Abstract:

Event detection requires interpretation of the "semantically meaningful" object actions. To achieve this task, the gap between the numerical features of objects and the symbolic description of the meaningful activities needs to be bridged. We develop an ambiguity detection framework that has two significant advantages over past work. First, we introduce a fusion method for a set of time-wise and object-wise features including not only the trajectory coordinates but also the histograms and HMM based representations of object's speed, orientation, location, size, and aspect ratio. This fusion method enable detection of events that cannot be detected with the existing trajectory features reported so far. Second, we improve existing spectral clustering algorithms by automatically estimating the optimal number of clusters. Furthermore, we determine the conformity of the objects within the given data space. We compute a separate HMM for each object using a time-series that is composed of the mixture of its features. Then, we construct an aggregated affinity matrix from the pair-wise similarity scores of objects using the HMM's. We apply eigenvector decomposition and obtain object clusters. We show that the number of eigenvectors used in the decomposition is proportional to the optimal number of clusters. We examine the affinity matrix to determine the deviance of objects from common assemblages within the space. Our simulations reveal that the proposed detection methods accurately discover both usual and unusual events.