Salient points driven pedestrian group retrieval with fine-grained representation.

• Explore a less studied issue for the group retrieval in crowd video analysis field. • A non-uniform sampling processing is use to divide group into fine granularity units. • A cost-trim processing is use to lighten the influence caused by high cost pairs. People often take part in various social activities in the form of groups in public area. As the primary constituent units of crowd, groups retrieval has become one of the urgent issues for the security departments. In this paper, collection of stable individuals with some social relationship, called group, is selected as the research object, and a novel task of pedestrian group retrieval is introduced. Different from the individual person matching, groups often show high aggregation due to their inherent characteristics, occlusions in group individuals therefore are more serious. As a result, the performance of individual person based detection and matching will be affected. Meanwhile, group matching also needs to address the problems like variations in the shape or configuration. Therefore, we suggest that the group entirety may be disassembled into fine-grained representation and then design a salient points driven framework for pedestrian group retrieval. The work focuses on the problems of overall appearance characteristics extraction of a deformable pedestrian collection and matching of groups at varying scales. Experiments on Pedestrian-Groups2 dataset and Road Group dataset demonstrate the effectiveness of our proposed framework for Pedestrian Group retrieval. [ABSTRACT FROM AUTHOR]