University of Houston

Network of multiple cameras are typically used to monitor large geospatial areas by law enforcement and security authorities in public and private facilities. Such networks are typically under utilized as the manual monitoring of huge amounts of video data is not feasible. Automated video surveillance techniques hold the promise to efficiently process video data and enable a proactive approach to detecting events that need the attention of security personnel. The detection, tracking and analysis of pedestrians are among key issues to be addressed in video surveillance systems in order to detect abnormal or suspicious human activity.

Re-identification is a fundamental task necessary for analysis of long term behavior of individuals or connecting interrupted tracks for multi-camera tracking. Re-identification is defined as matching of images/videos of objects taken from different cameras and has applications in several surveillance tasks. It is a challenging task to accomplish due to the changes in camera viewpoints, pose and illumination conditions across different cameras.

In this work, we address the problem of person re-identification in distributed camera networks for the purpose of automated video surveillance. We develop a person model and study its effectiveness in the context of multi-camera tracking. Our person model is part-based and incorporates multiple cues like clothing appearance and facial features. The model is learnt over multiple images of the person developing a truly spatio-temporal model. The utility of the model is tested in the multiple person re-identification scenario. We solve the multiple person re-identification as a combinatorial optimization problem of rectangular assignment.

Moreover, we try to implicitly incorporate a false match rejection technique within the re-identification framework which is a relatively unexplored aspect of the re-identification problem. The aim is to devise strategies to reject false matches purely based on the person model matching cost in order to alleviate the need to leverage prior knowledge of spatio-temporal relationships between cameras in the network.