Title: A Review of Computer Vision Approaches for Non-Invasive Animal Re-Identification in Large-Scale Camera Networks
Authors: Kingsley Chinazaekpere Ndupu, Serif Oyindamola Oyesiji, Stanley Nwakamma, Harouna Wendpanga Yann Christian Sankara.
Volume: 10
Issue: 9
Pages: 114-132
Publication Date: 2026/09/28
Abstract:
The ability to recognize individual animals repeatedly and reliably underpin a wide range of scientific and operational activities, from population ecology and conservation management to precision livestock farming and biosecurity. Traditional approaches to individual identification, including physical tagging, branding, collaring, and manual photographic matching, are invasive, labor intensive, or both, and they scale poorly to the camera networks now deployed across protected areas, rangelands, and farms. This review synthesizes the computer vision literature on non-invasive animal re-identification, with particular attention to the pipeline that has become dominant in large-scale deployments: convolutional object detection, most prominently the YOLO family of single-stage detectors, followed by deep feature extraction into high-dimensional embedding spaces in which individual identity is resolved by similarity search. The review traces the development of the field from early pattern-matching systems built on hand-crafted features through the adoption of deep metric learning, examines species-specific applications spanning big cats, primates, cetaceans, pinnipeds, elephants, bears, and domestic cattle and pigs, and analyzes the systems-level challenges introduced by distributed multi-camera networks, including open-set recognition, cross-camera appearance variation, embedding indexing at scale, and edge deployment under constrained power and bandwidth budgets. Persistent obstacles are identified in dataset scarcity and imbalance, domain shift between capture locations, the absence of standardized evaluation protocols, and the integration of algorithmic outputs with statistical population models. The review closes with directions for future research, emphasizing self-supervised representation learning, cross-species transfer, and uncertainty-aware identification suitable for downstream ecological inference.