Research On Human Interaction Recognition Based On Space-time Feature

Behavior recognition, especially recognition of the interactive behavior has an importantsignificance in visual surveillance. In recent years, the researchers are actively seeking new orimproved ways to use technology to deter and respond to accidents, crime, terrorism, and soon. The technology of behavior recognition can be used proactively for prevention ofincidents or reactively for investigation after the fact. We focus on the recognition of theinteraction between two persons. This paper mainly describes the algorithms about theextraction of the feature points and the algorithms about the technology of behavior modeling.And it also includes the evaluation of these algorithms in this paper. For the extraction offeature points, we mainly describe the static characteristics, the dynamic characteristics andthe spatial-temporal characteristics. And for the approach of behavior modeling, this articlehas introduced the template matching model, the probabilistic graphical model, the modelbased on the grammar and the model based on the statistical learning algorithm.In this paper, it proposes a three-level behavior recognition system, and the three levelsare the extraction of the behavioral characteristics, the classification of the single peopleaction and the recognition of the interactive behavior.For the extraction of the behavioral characteristics, this paper proposes two algorithms ofthe spatial-temporal feature points extraction. One is the feature points detection based on the3D Harris corner points. And the other is the detection of spatial-temporal feature points basedon the MIC corner and1D Gabor filter. Then the article describes the classification algorithmof the spatial-temporal feature points.For the classification of the single people action, the paper proposed a behavior modelingalgorithm using the Restrictive Boltzmann Machine. The input of the algorithm is thespatial-temporal feature and the output is the classification of the single person action. TheRBM network is a probabilistic graphical model. The Restrictive Boltzmann Machine iscomposed of the visible units and the hidden units. The RBM convergence is faster than theBM and has the same classification ability with the BM.For the recognition of the interactive behavior, the article proposed a model based onMarkov network and First-order logic. And the input of the model is the describes of the single person actions. The output is the result of the interactive behavior recognition.In this paper, we have designed a series of experiments. The training data include thevideo of boxing, handshaking, kicking, embracing and pushing. And the result of experienceshows that the algorithm we proposed is effective in the training videos.