Deep learning in multiple animal tracking: A survey

Abstract

Multiple animal tracking based on deep learning is a crucial and demanding task with extensive applications in agriculture, livestock farming, and ecological research. The primary objectives comprise monitoring livestock health, preserving endangered wildlife, and automatically detecting animal behaviors. The task involves a two-step procedure of localizing and tracking individual animals across sequential video frames. However, tracking animals in complex scenarios is still prone to challenges like occlusion and identity switches due to intensive and similar motion patterns. Conventional animal tracking methods fail to meet the precision and real-time speed requirements of real-world applications. In recent years, deep learning has attained remarkable achievements in multiple animal tracking by leveraging deep neural networks for feature extraction. This paper presents a comprehensive review of the development and applications of various multiple animal tracking approaches over the past five years. These methods are categorized according to tracking paradigms, covering diverse animal species comprising both livestock and wildlife. To begin with, we analyze various animal tracking algorithms constructed upon diverse tracking paradigms and summarize their strengths and weaknesses concerning animal subjects. Subsequently, we present open-source benchmark datasets and evaluation metrics, while also discussing and analyzing their applicability in animal tracking. In addition, we review practical applications of animal tracking in agricultural, wildlife, and laboratory environments. Finally, we present an in-depth discussion of datasets, tracking approaches, and promising future perspectives. We hope that our work can make some contributions to the field of multiple animal tracking.