Abstract
In this paper, we address the problem of multiple person re-identification in the absence of calibration data or prior knowledge about the geospatial location of cameras. Multiple person re-identification is a open set matching problem with a dynamically evolving gallery and probe set. We present a part-based spatio-temporal model that learns a person's characteristic appearance as well as it's variations over time. The model is based on 2 distinct color features that capture the distribution of chromatic content and generates a signature of representative colors from a person's appearance. The model implicitly retains the meaningful variations while discarding the repetitive and noisy information and outliers. Re-identification is established based on solving a linear assignment problem in order to find a bijection that minimizes the total assignment cost between the gallery and probe pairs. Open and closed set re-identification is tested on 17 videos collected with 9 non-overlapping cameras spanning outdoor and indoor areas, with 25 subjects under observation. A false match rejection scheme based on the developed appearance model is also proposed.