Artificial Intelligence and Camera Trapping: Investigating computer-based approaches to accelerate camera trap data processing

Abstract

Global biodiversity is declining with serious implications for ecosystem stability, ecosystem services, and the livelihoods and wellbeing of humans. More effective conservation is needed, which requires data to form evidence-based decisions. An increasingly popular method of collecting data on fauna uses motion-sensing cameras called camera traps (CTs). With increased monitoring effort comes increased volumes of data, which are logistically challenging to analyse in timely manner. This thesis investigates advancements in a subsection of Artificial Intelligence targeting the processing of visual data called Computer Vision (CV). More specifically, I combine developments in Computer Science, Ecology, and Citizen Science (CS) in an interdisciplinary approach to investigate the performance of novel approaches to processing CS CT data. Recognising the importance of depth perception for many approaches to CT-based density estimation, I begin by assessing CV methods to replace the human labour currently needed to extract distance measurements from CT images. I find it possible to eliminate human effort beyond CT deployment, but widespread adoption of these methods requires development of new commercial hardware or CV algorithms. Next, I describe the creation of a dataset designed for use with a set of CV algorithms called Deep Neural Networks (DNNs). I go onto compare the performance, when applied to CT data, of DNNs designed for image classification and object detection, finding the latter overall more suitable. An exploration of deployment strategies follows, showing a single multi-class network more appropriate than an ensemble of binary networks. Next, using CV predictions to calculate ecological outputs is shown to be possible depending upon the analysis. After, I detail a proof-of-concept algorithm to link single-frame analysis across multiple frames, and how to combine this multi-frame output with CS outputs. The final Chapter summarises my findings and discuses implications, followed by highlighting future avenues of research to continue progress.