Department of Pattern Recognition

Abstract:

During the last years, significant improvements were made in people detection, tracking and high-level recognition of events. At the same time, several challenges remains and one of them is error management. Evaluating the quality of tracking outputs is an important task in video analysis. However this task is usually performed offline and need additional data to be used for comparison.

In this presentation we explore a method to automatically detect tracking output errors during runtime by extracting a feature vector that captures the appearance, the motion and the interactions with the other elements of the scene (other actors or background elements). We then try to correct some of these errors by using a re-identification method to identify people who are hidden for a long time due to occlusions or who leave and re-enter the scene.

Short bio:

Julien Badie is a PhD student at INRIA Sophia-Antipolis, France in the STARS team (Spatio-Temporal Activity Recognition Systems). He receive in 2011 a Master’s Degree in Computer Science and Telecommunication specialized in Multimedia, Speed and Image Processing. His main research interests are in people tracking in video surveillance, quality estimation of tracking algorithms and automatic correction of errors. He is now working on the European project CENTAUR on behaviors analysis of crowds.