Path Modeling and Retrieval in Camera Networks
We defined a framework implementing filtering and querying capabilities to execute data mining procedures for high-level analysis on surveillance databases.
The analyst is presented with the most likely path that a queried person could have followed starting from (or ending to) a specified FOV during a predefined temporal window (research at Boston University + University of Palermo).
Lo Presti L., Sclaroff S., La Cascia M., “Path Modeling and Retrieval in Distributed Video Surveillance Databases”, IEEE Transactions on Multimedia (TMM), April 2012, (pdf)
Appearance Description/Matching across Camera Networks
We used a Latent Dirichlet Allocation (LDA) machine to learn, from a training set, the latent structure of appearance features (i.e. combinations of colors and textures) in the data and used this machine to describe the appearance of the incoming objects.
We also developed a method to establish matches among objects detected in different FOVs within the camera network by finding correspondences among the LDA machines maintained at each camera node (research at Boston University + University of Palermo).
Lo Presti L., Sclaroff S. and La Cascia M., “Object matching in distributed video surveillance systems by LDA-based appearance descriptors”, ICIAP 2009 (pdf)
Camera Networks and Embedded Systems
We developed methods and techniques to manage distributed embedded surveillance systems that are able to fuse the collected local information to globally understand the activity in the whole site.
Overlapping Camera FOVs
We proposed a method to learn the correspondences among FOVs on-line while people are moving in the site. This provides geometrical constraints for people re-identification.
Lo Presti L. and La Cascia M., “Real-time object detection in embedded video surveillance systems”, WIAMIS 2008 (pdf)
Lo Presti, L., M. La Cascia, “Real-time estimation of geometrical transformation between views in distributed smart-camera systems”, ICDSC 2008 (pdf)