Unusual Event Detection
The ultimate goal of most surveillance systems is automatic detection of events and suspicious activities thereby triggering alarms (detection) as well as reducing the volume of data presented to human operator (retrieval). Event detection requires interpretation of the "semantically meaningful object actions." To achieve this task, the gap between the numerical features of video objects and the symbolic description of their meaningful activities needs to be bridged. Highway monitoring, airport surveillance, building access control are just a few of the several important applications.
Background & Objective: Past work on event detection has mostly consisted of extraction of object trajectories followed by a supervised learning using parameterized models for actions. Models are usually predefined dynamic patterns of movements. Such methods assume that 'all' unusual event events can be modeled, which requires off-line training. However, it is not viable to foresee every possible event. Besides, the nature of event varies depending on the application, thus event modeling becomes even more challenging.
Unlike the past work cited above, we employ an unsupervised learning method. Our method does not require definition of what is usual and what is not. We define usual as the high recurrence of events that are similar. As a result, unusual is the group of events that are not similar to the rest. This enables as to detect multiple unusual events. The developed technique enables fusing different type of features, e.g. histograms, HMM's scalars, etc. as well.
Technical Discussion: We develop an event detection framework that has two significant advantages over past work. First, we introduce an extended set of time-wise and object-wise statistical 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. These features enable detection of events that cannot be detected with the existing trajectory features reported so far. Second, we introduce a spectral clustering algorithm that can automatically estimate the optimal number of clusters. First, we construct feature-wise affinity matrices from the pair-wise similarity scores of objects using the extended set of features. To determine the usual events, 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.
Publications:
Technology Area: Computer Vision
Modification Date: July 14, 2005