People search based on attributes description provided by an eyewitness for video surveillance applications

People search based on attributes description presents a paramount task for several forensics and surveillance applications. The aim is to locate a suspect or to find a missing person in public areas. However, semantic attributes provide a natural interface for this system as they present human understandable properties. These features can cover the whole body characteristics by describing the worn bags, carried objects, clothes, accessories, etc. Detecting semantic attributes under uncontrolled acquisition conditions still remains a challenging task. Most of state-of-the-art approaches assume independence among attributes where each attribute classifier is trained independently based on low-level features extracted from training samples. In this paper, we propose a novel people search system based on attributes description that relies on several components. An interactive query verification algorithm is introduced to prevent search failure. In addition, an attribute classification method that relies on two steps is introduced. We start by selecting the most relevant features in attribute adaptive way. Then, we explored the interactions among attributes to predict a semantic trait by involving the independent attribute classifier and the other correlated attribute classifiers. Several experiments were conducted to validate the effectiveness of the proposed people search system on the challenging VIPeR, CUHK, and HDA+ datasets benchmark.