Person Retrieval in Surveillance Videos Using Deep Soft Biometrics

Abstract

In the world of security and surveillance, it is very common to localize the person in the video. The localization turns to be non-trivial when the search is based on a linguistic query. This is due to the semantic gap between language based query and its processing by a machine. Typically query uses attributes like height, cloth color, cloth type, gender, and hair color, e.g., a tall male with a black t-shirt and pink short. Such attributes are known as soft biometrics. Conventionally, searching the person in the surveillance video is done manually by scouring through hours of videos, which is inefficient and time-consuming. Thus, an automatic person retrieval algorithm is an active area of research.