Machine vision algorithms for robust animal species identification

Abstract

Numerous military bases have a requirement, based on the Sikes Act, to maintain the base's natural environment while still meeting military mission objectives. One method used to accomplish this is by working towards the goal of achieving habitat and species sustainability. One difficulty is that there is currently no baseline of the ecosystem. Specifically, a critical need is the detection and identification of animals on Federal and State endangered lists. For instance, the U.S. Fish and Wildlife Service lists 130 animals as either endangered or threatened, including the desert tortoise, the Mohave ground squirrel, various species of fox, jaguar, mountain beaver, and wolf. In order to even begin to form an appropriate natural environmental baseline, the location and movements of these animals must be acquired, recorded, and made available for review. To this end, in this presentation we detail technology and machine vision algorithms that can be used to: 1.) Recognize animals that are on the endangered or threatened lists, 2.) Identification of animals without the need to track them in sequential image frames, 3.) Provide continual animal census surveillance for weeks at a time in operational environments, and 4.) Record video and still-image data along with annotations for later analysis. Specifically, present an extendable architecture for species identification and identification software truthing/training, and populate this architecture with three recognition modules: a Haar Cascade classifier, a Local Binary Pattern cascade classifier, and a neural network. We also detail the results of our work, current challenges, and future approaches we are taking with our research.