Selecting informative samples for animal recognition in the wildlife

Abstract

Observations in the wildlife using cameras traps are very useful in ecological, conservation, and behavioral research of animals and birds. However, a large number of recorded images do not contain the objects of interest, and manual removal of such images is a highly difficult and durable process. We suggest an automatic selection of relevant images in order to prepare the informative samples for following animal recognition and a set of representative images for manual detailed analysis if it is necessary. In this research, we propose two methods based on the background model constructed “on the fly” and Gaussian mixture model. The distorted images by visual artifacts are removed preliminary. The experiments were conducted on 30,000 images captured by camera traps in Ergaki national park, Krasnoyarsky Kray, Russia, 2012–2018. The best accuracy result for selecting informative samples achieved 96% regarding the human estimates.