Video Gesture Tracking And Handwritten Numeral Recognition

In recent years, with the continuous improvement of computing capacity, target tracking and recognition technology develop rapidly, and get the wide attention of many researchers. Now the interaction research of human and machine is in full swing, gesture is an important way of communication, to realize interaction of human-computer more humanized, we must do some in-depth study of gestures, so gesture tracking and gesture recognition have get the wide attention. For gestures using in daily life is complicated, it also makes the research work very challenging.In this paper, tracking and recognizing gestures which changing constantly in the real-time video, the main work and research results are as follows:1. Due to the hand skin is stable relative to the changing gestures, so this article uses the color as the feature to track gestures in real-time video, because the RGB color space is not suitable for detection of skin color, replacing it with the HSV color space in this paper.2. This paper detailed introduces the Mean-Shift algorithm, and aims at Mean-Shift algorithm initial target must be selected manually, proposing the maximum connected domain for the initial gestures selected, so as to realize the initial gesture selecting automatically.3. Due to Mean-Shift algorithm lack of updating mechanism for tracking window, it can’t realize tracking window rotating and scaling along with the gestures, and sometimes because the tracking target changing so quickly that lead to gesture tracking failure. According to the above problem, this paper uses the method of image geometric moment to realize the tracking window changing self-adapting along with the target, and the simulation experiment results show that this method has good performance.4. Because the taken samples and the tracking gesture in the real-time video may have many similar skin color area or miscellaneous class that has nothing to do with the skin color, to remove these miscellaneous class when we in image processing, given that the extracted contour of gestures accounted for most of the area in general, firstly find the largest contour area, so that it can remove most noise, then can extracting relatively accurate feature of the gesture found on above. Owing to different gesture area often have different ratio, and it can be used as a feature. At the same time, using the invariant moment and Fourier descriptor as to the characteristic of gestures.5. In this paper, we using the tracking gestures to control its trajectory, and the gesture trajectory collected by camera in the air writing numbers, etc. just like writing on the paper using the pen. Thus, at first, extracting the stroke number of handwritten numbers and the Fourier descriptor as features for recognizing the handwritten numbers on-line in this paper. And then using the random forest algorithm to train the extracted characteristics, and finally using the learning outcome to recognize gestures and handwritten numbers. The experiments prove that the same idea also applies to handwritten letters.6. Experiments prove that the image preprocessing method, an improved tracking method, extracted features for gesture tracking and identification as well as classification and recognition method used in this paper can realize the accurate tracking and recognition for dynamic hand gestures in real time video; and it also can recognize the on-line handwritten digital quickly and accurately.