Detection of People in Military and Security Context Imagery

摘要

A high level of manual visual surveillance of complex scenes is dependent solely on the awareness of human operators whereas an autonomous person detection solution could assist by drawing their attention to potential issues, in order to reduce cognitive burden and achieve more with less manpower. Our research addressed the challenge of the reliable identification of persons in a scene who may be partially obscured by structures or by handling weapons or tools. We tested the efficacy of a recently published computer vision approach based on the construction of cascaded, non-linear classifiers from part-based deformable models by assessing performance using imagery containing infantrymen in the open or when obscured, undertaking low level tactics or acting as civilians using tools. Results were compared with those obtained from published upright pedestrian imagery. The person detector yielded a precision of approximately 65% for a recall rate of 85% for military context imagery as opposed to a precision of 85% for the upright pedestrian image cases. These results compared favorably with those reported by the authors when applied to a range of other on-line imagery databases. Our conclusion is that the deformable part-based model method may be a potentially useful people detection tool in the challenging environment of military and security context imagery.