The feature extraction device according to one aspect of the present disclosure comprises: a reliability determination unit that determines a degree of reliability with respect to a second region, which is a region that has been extracted as a foreground region of an image and is within a first region that has been extracted from the image as a partial region containing a recognition subject, said degree of reliability indicating the likelihood of being the recognition subject; a feature determination unit that, on the basis of the degree of reliability, uses a first feature which is a feature extracted from the first region and a second feature which is a feature extracted from the second region to determine a feature of the recognition subject; and an output unit that outputs information indicating the determined feature of the recognition subject.