Automatic lameness detection based on consecutive 3D-video recordings

Manual locomotion scoring for lameness detection is a time-consuming and subjective procedure. Therefore, the objective of this study is to optimise the classification output of a computer vision based algorithm for automated lameness scoring. Cow gait recordings were made during four consecutive night-time milking sessions on an Israeli dairy farm, using a 3D-camera. A live on-the-spot assessed 5-point locomotion score was the reference for the automatic lameness score evaluation. A dataset of 186 cows with four automatic lameness scores and four live locomotion score repetitions was used for testing three different classification methods. The analysis of the automatic scores as independent observations led to a correct classification rate of 53.0% on a 5-point level scale. A multinomial logistic regression model based on four individual consecutive measures obtained a correct classification rate of 60.2%. When allowing a 1 unit error on the 5-point level scale, a correct classification rate of 90.9% was obtained. Strict binary classification to Lame vs. Not-Lame categories reached 81.2% correct classification rate. The use of cow individual consecutive measurements improved the correct classification rate of an automatic lameness detection system. •Three regression models used with four consecutive measurements of 186 cows.•Independent cow-observation analysis resulted in 53.0% correct classification.•Binary classification scale gave better classification results than a 5-point scale.•A tentative proposal for a tolerant classification reduced observer-related errors.•Four consecutive measurements improved classification of automatic output to 90.9%.