Individual Feral Cat Identification in Camera Trap Imagery using Deep Learning

Rio Akbar

doi:10.60867/00000054

The management of feral cats (Felis catus) as invasive species relies on population monitoring techniques such as Spatially Explicit Capture–Recapture (SECR), which usually involves automatic camera traps to record individual identifications in the wild. Using camera trap imagery still requires manual identification of cats in each image, which is time-consuming and limits scalability for large-scale projects. Deep learning, widely applied in wildlife management for species classification, has recently been used to identify individual animals. However, most applications of individual identification have been conducted in controlled or semi-controlled environments and are not directly applicable to the practical needs of feral cat monitoring. Camera traps often produce images with occluded or partial visibility of the cat, as well as various other issues that hinder accurate individual identification by deep learning models. In addition to image quality challenges, in-the-wild monitoring also means new cats can appear. In this thesis, we investigated a ResNet50-based identification model (Feature Concatenation Model; FCM) that learns features from multiple body parts of the cats to identify individuals, as well as a YOLOv12 object detection model to automatically detect those body parts. For experimental evaluation, four body parts from a consistent side view of the cat were used (body, back leg, front leg and tail) from a subset of feral cat camera trap images collected across the Glenelg and Otway regions of Victoria, Australia. A curated subset of 1,050 images from 10 individuals was selected from a dataset of 15,881 images and 135 individuals, based on sufficient visibility of the specific body parts. We found that the Feature Concatenation Model (FCM) effectively identified individual cats from camera trap images using all or a subset of body parts, which provides flexibility when some parts are not visible, with the body (flank) being the most distinctive region. YOLOv12 also proved capable of being an effective detector of body parts with further parameter tuning and training with a larger separate dataset. However, regardless of detector performance, camera trap datasets are inherently imbalanced, with some body parts appearing only a few times. This imbalance creates a bottleneck for the early fusion FCM, which requires equal numbers of samples across body-part classes before combination. Consequently, many images are excluded to balance, and the resulting smaller datasets make the identification performance highly dependent on the quality of the available annotations. We suggest that a late fusion identification model within the pipeline can be further experimented with to overcome this bottleneck. In addition, this thesis focuses on techniques to overcome camera trap image issues. However, for a truly automated system, further work must explore ways to handle new, unseen cats during deployment.

Individual Feral Cat Identification in Camera Trap Imagery using Deep Learning

Files and links (1)

Abstract

Details

UN Sustainable Development Goals (SDGs)

Metrics