Research On Unsupervised Activity Recognition Based On Spatio-Temporal Interest Points

Human activity recognition is already one of the important research areas in the domainof computer vision and pattern recognition. In recent years, human activity recognitiontechnique has made remarkable progress and has broad application prospects and economicbenefits in the fields of human production, life and scientific research and other high-techfields. Activity recognition is applied to many domains, such as intelligent video surveillance,visual reality, human-computer interface and motion analysis. This paper mainly solve theproblems of activity representation and activity recognition based on spatio-temporal interestpoints, including the detection of spatio-temporal interest points, spatio-temporal interestpoints descriptor, and activity recognition system based on the unsupervised model.In the detection of spatio-temporal interest points, the objective function of the detectionalgorithm is influenced by the change of light intensity easily. Spatio-temporal interest pointsmay be detected in the background by mistake. This paper proposes a novel detectionalgorithm of spatio-temporal interest points based on the region of interest. On the basis offoreground interest region extracted, dense spatio-temporal interest points are obtained by theseparable linear filter method.In terms of the feature description, as bag of words model ignores the spatial andtemporal correlation between features, a novel four-dimensional relative coordinate featuredescriptor is proposed in this paper. The feature is combined with3D-SIFT descriptor torepresent the activity. The method can effectively solve the problem, which is a single bag ofwords cannot reflect the spatial and temporal position relation of the keywords. Then,k-means clustering algorithm is employed to generate double codebooks, which are thefour-dimensional relative coordinate codebook and the3D-SIFT codebook. Finally, theunsupervised probabilistic latent semantic analysis model is used to recognize activity.Experiment results on two activity datasets are provided to demonstrate theeffectiveness.