Correctly identifying the observed subjects is an important problem camera networks. Prior art[1, 5] has demonstrated that this data association problem is indeed very difficult when working solely with visual information provided by the cameras, because the appearance of the subjects are highly variable. Visual data provided by surveillance cameras are in general noisy, low resolution, prone to degradation due to lighting and other adverse effects. We hypothesize that knowing the social associations of people can improve the recognition performance of a given visual-only matching metric. We cast the problem as bipartite graph matching problem between the observed people in the camera network and a database of identities and appearance models with an additional pairwise configuration cost on the set of identities. The effectiveness of our claim is demonstrated on a dataset synthesized from UC Irvine Pedestrian Recognition Dataset (VIPeR[3]) (for visual data) and Enron Email Dataset (for social network data).
展开▼
