Animal detection and species classification on Swiss camera trap images using AI

Motion-triggered camera traps are essential for the monitoring and management of wildlife. As per today in Switzerland, a high number of pictures is manually processed (annotated and classified). We study the utilization of available detection and classification models to (semi-)automatize this process. Two main aspects were investigated: 1) evaluate the feasibility of a non-expert local application (with Microsoft's MegaDetector model), and 2) quantify model performance using several labelled datasets of varying quality and content. Our results show a highly accurate (sensitive and specific), and reliable, fast inference which efficiently allows the automatic pre-discarding of all non-animal images. Further, the MegaDetector turns out to be both, user-friendly and highly performant and thus, an ideal tool for Swiss wildlife experts and stakeholders. Incentives (educational and financial) are required to promote knowledge transfer to this field