Analyzing activities and events in video from motion content

Object motion in video contains important information for video content analysis especially video event detection. The motion contents in its raw form is real-value multidimensional time series. Processing and analyzing this kind of data is not trivial since most standard machine learning algorithms can only be applied to data in vector space. In this work, we explore the techniques to solve three problems related to motion content in video: object motion estimation, motion representation, and motion analysis to recognize video activities. We first present our work on object detection and tracking. These algorithms help us to estimate the object motion, which is the input data for motion representation and analysis. We propose an edit distance based approach to measure the similarity between motion trajectories in its raw form. A stochastic transition model is developed to learn the model parameters from the training data. We develop a novel trajectory representation framework, “bag of segments,” by which trajectories are transformed to a frequency in vector space so that many traditional machine learning algorithms can be directly applied to the motion trajectory data. We present our work on using the Granger causality test to analyze multiobject interactions from motion. By estimating the causality feature of interactive objects, some important video activities can be rigorously defined and recognized using machine algorithms. Finally, a correlation based feature weighting algorithm is presented with application to video activity classification.