LAUSR

The mounting behaviour of pigs is a symptom of oestrus and is also related to their welfare. It can be recorded by surveillance video, which is a kind of streaming data. This paper proposes a pig mounting behaviour recognition algorithm by exploiting video spatial–temporal features. First, a pig detector is constructed based on Faster R–CNN to locate the pig's body, head, and tail from the image frames. Then, the distance, the overlap area and the intersection angle between two pigs in a single frame are selected as the spatial features related to mounting behaviour. The changing rate of these variables in adjacent frames is considered as the temporal feature. Data mining methods are applied to handle these features and a classifier is built based on XGBoost for distinguishing pigs' mounting and non-mounting behaviour. Finally, for the video sequences, measures such as the merging of adjacent frames and noise filtering are taken to achieve a dynamic and continuous mounting behaviour recognition algorithm. The results show that the accuracy of our pig detector is 97% and the average accuracy of the pig mounting behaviour detection in videos is 95.15%, which can be effectively applied to the mounting behaviour recognition in the video sequence.