Gesture Recognition Research Based On Depth Image

Gesture recognition is more and more important in many fields such as sign language recognition, video game and virtual control. Traditional recognition methods based on data glove and color image have their limits. The Data glove needs much adjustment and limit hands’natural movement. Besides, the data glove is very expensive. The color image make it difficult to segment hand out. It will be more difficult in scene where overlapping, light changing, quick movement and objects whose color is similar to hand’s occur.Depth cameras like Kinect offer new way for gesture recognition.In the depth image, pixel value is the distance from object to camera. Depth image makes it easy to segment hand out, and gives 2.5D geometry information. Kinect can not be affected by light, and can even work in dark environment, we design algorithms for static gesture recognition and dynamic recognition. Further, we use depth learning framework to find out underlying features of hand and use these features for gesture recognition.For static gesture recognition, we design a recognition algorithm based on hand dominant line. Our algorithm keeps scale and rotation invariant. We propose the nota-tion of hand dominant line(HDL) and its computational method. The HDL achieves ro-tation invariant. We evaluate our algorithm on two public available dataset NTU dataset and sASL dataset. In cross validation wherehalf samples are used for training and another samples are used for testing, we achieve 97.1% and 96.2% recogniton rate, re-spectively. For dynamic gesture recognition, we design a recognition algorithm based on UV feature and random forest. Our algorithm skillfully avoid the problem of video series alignment and compression. UV features can grasp the relative depth informa-tion between the feature point and the points around the feature points. When there is only labeled dynamic gesture per class, we achieve 85% recognition rate on ChaLearn dataset. Deep learning framework can automatically find out features. We use DL to learn features from the depth hand images. We choose 8*8 patches as training dataset. we choose softmax classifier as classifier. On NTU dataset, the method achieves 93% recognition rate.