Detektion och Klassificering av Svenska Vilda Djur med hjälp av Maskininlärning

Monitoring wildlife is essential for biodiversity research and for managing threats to agriculture and forestry. However, manual analysis of large volumes of camera trap images is both time-consuming and inefficient. This thesis presents the development and evaluation of a machine learning model designed to detect Swedish wild animals in natural environments. The aim is to support farmers, foresters, and cattle owners in protecting their livelihood by enabling real-time animal classification. The model focuses on recognizing key species such as predators (foxes, wolves, lynxes, bears, and wolverines), ungulates (roedeer and moose), large birds, wild boars, and hares. The work began with a detailed analysis of available data sources, including imagery from Qulinda, Artdatabanken, eBird, GBIF, and MS COCO, to construct a robust and diverse training dataset. Multiple YOLO versions, ranging from YOLOv8 to YOLO12, and their variants were evaluated based on their detection accuracy, inference speed, and class generalization. YOLO11s was ultimately selected for its balance of performance and efficiency. Iterative experiments with class structures and training configurations refined the model further. Evaluation results show reliable detection performance across various environments, including scenarios with out of distribution objects such as humans, vehicles and other animals. This work offers a structured approach to wildlife detection model development and provides considerations for deploying AI-based support tools in agriculture, forestry, and wildlife monitoring.