Reliable object tracking in video sequences is a very important issue in most video surveillance applications as it constitutes a preliminary step for the interpretation of the scene. Among the many issues that affect the correct determination of the trajectory followed by the objects while moving into the scene, the most relevant ones are represented by those events that break the continuity of the trajectory, such as occlusions or temporary exits of the objects from the scene. In this paper we introduce a method that relies on a 3D model representation, called MultiView Appearance Model, that proved to be effective in the re-identification of the object after a long lasting occlusion or when it re-enters into the scene. We also present the proposed procedure for the management of the model, from the creation, to the update and the removal, while taking into account the problem of object apparent size changes due to perspective. The method has been tested on the public PETS2010 dataset. The results show significant performance improvements to other appearance representation models.
A 3D Appearance Model for Object Tracking in Video Surveillance Applications
CONTE, Donatello;FOGGIA, PASQUALE;PERCANNELLA, Gennaro;VENTO, Mario
2012-01-01
Abstract
Reliable object tracking in video sequences is a very important issue in most video surveillance applications as it constitutes a preliminary step for the interpretation of the scene. Among the many issues that affect the correct determination of the trajectory followed by the objects while moving into the scene, the most relevant ones are represented by those events that break the continuity of the trajectory, such as occlusions or temporary exits of the objects from the scene. In this paper we introduce a method that relies on a 3D model representation, called MultiView Appearance Model, that proved to be effective in the re-identification of the object after a long lasting occlusion or when it re-enters into the scene. We also present the proposed procedure for the management of the model, from the creation, to the update and the removal, while taking into account the problem of object apparent size changes due to perspective. The method has been tested on the public PETS2010 dataset. The results show significant performance improvements to other appearance representation models.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.