Idiap on LinkedIn Idiap youtube channel Idiap on Twitter Idiap on Facebook
Personal tools
You are here: Home Awards and Jobs News Jagan Varadarajan successfully defended his PhD thesis

Jagan Varadarajan successfully defended his PhD thesis

On 30th August 2012, Jagan Varadarajan successfully defended his PhD thesis entitled "Sequential Topic Models for Mining Recurrent Activities and their Relationships: Application to long term video recordings ", supervised by Dr. Jean-Marc Odobez.

The thesis took place in the context of the SNF funded HAI project as well as the FP7 European project Vanaheim.

The private defense was held on 9 July 2012 and was successfully defended in the presence of an international jury committee consisting of five experts in the field: Prof. Pascal Fua (EPFL, Switzerland), Prof. Shaogong Gong (QMUL, London), Prof. Bernt Schiele (MPI, Germany), Prof. Pascal Frossard and the thesis advisor Dr. Jean-Marc Odobez (Idiap, EPFL, Switzerland).

PhD summary:

The work addresses the problem of mining complex data streams such as videos coming from public surveillance scenes. The thesis proposes several probabilistic graphical models that can, for example, take a video of a traffic junction or a metro station scene and automatically discover activity patterns that occur in the scene. Using this modeling we can answer questions such as: What are the dominant activities occurring in the scene? When do they start and end? Are there causal and co-occurrence relations among activities? Are there cyclic phases given by the traffic signals in the scene? Are there abnormal events occurring in the scene? and so on.

One of the main contributions of this thesis is a novel method called Probabilistic Latent Sequential Motifs (PLSM) that automatically extracts and locates sequential activity patterns from time series data caused due to multiple activities simultaneously. The method is completely unsupervised i.e., uses simple low-level visual features and no human intervention to obtain activities with temporal structure and semantic significance.

The Mixed Event Relationship (MER) model, which is another notable contribution of the thesis, builds on top of the PLSM's detected activities to infer higher level scene semantics such as scene cycles and event relations. While these methods can be extended to mine activities from multi-dimensional time-series data in general; immediate implications are in the domain of surveillance video summarization, indexing and retrieval, abnormal event spotting and automatic stream selection.

For more details, please download the thesis from http://publications.idiap.ch/index.php/publications/show/2394

Sep 07, 2012
Document Actions