指势交互用户对象识别

he daily ability of human being can be fully exerted to get rid of bondage of current routine input devices if hand pointing gesture is selected as one of human computer interaction (HCI) models. A key problem is how to extract efficiently pointing object from clutter dynamic circumstance. A novel approach to recognizing the pointing object is developed based on spatial-temporal motion features in this paper. Multi-scale wavelet transform with some outstanding local characteristics both in temporal and spatial fields is adopted to extract foreground motion objects from a clutter scene, which overcomes some deficiencies in environment conditions, dynamic circumstance variations, prior hypotheses, and so on. According to the characteristics of the pointing behavior, some motion variation states are selected to count the accumulative frames in a certain direction. The action with pointing behavior is detected from the extracted foreground motion objects. Some stable and credible HOG vectors in the pointing hand are fast gotten based on gradient integral map with multiple scale variations. Machine learning method is used to classify and train the HOG vectors. The object with pointing gesture is extracted according to spatial relevance of the pointing hand to the object with pointing gesture. Experiment results have shown that the proposal is efficient and feasible.