“Déjà vu” – Visual Search for Efficient Reidentification of Pedestrians in Huge Surveillance Databases  

Csaba Beleznai (1) , Martin Winter (2), Martin Hirzer (2), Horst Bischof (2), Josef Birchbauer (3)
(1) Austrian Research Centers GmbH – ARC, Vienna, Austria
(2) Institute for Computer Graphics and Vision, Graz University of Technology, Austria
(3) Siemens Corporate Technology Central Eastern Europe, Siemens AG Austria

Abstract:
We demonstrate an interactive visual search method that finds a given pedestrian in a huge archive of other camera views efficiently. A user-selected pedestrian image or sequence is used to obtain initial discriminative features and an initial ranked list of hypothetical matches. A discriminative pedestrian recognition model is learned in an on-line manner by user interaction assigning positive and negative labels to the initially retrieved results and on-line boosting for feature selection. This enables that the best discriminative features for the current query are selected. The framework typically retrieves the correct match after a few iterations.

The individual steps of the search process are illustrated in the following slides:

Content on this page requires a newer version of Adobe Flash Player.

Get Adobe Flash player

 

The below video demonstrates the search mechanisms of the visual search framework (Note that individual steps depicted in the video are edited/shortened in order to provide a concise summary):

  1. An arbitrary pedestrian (which has been previously detected and tracked by a surveillance system) is selected manually to form the query.
  2. An initial search is performed using a fixed set of features and a first ranked list of hypothical matches is returned (40 images shown in two rows).
  3. If the searched pedestrian is not among the returned hypotheses, a limited set of samples - representing the worst matches - are automatically labelled as negatives. On-line boosting is used for feature selection and a new ranked list of possible matches is generated.
  4. If the search is unsuccessful, an interactive labelling is performed, marking pedestrian images which look similar to the query as positive samples (marked green in the video) and dissimilar ones as negative samples (marked red in the video).
  5. Typically, after few iterations the correct match is retrieved.

Content on this page requires a newer version of Adobe Flash Player.

Get Adobe Flash player