Dynamic Gesture Recognition Based On Visual Research

The gesture recognition is a priority of the new generation of human-computer interaction research. In paper, we mainly focus on the recognition system based on the apparent of dynamic gestures under a monocular vision. We will talk about the studies of all aspects of gesture recognition. And finally, we established an available dynamic gesture recognition system.In the first place, we introduce the framework of the gesture recognition system. Then in order of that image sequences are processed, we discuss technologies used by each module, compare their performances and intro some improvements, mainly include:(1) we analyze characteristics and the nature of dynamic gestures as a means of human communication, and construct a dynamic gestures vocabulary for a monocular visual recognition. This vocabulary would also be useful material for future researchers.(2) Segment hands with motion detection and skin-color detection. We tried a new threshold updated method based on genetic algorithm for skin-color detection. Integrating motion detection and color detection, an integrative method based on fuzzy mathematical theory was put forward. The new method get good results.(3) We prove that shape and trajectory are adaptive for gestures classification, and introduce several widely used methods of extracting these features. Then we propose a way omitting time of extracting trajectory features and improving the timeliness of features extraction, which gain trajectory from the process of computing moments.(4) We regularize parameters of the input interface of a SVM classifier, so that it will be suit to quantify invariant moments. Then we set up a SVM-HMM2-layer classification system. In the first layer, the SVM classifier quantize high dimensional feature vectors to some values that are suitable for input of HMM. The final results can be generated by connected the two classifiers in series.