Early and non-intrusive lameness detection in dairy cows using 3-dimensional video

• Published in: Biosystems engineering Vol. 153; pp. 63 - 69
• Main Authors: Abdul Jabbar, K., Hansen, Mark F., Smith, Melvyn L., Smith, Lyndon N.
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.01.2017
• Subjects: 3D computer vision; Animals; Automation; Classification; Dairy farming; Early lameness detection; Gait asymmetry; Locomotion; Locomotion analysis; Support vector machines; Thresholds; Walking; Gait asymmetry; Locomotion analysis; 3D computer vision; Early lameness detection
• ISSN: 1537-5110; 1537-5129
• DOI: 10.1016/j.biosystemseng.2016.09.017

Lameness is a major issue in dairy herds and its early and automated detection offers animal welfare benefits together with potentially high commercial savings for farmers. Current advancements in automated detection have not achieved a sensitive measure for classifying early lameness. A novel proxy for lameness using 3-dimensional (3D) depth video data to analyse the animal's gait asymmetry is introduced. This dynamic proxy is derived from the height variations in the hip joint during walking. The video capture setup is completely covert and it facilitates an automated process. The animals are recorded using an overhead 3D depth camera as they walk freely in single file after the milking session. A 3D depth image of the cow's body is used to automatically track key regions such as the hooks and the spine. The height movements are calculated from these regions to form the locomotion signals of this study, which are analysed using a Hilbert transform. Our results using a 1–5 locomotion scoring (LS) system on 22 Holstein Friesian dairy cows, a threshold could be identified between LS 1 and 2 (and above). This boundary is important as it represents the earliest point in time at which a cow is considered lame, and its early detection could improve intervention outcome thereby minimising losses and reducing animal suffering. Using a linear Support Vector Machine (SVM) binary classification model, the threshold achieved an accuracy of 95.7% with a 100% sensitivity (detecting lame cows) and 75% specificity (detecting non-lame cows). •Completely covert and non-intrusive 3D video system to detect lameness in cows.•A novel proxy of symmetry used from height variation of cow hooks.•Automated feature extraction of spine and hooks using shape index and curvedness.•A threshold was identified to make an early stage lameness classification.•Using a linear SVM model, overall binary classification accuracy of 95.7% achieved.