​Behavior Recognition of Group-ranched Cattle from Video Sequences using Deep Learning

Abstract

Background: One important indicator for the wellbeing status of livestock is their daily behavior. More often than not, daily behavior recognition involves detecting the heads or body gestures of the livestock using conventional methods or tools. To prevail over such limitations, an effective approach using deep learning is proposed in this study for cattle behavior recognition. Methods: The approach for detecting the behavior of individual cows was designed in terms of their eating, drinking, active, and inactive behaviors captured from video sequences and based on the investigation of the attributes and practicality of the state-of-the-art deep learning methods. Result: Among the four models employed, Mask R-CNN achieved average recognition accuracies of 93.34%, 88.03%, 93.51% and 93.38% for eating, drinking, active and inactive behaviors. This implied that Mask R-CNN achieved higher cow detection accuracy and speed than the remaining models with 20 fps, making the proposed approach competes favorably well with other approaches and suitable for behavior recognition of group-ranched cattle in real-time.