A computer vision-based approach for respiration rate monitoring of group housed pigs

Abstract

In recent years, respiration rate (RR) monitoring using video data has been explored by researchers with relatively good success. However, the approaches used so far require the manual identification of the region of interest (ROI) in the image of the animal. When monitoring farm animals, typically housed in groups, such manual actions would entail an excessive time investment and increase the level of subjectivity in the application of the method. The aim of this study was to design a RR monitoring system targeted at group-housed animal applications. The developed system first selected video clips where pigs were in a resting status. Then an oriented object detector was used to detect each animal and select the ROI without manual intervention. Finally the RR is estimated by analyzing the time-varying features extracted from the ROI. Videos of group-housed pigs were collected to test the method, in which 5 pigs were included. Four pigs wore an ECG belt around the abdomen to collect the gold standard (GS) RR measures while a control pig did not wear a belt, with the GS for the control pig obtained through manually observation. The comparison between RR obtained by the computer vision (CV) method and the GS showed good agreement with a mean absolute error of 2.38 breath per minute (bpm) in the 4 pigs wearing belts and 1.72 bpm in the control pig, a root mean square error of 3.46 bpm in the 4 pigs wearing belts and 2.26 bpm in the control pig, and a correlation coefficient of 0.92 in the 4 pigs wearing belts and 0.95 in the control pig.