Loading
Research Projects
Person Re-Identification
|
Abstract
Tracking multiple people across cameras with non-overlapping fields of view is a challenging problem in the domain of visual surveillance. Most of the existing work deals either with designing appearance signatures or learning a discriminative appearance model that can stably re-identify the same person on different cameras. Our work on the other hand is focused on making inferences based on both appearance and spatio-temporal information. Given a collection of observations, we cluster the observations based on appearance similarity and region of surveillance. Then, on each cluster we find an optimal partitioning of observations that best justifies the identity assignment. We also propose an error correction strategy to account for noise in the appearance descriptor of a person.
Advisor(s): Carlo Tomasi
Ronald Parr, Pankaj Agarwal