Automatic detection of locomotor play in young pigs: A proof of concept. (May 2023)

• Record Type: Journal Article
• Authors: Larsen, Mona L.V.; Wang, Meiqing; Willems, Sam; Liu, Dong; Norton, Tomas
• Is Part Of: Biosystems engineering. Volume 229(2023)
• Page Start: 154
• Page End: 166
• Publication Date: 2023-05
• Subjects: Technology -- Precision Livestock Farming -- Computer vision -- Gaussian Mixture Model -- Animal welfare -- Animal behaviour; Bioengineering -- Periodicals; Agricultural engineering -- Periodicals; Biological systems -- Periodicals; Génie rural -- Périodiques; Systèmes biologiques -- Périodiques; 631
• DOI: 10.1016/j.biosystemseng.2023.03.006
• Languages: English
• ISSNs: 1537-5110

Abstract

Abstract : Play behaviour is considered an indicator of animal welfare in young pigs. However, as play behaviour events are short-lasting and occur sporadically, continuous monitoring is necessary. This study presents a first attempt at automatic detection of locomotor play behaviour in young pigs from video by classifying locomotor play from other solitary behaviours including standing, walking, and running. Two methods were developed, compared, and sequentially combined: (1) a less computational heavy method utilising the Gaussian Mixture Model for quantification of movement combined with the calculation of contour features and standard machine learning classifiers ( FEATURES ); (2) a computational heavy method utilising a deep learning classifier taking both spatial and temporal features into account ( DEEP ). The DEEP classifier outperformed the FEATURES classifier and obtained values of internal validation recall, precision, and specificity of 94%, 88% and 96%, respectively. When combining the two classification methods, almost similar performance was retained, whilst 44% of the other behaviours were correctly classified without the need for deep learning methods. The combination thereby decreased the computational power needed to run the algorithm. Thus, locomotor play can be automatically detected in young pigs and the combination of a less computational heavy method with a deep learning method can reduce the computational requirements for the classification and detection of complex behaviours. Future work should focus on the segmentation of single pigs during high-speed activity in order to enable the play detection algorithm to work in real-life settings. Highlights: Locomotor play in young pig can be classified from other solitary behaviour. Both spatial and temporal features needed to detect the complex behaviour. Recall, precision, and specificity of 94%, 88% and 96%, respectively. Combining methods have potential to decrease computational footprint. Remote tracking may be needed to segment playing pigs.