BEHAVIOUR RECOGNITION METHOD BASED ON HIDDEN STRUCTURE REASONING

摘要

Disclosed is a behaviour recognition method based on hidden structure reasoning. The method comprises the following steps: extracting local features of training data and clustering same so as to construct a codebook word bag, establishing a symbiotic relationship graph among codebooks by using the spatial and temporal distribution of the local features, reasoning a hidden structure in the symbiotic relationship graph in combination with the distribution characteristics of the local features in a human body and feeding back same to the symbiotic relationship graph, so as to obtain a graph model having multilayer structure information; using the graph model to acquire stability coefficients of the local features in the training data and constructing a feature vector of the training data, and training a classifier model; and extracting local features in an unknown video, using the graph model to calculate the stability coefficients of the local features, constructing a feature vector of the video, and using a classifier to classify. According to the method, by studying structure features of a human body to learn structural information among local features, the behaviour of a main body can be better described. The method can be used in the field of public security, such as abnormal behaviour detection and so on, and also can be used for man-machine interaction.