Small mammals and bird detection using IoT devices

Abstract

This article describes the application of computer vision to automatically detect pests in protected forests. One of the most threatening pests in NZ are rodents (rats, mice, possums etc). The objective of the application is to allow the deployment of Internet of Things (IoT) devices capable of detecting rodents while differentiating them from birds. As IoT devices have limited memory and CPU capabilities, it is important to find computer vision methods that are efficient and lightweight. Two main methods were employed for this work to test their suitability for this application. The first method is based on getting a profile of moving objects and classify them using Fourier descriptors (FD) and classifiers. The second method uses YOLO (You Only Look Once) to classify birds versus rodents. In order to increase the number of images for training YOLO, a semi-automatic labelling system was created, using the FD method to segment images of birds and rodents. The final accuracy rate using FD reached 83 percent with random forests, and YOLO reached 97 percent. The FD method can outperform YOLO in terms of speed when using a Raspberry Pi, achieving more than 6 frames per second.