Computerized identification and classification of stance phases as made by front or hind feet of walking cows based on 3-dimensional ground reaction forces

Highlights? The capturing of 3-dimensional ground reaction forces of walking cows was described. ? Computerized procedures to identify and classify stance phases were developed. ? Overall 62% of the detected stance phases were deemed valid for gait analysis. ? Human evaluation and computerized procedures agreed in 96.5% of stance phases. ? Sensitivity analysis showed that the classification criteria were robust to changes. Lameness is a frequent disorder in dairy cows and in large dairy herds manual lameness detection is a time-consuming task. This study describes a method for automatic identification of stance phases in walking cows, and their classification as made by a front or a hind foot based on ground reaction force information. Features were derived from measurements made using two parallel 3-dimensional force plates. The approach presented is based on clustering of Centre of Pressure (COP) trace points over space and time, combined with logical sequencing of stance phases based on the dynamics of quadrupedal walking. The clusters were identified as full or truncated (incomplete) stance phases furthermore the stance phases were classified as originating from a front or hind foot. Data from 370 walking trials made by 9 cows on 5 experiment days were used to test the method. Four cows were moderately lame at experimental onset. On average 5.1 stance phases per cow per trial were obtained of which 3.2 were classified as full stance phases and therefore appropriate for further gait analysis (the latter not being the scope of this study). Of the 2617 identified clusters 1844 were classified as stance phases, of these 1146 (62%) were automatically identified as full stance phases and classified as made by a front or hind foot. As intended, the procedures did not favour identification of stance phases of healthy cows over lame cows. In addition, a human observer evaluated the stance phases by visual inspection, revealing a very low discrepancy (3.5%) between manual and automated approaches. Further, a sensitivity test indicated large robustness in the automatic procedures. In conclusion, the experimental setup combined with the computerized procedures described in the present study resulted in a high number of stance phases obtained per trial. It is thus a combination which has the potential to enable unsupervised gait analysis based on data collected automatically on-farm.