Trajectory-based Gesture Recognition Research And Application

In recent years, with the rapid development of computers, humancomputer interaction or HCI, has become an increasingly important part of ourdaily lives. As we know, the most popular mode of HCI is based on simplemechanical device, e.g. keyboard, mice and touch screen. However, thesedevices inherently limit the speed and naturalness with which we can interactwith the computer. With dexterous functionality in communication andmanipulation, human hand has been regarded as the most effective,general-purpose interaction tool for HCI. As a result, the gesture recognitiontechnique becomes one of the most effective and most natural interactiontechnologies.For the self-complexity of hand gesture and obstacles from otherconditions, it is very difficult to realize a real-time, simple, rapid, and robustrecognition system. For instance, the change of the illumination conditionhinders greatly the hand localization. As a result, it is difficult to realize robusthand segmentation in real-time dynamic hand gesture recognition system.Furthermore, the selection of dynamic hand model is also significant forrecognition. For improving the recognition, in this paper, a gesture recognitionmethod based on hand trajectories is proposed in the view of optics and kinematics. First, a luminosity removing method is used on hand localizationstep for eliminating the influence of illumination. After hand localization,hand motion trajectory information is extracted for representing dynamicgesture, while a velocity filter is applied for consecutive hand gesturesegmentation. A standard vector database is then built by combining all thedefined gesture standard vector data. When an unknown gesture inputs, itsdistances to the defined gestures in standard database are calculated and usedto decide which types it belongs to. Experimental results show an averagerecognition rate of89.5%for10defined dynamic gestures. Additionally, anapplication system is also constructed by combining dynamic hand gesturerecognition and static hand gesture recognition in this paper. In the real-timetesting step, we verified the efficiency of our system.The innovations of our work include at least three points as follow：First,a luminosity removing method is used to remove the influence of illuminationcondition. Second, a normalization and interpretation method is employed toeliminate the differences in temporal and spatial dimension between differenthand gestures. Finally, a two-level classification method is proposed formaking different criterions for defined gestures. The testing result showsimprovement on recognition.