Machine learning for animal detection and recognition for European wildlife conservation

Abstract

This Master of Science thesis explores the potential of Computer Vision (CV) for wildlife research and conservation in Europe. I collaborated with Theodoros Kominos, a wildlife researcher who provided me images and video recordings from trap cameras he set in Greece at the past. The data includes various wild animals, some of which are protected in Greece, such as brown bear, wolf, chamois, and wildcat. Other species like roe deer, wild boar, red fox, European badger, other mustelids, and European hare are also present. The data also contains domestic animals (like cows, dogs, horses) and human activities (hikers, potential hunters, conservation workers, and vehicles). The main objective was to develop a dataset of annotated images from this raw data and implement ΑΙ algorithms to analyze it. The goals were to classify and detect different animal species and identify potential threats to their habitats. Additionally, I explored how these AI techniques could enhance existing conservation efforts. The process involved organizing and annotating the visual data provided by the researcher. I then employed computer vision techniques and trained models for accurately identifying and classifying the various subjects in the images and videos. The study also focused on detecting potential threats to wildlife habitats by identifying forbidden or concerning activities in the monitored areas. This includes detecting unauthorized human presence, illegal hunting, or other activities that could harm wildlife or their habitats. This aspect has significant implications for proactive conservation efforts and the development of early warning systems for habitat protection, allowing for timely intervention when threats are detected. I explored how these AI techniques could be integrated into existing conservation practices, developing user-friendly interfaces and workflows to incorporate these tools into daily operations. This integration aims to streamline data analysis, reduce manual labor, and provide more accurate and timely information for wildlife management decisions. Furthermore, I propose potential implementations of edge hardware for use by conservationists or in areas of human-wildlife conflict. The intended results of this research include practical AI tools and techniques for wildlife researchers and conservationists to better understand, protect, and manage biodiversity in Europe. In conclusion, this thesis demonstrates the potential of Computer Vision in enhancing wildlife research and conservation efforts, covering a wide range of species from large mammals to smaller, less studied animals. By providing more accurate, efficient, and scalable methods for monitoring wildlife and their habitats, this work could potentially be adapted for preserving Greece's and Europe's biodiversity.