Novel methods for Autonomous/Assisted Human Activity Modeling and Anomaly Detection

Recognition of human intention using infrared video sequences has applications in surveillance. Although motion trajectories extracted from IR sequences contain significant information about the underlying event, we need a methodology to connect trajectories and intentions. ImageCorp, Inc proposes to implement two deformable-shape based descriptions for characterizing human intentions. The first one is based on the factorization of motion trajectories in terms of basis shapes such that every activity is represented as a weighted sum of basis shapes. When a video is given, the method extracts motion trajectories and identifies the closest activity using a similarity measure. If an unforeseen event occurs, the corresponding motion trajectory will be poorly represented by the existing basis shapes and so will be seen as an outlier. The second approach uses statistical shape theory for modeling the mean shape of the configurations of moving humans and its deformation over time. The mean shapes are represented using Markov processes, which enable us to derive conditions for identifying similar intentions as well as abnormal ones. Both approaches lend themselves for operator intervention and on-line training. We will evaluate the effectiveness of these two approaches using in house data and any sponsor provided data.