Application of deep learning for livestock behaviour recognition: A systematic literature review

Livestock health and welfare monitoring is a tedious and labour-intensive task previously performed manually by humans. However, with recent technological advancements, the livestock industry has adopted the latest AI and computer vision-based techniques empowered by deep learning (DL) models that, at the core, act as decision-making tools. These models have previously been used to address several issues, including individual animal identification, tracking animal movement, body part recognition, and species classification. However, over the past decade, there has been a growing interest in using these models to examine the relationship between livestock behaviour and associated health problems. Several DL-based methodologies have been developed for livestock behaviour recognition, necessitating surveying and synthesising state-of-the-art. Previously, review studies were conducted in a very generic manner and did not focus on a specific problem, such as behaviour recognition. To the best of our knowledge, there is currently no review study that focuses on the use of DL specifically for livestock behaviour recognition. As a result, this systematic literature review (SLR) is being carried out. The review was performed by initially searching several popular electronic databases, resulting in 1101 publications. Further assessed through the defined selection criteria, 126 publications were shortlisted. These publications were filtered using quality criteria that resulted in the selection of 44 high-quality primary studies, which were analysed to extract the data to answer the defined research questions. According to the results, DL solved 13 behaviour recognition problems involving 44 different behaviour classes. 23 DL models and 24 networks were employed, with CNN, Faster R-CNN, YOLOv5, and YOLOv4 being the most common models, and VGG16, CSPDarknet53, GoogLeNet, ResNet101, and ResNet50 being the most popular networks. Ten different matrices were utilised for performance evaluation, with precision and accuracy being the most commonly used. Occlusion and adhesion, data imbalance, and the complex livestock environment were the most prominent challenges reported by the primary studies. Finally, potential solutions and research directions were discussed in this SLR study to aid in developing autonomous livestock behaviour recognition systems.