The University of Massachusetts Amherst
University of Massachusetts Amherst

Search Google Appliance


Yu Awarded “Best Paper Contender”

Electrical and Computer Engineering (ECE) Ph.D. student Xunyi Yu and advisor, ECE Professor Aura Ganz (pictured), won a Best Paper Contender prize at the 7th International Conference on Advanced Video and Signal-Based Surveillance, held from August 29 to September 1 in Boston. Their paper, "Global Identification of Tracklets in Video Using Long Range Identity Sensors," developed a new system, using RFID sensors, to identify people captured in outdoor surveillance videos. This year at the conference, one paper out of the 63 accepted received the Best Paper Award, while another five received Best Paper Contender prizes.

The conference, staged by the Institute of Electrical and Electronics Engineers, is a forum bringing together participants from the worlds of research, industry, and government agencies sharing interest in various forms of surveillance. The conference focuses on underlying theory, methods, systems, and applications of surveillance.

Yu developed his system for Dr. Ganz to aid the response to disasters with many casualties. Surveillance systems based on video input alone have inherent limitations in reliable tracking of people and recovering their identities. Yet identities of the targets are of particular importance for video analytics systems in mass casualty incidents.

Victims of different triage levels need to be highlighted to speed up evacuation, trespassers and authorized rescue personnel need to be differentiated in perimeter surveillance of open spaces to reduce false alarms, and identities of different units need to be fully recovered to annotate video footages for forensic and training purposes.

In Yu’s system, active RFID sensors are used to assist data association in tracking and target identification. Active RFID sensors can easily cover large open spaces, but only provide very coarse location information for identity association. A two-stage data fusion framework is proposed to fuse drastically different video and RFID sensor data.

Tracklets are locally formed from raw video observation at frame rate granularity in the first stage. Each identity is then globally resolved in the second stage using RFID sensor data and tracklet temporal, kinetic, and appearance constraints. The global data fusion is formulated into a network flow problem with side constraints, which can be solved efficiently in practice due to network structure pruning and the reduced problem size using the two-stage approach. (October 2010)