Comparison of strategies for automatic video-based detection of piling behaviour in laying hens

This study addresses piling behaviour in laying hens, a concern for producers due to welfare issues. Piling is commonly defined as the intense aggregation of birds into dense clusters, typically occurring in large group-housing systems. This behaviour can lead to severe welfare issues including suffocation for the affected birds, particularly those at the bottom of the pile. This is a preliminary study, using video data from 4 commercial Swiss layer flocks. The long-term goal of our research is to develop a machine learning-based method to automatically detect piling behaviour using a two-step approach: first, a pre-trained convolutional neural network model (VGG-16) is used to extract features from the video data. Second, a secondary model, trained specifically to detect piling, is applied to the extracted features. The specific aim of this preliminary study is to determine which secondary modelling strategy is best suited for the accurate and efficient detection of piling behaviour. Three secondary methods are explored, namely fully connected artificial neural networks (FC-ANN), long-short term memory (LSTM) networks, and convolutional neural networks (CNN).