Untitled Document

On-going projects:
Detecting unusual human activity

We present an unsupervised technique for detecting unusual activity in a large video set using many simple features. No complex activity models and no supervised feature selections are used. We divide the video into equal length segments and classify the extracted features into prototypes, from which a prototype-segment co-occurrence matrix is computed. Motivated by a similar problem in document-keyword analysis, we seek a correspondence relationship between prototypes and video segments which satisfies the transitive closure constraint. We show that an important sub-family of correspondence functions can be reduced to co-embedding prototypes and segments to N-D Euclidean space. We prove that an efficient, globally optimal algorithm exists for the co-embedding problem. Experiments on various real-life videos have validated our approach.
[Details] [Result Videos] [Slides] [Demo application (Matlab)]

Summarizing long videos

With the growth of the storage capacity of digital devices the popularity of long digital videos is increasing. The lack of a priori knowledge of structure and the content of long videos demands new algorithms to summarize, search and organize videos without or with minimal human interaction. The size of such data prohibits the use of exhaustive search algorithms, e.g. watching through hours of video (to find a single event) is not a feasible option. Thus, we investigate unsupervised and semi-unsupervised methods for summarizing videos, building an index (for faster search), scene retrieval and categorizing videos. We base our approach on the visual content of the video (image features, objects, events happening in the video, etc.) rather than relying on the title and other verbal classifications.

Identifying the rhythm of activities

Many events in real life are periodical. Finding such periodical patterns can lead to important observations, like discovering the connection behind seemingly unrelated events, building complex activity models and perhaps predicting them. Further application could be measuring the ability of humans to repeat tasks multiple times and comparing their performance in time - a way of measuring fatigue. Unfortunately, real life patterns are not ideal, the periods can be shorter or longer and might not follow each other in equal time steps. Nevertheless, they have common features which makes them recognizable by humans. Our goal is discover such patterns from different human activities, identifying the periodicity, what we call the rhythm of the activity.