(I) Technologies to be developed

• Inter-camera correspondence


There are several cues for inter-camera correspondence: face recognition algorithm can provide a similarity score for the faces captured from multiple cameras; space-time cue can estimate probabilities of a human entering a certain camera at a certain time given the location, time and velocity of its exit from other cameras based on the scene modeling results; human appearance cue can be represented by using the distance in color space. In this project, we will develop the algorithms to quantize those cues and method to fuse those factors together


• Self calibration and site modeling


Without manual inter-camera calibration, this system needs self calibration and modeling the scene by training session to learn the camera topology and path probabilities of targets. During the training stage, it is not necessary to track all persons across cameras. Only the best matches (those closest in face and appearance) will be used for learning. Related algorithms will be developed in this project.

(II) Innovation use of existing technologies

• Face Recognition


Frontal faces will be generated by our formerly developed face reconstruction [15-17], which will be adopted as input for face recognition module to get similarity score of a pair of faces. We will select appropriate face recognition algorithm and implement it in this project.


• Human tracking in single camera level


We will follow our previous research on human tracking at single camera level [11-14]. Furthermore, for those severely occluded case, regular tracking algorithms may fail to track the targets even at single camera level. With the assistance of face recognition results, advanced tracking method will be developed in this project to handle these extreme cases.