Part-based spatio-temporal model for multi-person re-identification

In this paper we propose an adaptive part-based spatio-temporal model that characterizes person’s appearance using color and facial features. Face image selection based on low level cues is used to select usable face images to build a face model. Color features that capture the distribution of colors as well as the representative colors are used to build the color model. The model is built over a sequence of frames of an individual and hence captures the characteristic appearance as well as its variations over time. We also address the problem of multiple person re-identification in the absence of calibration data or prior knowledge about the camera layout. Multiple person re-identification is a open set matching problem with a dynamically evolving and open gallery set and an open probe set. Re-identification is posed as a rectangular assignment problem and is solved to find a bijection that minimizes the overall assignment cost. Open and closed set re-identification is tested on 30 videos collected with nine non-overlapping cameras spanning outdoor and indoor areas, with 40 subjects under observation. A false acceptance reduction scheme based on the developed model is also proposed.

► Adaptive spatio-temporal model for re-identification is proposed.
► Model leverages color and facial features.
► Multiple person re-identification as open set matching is studied.
► False acceptance reduction criteria proposed.
► Facial features inclusion improves re-identification performance.