Person re-identification involves matching the identity of an individual as they appear at different times and locations across a network of cameras. This is useful for monitoring and understanding people’s behaviour in an environment, and can assist operators in tracking the movements of people over a large camera network and can enable the retrieval of video sequences where an individual of interest appears.
In the work of this project, a person’s whole body appearance is described using “soft biometrics”, which are traits that describe an individual, but lack the permanence and distinctiveness to uniquely identify them (e.g. height, hair and skin colour as well as clothing characteristics). Soft biometrics are particularly well suited to surveillance applications as they can be acquired unobtrusively in complex environments and when fused together can provide course authentication and identification at long range. In addition, these features can be used to locate a person based on a description.
We have released a dataset of 150 subjects, to evaluate person re-identification and soft-biometric models in a real environment, to determine and evaluate which features are best for recognising the identity of a person in low resolution footage across different camera views, illumination conditions and pose.