Person re-identification is the task of recognizing an individual who has previously been observed over a camera network. It is a challenging computer vision task that can provide useful tools for many security applications of video-surveillance, e.g., on-line tracking of individuals over different, non-overlapping cameras, and off-line retrieval of the video sequences containing an individual of interest, whose image is given as a query.
Clothing appearance is the most widely used cue, since the low image resolution and the variety of poses typical of video-surveillance settings make face recognition ineffective. State-of-the-art re-identification methods build descriptors of clothing appearance often based on a subdivision of the body into parts (e.g., torso and legs), and extract a set of low-level features (e.g., related to colour or texture) from each part (or form the whole body), like SIFT points, or small image patches.
At PRA Lab, we developed a new, dissimilarity-based descriptor. It turns any
multiple part/multiple component descriptor into a vector of dissimilarity values between each body part and a set of visual prototypes
, i.e., sets of low-level features representative of clothing characteristics of the corresponding body part of several individuals. Our dissimilarity-based descriptor attains a considerable speed-up of the matching step of person re-identification systems, enabling real-time applications
, without losing accuracy. We also exploited it to implement the new people search
A typical application scenario of person re-identification: a network of video surveillance cameras monitoring a large public space. The girl is seen at first by the camera in the upper-left corner, then by a second camera of the network (lower-right corner). A person re-identification system should associate these views to the same identity.