Field performance evaluation of a computer vision-based system for the automatic detection of dairy cow behaviour housed in free-stall barns

In the literature dairy cow behaviours are generally analysed by skilled operators who observe the herd directly in the breeding environment or within digital images acquired from time-lapse video recordings. Since this method could suffer from limitations widely discussed in literature, the research activity described in this paper aimed at developing a computer-vision based system (CVBS) for the automatic detection of dairy cow behaviours within digital images. The system was composed of a multi-camera video-recording system and a software component which included four cow behaviour detectors (i.e., cow lying, cow standing, cow feeding, and cow perching detectors) modelled by using the Viola-Jones's algorithm. The number of true positives (TP), false negatives (FN) and false positives (FP) produced by the CVBS was obtained from a validation phase which involved the work of an operator who checked the CVBS detection results. Next, the field performances of the CVBS were evaluated by using the following accuracy indices: sensitivity index, i.e., the ratio between TP and the number of cows really existing in the monitored area; branching factor (BF), i.e., the ratio between FP and TP; miss factor (MF), i.e., the ratio between FN and TP. The highest value of the sensitivity index was recorded for the lying behaviour detector (0.92), whereas the lowest one (0.86) for the standing behaviour detector. The minimum value of BF and MF were obtained for the cow lying behaviour detector (0.08 and 0.09, respectively), whereas the maximum values were recorded for the standing behaviour detector (0.17 for both).