ISBN 978-3-8439-2491-7

978-3-8439-2491-7, Reihe Informatik

Mahesh Venkata Krishna
Unsupervised Event Detection in Video Sequences

252 Seiten, Dissertation Friedrich-Schiller-Universität Jena (2015), Softcover, A5

Zusammenfassung / Abstract

In applications of Computer Vision ranging from surveillance to microscopy video analysis, an important challenge is to detect events happening in the video sequences. An event may be defined as a deviation from the normal happenings in a scene. It may be in the form of an activity not seen before or a rapid succession of movements by an actor, or even a complete change in the scene.

Traditional machine learning systems tackling this problem rely on the availability of sufficient amount of labeled training data for constructing a model of the scene. However, in many application scenarios, such a training dataset is either not available or is extremely difficult to create. In such scenarios, unsupervised event detection schemes become imperative.

The focus of this thesis is to develop such unsupervised event detection systems. In the first approach we propose in this work, we use a One-Class-Classification based novelty detection scheme to perform the task of video segmentation, where segmentation boundaries denote temporal points where events take place.

Further on, we present existing possibilities for event detection systems based on the Hierarchical Dirichlet Processes (HDPs), and note that HDPs have some shortcomings in terms of execution time, modeling of temporal information and continuous feature spaces. To tackle the issue of computational complexity, we develop a combined generative-discriminative model, where the Bayesian inference step is limited to the training stage and in the testing stage, a fast discriminative classifier is used.

This, we show, results in order of magnitude improvement in execution times.

Next, we tackle the issues of imbibing temporal information and continuity feature spaces in HDP models.

For this purpose, we combine ideas of HDP-HMM and wordless LDA to arrive at the wordless HDP-HMM. We go on to derive an inference scheme for this model using the Gibbs' sampler and demonstrate the performance of the model through the application of clustering cells according to the stage of mitosis they are in. The results demonstrate that our model matches the state-of-the-art and in many situations, outperforms it.