Person re-identification in multi-camera system by signature based on interest point descriptors collected on short video sequences

Abstract

We present and evaluate a person re-identification scheme for multi-camera surveillance system. Our approach uses matching of signatures based on interest-points descriptors collected on short video sequences. One of the originalities of our method is to accumulate interest points on several sufficiently time-spaced images during person tracking within each camera, in order to capture appearance variability. A first experimental evaluation conducted on a publicly available set of low-resolution videos in a commercial mall shows very promising inter-camera person re-identification performances (a precision of 82% for a recall of 78%). It should also be noted that our matching method is very fast: ~ 1/8s for re-identification of one target person among 10 previously seen persons, and a logarithmic dependence with the number of stored person models, making reidentification among hundreds of persons computationally feasible in less than ~ 1/5 second.

Multiple-shot person re-identification by chromatic and epitomic analyses

“…The former models a person analyzing a single image for each individual, not exploiting the temporal information provided by tracking, such as Prosser et al (2010), Gray and Tao (2008), Schwartz and Davis (2009), Bak et al (2010), and Zheng et al (2009). The latter group, instead, employs multiple images of a person (obtained via tracking) to build the descriptor used for re-id, such as Bird et al (2005), Gheissari et al (2006), Hamdoun et al (2008), Nakajima et al (2003), and Farenzena et al (2010). Bird et al (2005) defines a descriptor built by subdividing the person in horizontal stripes, keeping the median color of each stripe accumulated over different frames.…”

“…A matching between decomposable triangulated graphs, capturing the spatial distribution of local temporal descriptions, is presented by Gheissari et al (2006). Hamdoun et al (2008) uses SURF interest points, collected over short video sequences. Another supervised learning-based approach is proposed by Nakajima et al (2003): local and global features accumulated over time are fed into a multi-class Support Vector Machine (SVM) for recognition.…”

“…2) We do not simply accumulate local features with heuristics, as Hamdoun et al (2008) and Farenzena et al (2010), but we keep recurrent local aspects by analyzing the epitome result- ing from the images of several person, that may reappear with higher probability in novel instances of the person. 3) We do not employ discriminative learning techniques as in Nakajima et al (2003), that have to be re-trained each time a novel subject appears.…”

Person re-identification in crowd

“…The Euclidean distance between two colors is also included in an ad-hoc similarity measure created to compare two DCD feature sets (Bak et al, 2010). Alternative measures are the sum of quadratic distances (Oliveira and Luiz, 2009) and the sum of absolute differences (Hamdoun et al, 2008). Other distance measures include the Kullback-Leibler Distance, KLD (Jeong and Jaynes, 2008;Berdugo et al, 2010), and the Bhattacharyya Distance, BD (Prosser et al, 2008;Farenzena et al, 2010).…”

“…Finally, interest points can be used for re-identification in case of variations in scale, pose and illumination (Bauml and Stiefelhagen, 2011). Examples are SIFT (Teixeira and Corte-Real, 2009), SURF-like features (Hamdoun et al, 2008;Oliveira and Luiz, 2009) and the Hessian Affine invariant operator (Gheissari et al, 2006). When intra-camera tracking information is available, features extracted from single images can be grouped over time either by temporal accumulation (Hamdoun et al, 2008) or by clustering (Farenzena et al, 2010).…”