Assessing the reliability of camera-based identification, activity monitoring, and location in housing systems on dairy farms

Abstract Video-based livestock monitoring offers a noninvasive, cost-effective, and scalable alternative to direct human monitoring, but also to commonly used collar or ear tag devices on farms. It enables simultaneous real-time observation of multiple animals while avoiding stress and injuries from physical devices. However, single-camera systems face challenges such as blind spots and limited individual tracking, especially in barns lacking corridor layouts. These limitations can be overcome using multi-camera, multi-cow tracking (MCMCT) systems that integrate deep learning and statistical techniques to enable continuous detection, identification, activity classification, and zone location of animals in the barn, under commercial conditions. This environment is characterized by high stocking density (in m2 per cow), occlusions, and variable lighting. In this study, a commercial MCMCT system was tested over 31 d (May 2025) on 3 Holstein dairy farms in western France. Herd size ranged from 70 to 250 lactating cows and used automatic milking systems (AMS), which allowed identification of all animals when milked. Individual detection performance of this MCMCT system was then validated compared with official AMS records. A dedicated hybrid confusion matrix framework was developed to jointly assess detection and identification errors in the sequential process, allowing precise calculation of recall, precision, and F1-scores at both stages. Overall, this MCMCT system achieved over 90% detection recall and 87% to 93% precision, successfully detecting continuously more than 9 out of 10 cows daily. Identification was more challenging, with recall varying from 69% to 78% and precision above 83%, resulting in F1-scores of 79% to 82%. The performance of detection varied significantly between day and night in 2 out of 3 farms (H1 and H2), with recall rates dropping to 76% at night and exceeding 94% during peak daylight, underscoring the impact of lighting and activity patterns. Activity classification and zone location were robust, with F1-scores exceeding 87%, demonstrating the system's capacity to provide practical insights for herd management such as monitoring individual behaviors, identifying high-density zones around resources, and supporting daily management decisions. This work confirms the system's practical viability as a scalable, noninvasive monitoring solution effective under commercial farm complexities such as crowding, occlusion, and lighting variability. The integration of day–night performance analysis and the hybrid confusion matrix provide a rigorous and transparent framework for assessing system reliability, critical for deploying precision livestock farming technologies. Identification performance decreased under overcrowded conditions. Overcrowding is defined here as a surface area of less than 9 m2 per cow or less than than one cubicle per cow, as recommended by the EFSA Panel on Animal Health and Animal Welfare in 2023. The system demonstrates significant potential to support and enhance herd management, early disease detection, and animal welfare monitoring.