Towards an automatic camera trap pipeline

Abstract

This thesis delineates a robust pipeline aimed at bolstering ecological monitoring and con servation endeavors through automated camera trap analysis. The discourse is structured around three core segments: detection, species identification, and ecological inference analysis. In the detection phase, a novel model architecture inspired by the DETR model is proposed, which leverages multi-frame input akin to human analysis employed for iden tifying occluded animals. This augmentation yielded a marked improvement in detection performance. The narrative then transitions to species identification, where a unique applica tion of embedding models to the camera trap domain is introduced. A novel aspect of this segment is the development of an adaptive margin for triplet learning, grounded in phylogenetic tree structures, which enhanced the accuracy of subsequent classification tasks. By generating embeddings from extracted animal imagery and utilising K-nearest neighbours (KNN) for classification or similarity checking, this approach significantly mitigates the necessity for specialised classifiers, thereby aiding ecology teams devoid of in-house machine learning expertise. Lastly, the thesis introduces a method employing Bayesian Networks for ecolog ical inference, based on species observations across various locations. This innovative methodology facilitates the construction of models for intricate what-if analyses regard ing species interrelations and their spatial distributions. For instance, examining the implications of a Hedgehog sighting at a particular location on the likelihood of Fox observations in neighbouring areas. These substantial contributions allow ecologists to increase capacity in camera trap based research by automating large efforts required in the analysis of data, in addition providing robust tools for more insightful and scalable ecological monitoring.