First step to remote detection of locomotor play in young pigs: a preliminary study

TY - GEN

T1 - First step to remote detection of locomotor play in young pigs

T2 - 10th European Conference on Precision Livestock Farming

AU - Larsen, Mona Lilian Vestbjerg

AU - Wang, M.

AU - Willems, Sara

AU - Liu, D

AU - Norton, Tomas

N1 - Conference code: 10

PY - 2022/9

Y1 - 2022/9

N2 - The performance of play behaviour may be a measure of animal well-being in young pigs. However, play events are usually short in duration and occur sporadically in time. Thus, the use of play behaviour as a measure of animal welfare would benefit greatly from remote, real-time monitoring. We here present a preliminary study into remote detection of play behaviour with computer vision in young pigs, investigating the possibility of recognising locomotor play from other forward movement, using methods requiring low computing power. Two pens of pigs were observed between 1800-2200 h the day after weaning and labelled for events of locomotor play, running and walking. To quantify behavioural intensity, Gaussian mixture models were applied to the videos producing a low and high activity video stream with the background subtracted, from which 10 and 15 frames movement maps were produced for the events of behaviour. From the movement maps, contours were detected, and six contour features were calculated as behavioural descriptor. 1000 behavioural descriptors per class were used as inputs to several classification algorithms of which random forest had the highest accuracy (0.799). Only 76% and 72% of play and running contours were correctly classified, whereas recall was 91% for walking contours. Based on this, it was decided that stronger classifiers would be needed to recognise play from other high-activity forward movements such as running, and that the current method could function as initial classifier of locomotor play to reduce the data amount to be classified with more sophisticated models.

AB - The performance of play behaviour may be a measure of animal well-being in young pigs. However, play events are usually short in duration and occur sporadically in time. Thus, the use of play behaviour as a measure of animal welfare would benefit greatly from remote, real-time monitoring. We here present a preliminary study into remote detection of play behaviour with computer vision in young pigs, investigating the possibility of recognising locomotor play from other forward movement, using methods requiring low computing power. Two pens of pigs were observed between 1800-2200 h the day after weaning and labelled for events of locomotor play, running and walking. To quantify behavioural intensity, Gaussian mixture models were applied to the videos producing a low and high activity video stream with the background subtracted, from which 10 and 15 frames movement maps were produced for the events of behaviour. From the movement maps, contours were detected, and six contour features were calculated as behavioural descriptor. 1000 behavioural descriptors per class were used as inputs to several classification algorithms of which random forest had the highest accuracy (0.799). Only 76% and 72% of play and running contours were correctly classified, whereas recall was 91% for walking contours. Based on this, it was decided that stronger classifiers would be needed to recognise play from other high-activity forward movements such as running, and that the current method could function as initial classifier of locomotor play to reduce the data amount to be classified with more sophisticated models.

KW - Gaussian mixture model

KW - animal behaviour

KW - animal welfare

KW - contour detection

KW - machine learning

KW - movement map

UR - http://www.scopus.com/inward/record.url?scp=85172295637&partnerID=8YFLogxK

M3 - Article in proceedings

SN - 978-83-965360-0-6

SP - 91

EP - 98

BT - Precision Livestock Farming '22

A2 - Berckmans, M

A2 - Oczak, M

A2 - Iwersen, M

A2 - Wagener, K

PB - Vetmeduni

Y2 - 29 August 2022 through 2 September 2022

ER -