Turkey Gait Analysis: Predicting Expert Score With Machine Learning Based on IMU Data

In livestock production, locomotion is an important health & welfare trait. Breed4Food aims to improve locomotion through breeding and herd management using sensor technology for precision phenotyping. Usually, locomotion is scored as a one-off subjective snapshot by a human expert. Motion sensor application in poultry is largely underdeveloped. Therefore, this study aimed to predict expert locomotion scores based on inertial measurement unit (IMU) data in turkeys. In turkey breeding, breeding candidates are subject to selection on locomotion scores. Each bird is individually scored from poor (1) to good (6) by an expert whilst walking through a corridor. During this routine procedure, three IMUs were attached to each bird (N = 83) with Velcro straps, one on each leg and one on the neck. The IMU provides 3D accelerometer, gyroscope & magnetometer data. Gradient boosting was used for step recognition based on leg IMU data (F-score: 0.82, on an allowed distance of 0.2s). The steps (N=1736 after quality control) served as input for feature extraction, which were subsequently used for prediction of locomotion scores of individual steps with gradient boosting. The model had a mean per class error of 0.37 on the test set. The current approach shows promise in providing objective locomotion scoring, possibly leading to more frequent scoring or continuous scoring of locomotion. Knowledge gained could also enhance the application of motion sensor technology in other livestock species.