Improved Gesture Trajectory Reconstruction from Acceleration Values for a Device Independent Gesture Recognition Approach

With the persistent spreading of computer technology in our everyday life and the resulting interaction between human and computer in many different situations, hand gesture interaction is one of the many interaction techniques becoming more and more important. Several different gesture devices like cameras, devices with integrated accelerometers or computer mice can be used for hand gesture interaction. Each device has different properties (like intrusiveness or handling) which influence its usability for different applications and situations. However, most applications providing gesture-based interaction support only one gesture device which restrict users especially if the application should be used in different situations. One reason is that developers have to pursue a higher implementation effort to support different gesture devices for the interaction process as each gesture device requires the utilization of a different gesture recognition algorithm. Furthermore, the effort of the user is also increased as every device and every recognition algorithm has to be trained individually. The most recent research solves the issue of device independence only for a limited group of gesture devices such as data gloves [3] or 3-axis accelerometers [4]. The approach presented in [5] goes one step further and accepts gesture input from cameras, accelerometers, and pen devices. It thus supports device independent hand gesture recognition for the most known gesture device types. The approach extracts user’s gesture trajectory performed in space by applying different algorithms for different types of gesture data from different devices (i.e. images, acceleration or position data). The extracted trajectory is a device independent data representation of the performed gesture. The reconstructed trajectories from different gesture devices are very similar which allows the application of a single training set for different gesture devices. In the next step, the reconstructed trajectories are resampled to a certain number of points and transformed into the Kendall shape space [2], a scaleand position-free representation of the resampled trajectory in the complex space. Finally, the shape representation is compared to each shape template stored in a database by a rotation invariant distance measurement. Further details of the approach and results of extensive experiments are 125