3D Continuous Hand Gesture Recognition Research

Hand gesture recognition has drawn much attention in 3D human computer interaction thanks to the emerging techniques of commodity 3D sensors. Traditional Begin-End dynamic time warping(Begin-End DTW) approach for gesture recognition could detect the beginning and end of a speci?c gesture from an in?nite trajectory gesture sequence but provide multiple different gesture labels for one trajectory segment. This paper presents a Windowed Dynamic Time Warping(WDTW) approach for 3D continuous hand trajectory gesture recognition. The main contribution is as follow.By proposing WDTW approach, we introduce a parameterized searching window in the cost matrix of traditional DTW approach to detect the beginning and end of the speci?c template series curve from an in?nite trajectory gesture sequence. By doing so, we formulate continuous time series curve recognition into online parameter estimation of the searching window. Furthermore, we use a penalty item in the cost matrix to prevent the searching window from falling into the wrong path. Thus, the proposed WDTW approach signi?cantly improves the robustness of traditional Begin-End DTW approaches for noises.Firstly, we apply WDTW to user-defined continuous hand gesture recognition. WDTW can handle partial similar gestures. Even though we detect a multi-stroke gesture in which each stroke corresponds to a single-stroke gesture in worse case, we still can recognize this complex gesture not outputting multiple gesture labels for one trajectory segment. We evaluate WDTW and Begin-End DTW in continuous hand trajectory gesture recognition research in our two gesture dataset. WDTW gets 85% and 58.8% accuracy and gets 12.2% and 39.7% error rate respectively. While Begin-End DTW gets 86.7% and 57.6% accuracy and gets 45% and 139.8% error rate respectively.Secondly, we apply WDTW to letter recognition. In letter recognition, we perform an online research and get 75% mean accuracy and 15% mean error rate respectively. While Begin-End DTW gets 80% mean accuracy and 70.4% mean error rate respectively. The experimental results show that the proposed WDTW can signi?cantly improve the Begin-End gesture recognition performance.