Research On Depth Information Based Dynamic Hand Gesture Recognition

With the development of human-computer interaction technology, dynamic hand gesture recognition was gradually accepted by users with a natural and convenient experience. It is showing broad prospects in the new multimedia technologies such as virtual reality and augmented reality. Vision based dynamic hand gesture recognition algorithms can reduce the complexity of system, which is becoming the main trend of hand gesture recognition technology. The importing of the depth information can not only overcome the influence of environment illumination changing and background complexity in color image, but also can provide plentiful 3D information to the hand gesture features. However, dynamic hand gesture recognition algorithms based on depth sequences are not fully analysis the characteristics of hand motion, which leads to two defects. Firstly, the features extracted from hand gestures suffer temporal redundancy. Secondly, they have low robustness to the same hand gestures with different speed. This research is launched based on the above questions.First of all, a hand gesture decomposition algorithm based on depth sequence is proposed in this paper. The algorithm firstly extracts direction based hand shape features from depth images, which can describe edge direction information of hands. Then, the problem of hand gesture decomposition is solved using clustering method, and hand gesture decomposition algorithm in depth sequence based on spectral clustering algorithm is proposed. This algorithm computes weighted similarity matrix of adjacent frames to construct undirected graph model. To solve the problem of graph partition, the similarity matrix is further converted to graph Laplacian. It is meaningfully to simplify the graph optimization problem to a matrix solution problem. Then Laplacian matrix is solved, and eigenvector is selected for the binary iteration clustering which can overcome the problem of uncertain category number. The most representative frames are selected from each category to compose key node sets in the end.Secondly, dynamic hand gesture recognition algorithm based on hand gesture decomposition is proposed. And the effect of parameters in the algorithm is analyzed in detail on the common data set and our data set. What's more, the comparison of this algorithm with state-of-the-art algorithms shows that the proposed dynamic hand gesture recognition algorithm based on hand gesture decomposition can remove temporal redundancy information, and improve the robustness to different speed hand gestures. The proposed method significantly improved the recognition accuracy.Last but not least, depth information based dynamic hand gesture recognition algorithm is converted to key technology in stereo scene interaction. A virtual 3D meeting scene interaction system is realized by applying depth information based dynamic hand gesture recognition algorithm. It is a transformation of the theoretical research to practical application. And it is also a perfect combination of dynamic hand gesture recognition and stereo display technology.