented.