Smartphone-based Automated Livestock Welfare Monitoring System using Computer Vision

Thesis Abstract

The welfare of intensively reared livestock is increasingly compromised by delayed identification of stress, illness, and inadequate environmental conditions, which can lead to reduced productivity, increased mortality, and economic losses. This study addresses the need for scalable, rapid, and objective welfare assessment by developing a smartphone-based automated monitoring system that leverages computer vision to quantify behavioral and physiological indicators in farm animals. The aim is to design, implement, and validate a mobile ICT tool capable of real-time welfare assessment across dairy cows and beef cattle under typical commercial conditions. Specific objectives include (1) to develop a computer-vision pipeline that extracts salient welfare indicators (posture, gait, grazing/resting patterns, heat stress cues, and ocular/respiratory signs) from field-acquired video frames; (2) to integrate the pipeline with a user-friendly smartphone interface for farmers and veterinarians; (3) to evaluate the system’s diagnostic accuracy against ground-truth veterinary assessments and sensor-based proxies; (4) to assess the system’s operational performance in terms of latency, battery consumption, and robustness to lighting and occlusion; and (5) to model the relationship between automated welfare scores and production metrics (milking yield, feed conversion ratio, somatic cell count). The study adopts a mixed-methods approach underpinned by the Theory of Planned Behavior to understand adoption factors among producers, and the Ecological Modernization Theory to contextualize technology-enabled welfare improvements. The research design combines a developmental, action-research cycle with a validation study in two commercial farms over a 12-month period. The population includes lactating dairy cows and finishing beef cattle housed in free-stall and pasture-based systems, with a targeted sample of 480 animals across 8 herds. Data collection instruments comprise (a) a smartphone-based recording app and embedded computer-vision module for real-time video capture and processing; (b) ground-truth welfare assessments by qualified veterinarians using standardized welfare scoring sheets; (c) conventional sensor data (accelerometers, temperature-humidity trackers) for cross-validation; and (d) production records (milk yield, feed intake, weight gain, somatic cell count). Data analysis follows a sequential framework first, computer-vision outputs are validated against expert scores using Bland-Altman analysis and intraclass correlation coefficients to establish agreement; second, diagnostic performance is evaluated with receiver operating characteristic (ROC) curves and area under the curve (AUC) metrics, complemented by confusion matrices for distress and illness detection. Multivariate regression analyses and generalized linear models examine associations between automated welfare scores and production outcomes, controlling for farm-level fixed effects. A repeated-measures ANOVA investigates temporal changes in welfare indicators in response to management interventions. The study also employs thematic analysis of farmer interviews to elucidate perceived usability, perceived usefulness, and barriers to adoption. Expected findings include robust correlation between computer-vision-derived welfare indices and veterinary assessments (ICC > 0.80), high diagnostic accuracy for acute distress events (AUC > 0.85), and meaningful associations between welfare status and production metrics, such as reduced somatic cell counts and improved feed efficiency in animals identified and managed through the system. The contribution to knowledge lies in (i) empirically demonstrating the feasibility and accuracy of a mobile, vision-based welfare monitoring solution in real-world farming contexts, (ii) providing a scalable framework for low-cost welfare surveillance that minimizes manual workload, and (iii) offering an empirically grounded model of technology acceptance among livestock producers. The study concludes that smartphone-based automated welfare monitoring can significantly augment early detection of welfare issues, support data-driven management decisions, and enhance overall animal well-being while delivering measurable productivity benefits. Recommendations include integrating the system with existing farm-management software, conducting longitudinal cost-benefit analyses, exploring augmentation with infrared thermography for improved thermal stress detection, and expanding validation to additional species and housing systems to enhance generalizability.