A Deep Neural Network Machine Vision Application for Preventing Wildlife-Human Conflicts

Abstract

Most wildlife-human conflicts can be prevented if humans, who could potentially be affected, can be alerted about the presence of wildlife nearby so that they can take avoidance measures. The alerts must be accurate and timely so that such measures can be taken. We propose a Deep Neural Network consisting of two stages, that we call ‘WildlifeNet’, to automatically detect the presence of specific wildlife. WildlifeNet is optimized for low power and low memory so that it can be embedded in edge devices such as surveillance cameras or low cost special-purpose cameras. The first stage in WildlifeNet is an object detection system using the MobileNet model in TensorFlow that detects animals in an image. This is followed by our custom Convolutional Neural Network classification system that identifies specific animal species from the animals detected in the first stage. WildlifeNet uses images from surveillance cameras or low cost cameras placed near typical animal paths to detect the presence of wildlife. The components surrounding WildlifeNet in the machine vision system presented in this paper can quickly alert those living near the specific location where detections occur via their mobile phones. The custom Convolutional Neural Network model in WildlifeNet’s second stage was trained using a large number of coyote images from the Caltech wildlife image dataset to demonstrate its usefulness in detecting specific wildlife. We observed a consistently high accuracy of coyote detection with a potential towards even higher accuracies with user feedback. Therefore, this system is a viable candidate for consideration as an effective, fast, low-cost technology to assist in preventing wildlife-human conflicts.