In this paper, we propose a system capable of detecting and identifying people in videos captured by smart phones. We discuss the challenges to extend existing location aware multimedia applications from annotating static landmarks in distance to annotating dynamic people in close range with significant pose variations. We propose to use a hybrid video and RF tracking system to enable accurate observer and target localization, and extract part models comprised of Maximally Stable Color Regions for each target. The model can efficiently detect possible positions of targets in the video, which are then used as dynamic landmarks to calibrate the camera orientation. Finally, positions of all targets in the video are jointly estimated using both visual features and spatial constraints. Experiments show that our approach can locate identified targets in video with significantly higher accuracy than back-projection using camera orientation estimations from accelerometers and magnetometers.

Reliable tracking of people in video and recovering their identities are of great importance to video analytics applications.For outdoor applications, long range identity sensors such as active RFID can provide good coverage in a large open space, though they only provide coarse location information. We propose a probabilistic approach using noisy inputs from multiple long range identity sensors to globally associate and identify fragmented tracklets generated by video tracking algorithms. We extend a network flow based data association model to recover tracklet identity efficiently. Our approach is evaluated using five minutes of video and active RFID measurements capturing four people wearing RFID tags and a couple of passersby. Simulation is then used to evaluate performance for larger number of targets under different scenarios.identities are of great importance to video analytics applications.For outdoor applications, long range identity sensors such as active RFID can provide good coverage in a large open space, though they only provide coarse location information. We propose a probabilistic approach using noisy inputs from multiple long range identity sensors to globally associate and identify fragmented tracklets generated by video tracking algorithms. We extend a network flow based data association model to recover tracklet identity efficiently. Our approach is evaluated using five minutes of video and active RFID measurements capturing four people wearing RFID tags and a couple of passersby. Simulation is then used to evaluate performance for larger number of targets under different scenarios.