Use of convolutional neural networks to automate the detection of wildlife from remote cameras

Determining the abundance, distribution, and habitat associations of wildlife species is important for understanding their ecology, behavior and conservation. For mobile, rare, and wide-ranging species, biologists often obtain this information from remote cameras and time-lapse photography. The captured images are then visually inspected to identify those that contain useful information. Due to the large number of images to be processed, the task of visual inspection is painstaking and tedious. In this paper, we describe preliminary results of an automated screening system that is intended to alleviate this problem. Specifically, we study the problem of detecting grizzly bears (Ursus acrtos) in still images, using a convolutional neural network (CNN). Given each image, we first use the Maximally Stable Extremal Regions (MSER) to segment sub-regions that potentially contain a bear and then apply a pre-trained convolutional neural network as the classifier to determine if a bear is present in a sub-region. Experimental results from a real-world dataset demonstrate that our system is able to eliminate over 90% of the images from human inspection while recalling over 60% of the positive images that contain a bear, at a rate of approximately one minute per image.