Automated Detection of Flock Health, Behaviour and Weight with Machine Vision in Commercial Broiler Sheds

Visual assessments for health and welfare of broiler flocks are regularly performed by farm staff walking through the shed. However, common welfare indicators like footpad dermatitis and hockburn require manual handling of the flock and an automated measurement is desirable. There is potential for automated camera monitoring to complement human inspections for detection of health and welfare conditions and perform additional flock monitoring tasks. A proof-of-concept machine vision system has been developed for the detection of the health, welfare and weight of meat chicken flocks in commercial sheds using low-cost video cameras.

Video imagery was collected for six flocks in four commercial sheds using a camera (Galaxy Mini J105, Samsung, South Korea) that continuously recorded video clips for an average of 45 days per flock. The collected video was used for development of novel image analysis algorithms using the Open Computer Vision library (2015). Image analysis development focused on detecting attributes of single chickens such that flock motion could be expressed in terms of number of chickens. Automated chicken counting was achieved with root mean square error (RMSE; Microsoft Corporation, 2018) of 5%, compared to manually counted chickens in images, which was considered acceptable for quantifying flock motion.

Image analysis steps for the detection of additional flock parameters enabled identification of a novel metric which was highly correlated (R 2 0.7 to 0.9) with the flock health prediction technique of Dawkins’ optical flow (Dawkins et al., 2012), which has reported success with predicting gait and footpad dermatitis from flock motion expressed in pixels. The novel metric successfully ranked flocks as ranked by Dawkins’ optical flow and as compared with flock health data provided by the farm. The finding provides insight into the flock behavioural attributes detected by Dawkins’ optical flow for health prediction, which from previous studies was not known with certainty (Dawkins et al., 2013).

A novel image analysis algorithm for weight estimation achieved overall RMSE of 55 g and relative RMSE within 5% when compared with weekly weight measurements provided by the farm using standard commercial processes. The technique used imagery from a monocular low-cost video camera and has potential to provide automated and unobtrusive daily flock weight information for the farmer.

Automated classification of multiple bird behaviours including eating, pecking and sitting was achieved in commercial shed conditions with an average of 78% accuracy, which is considered within acceptable tolerance levels due to there being no other existing technologies to perform automated behaviour monitoring in commercial environments. Potentially, automated classification and quantification of behaviours provides an objective measure for assessing flock welfare, or a real-time sensor input for a climate controller inside a shed housing system. Commercialisation opportunities for the proof-of-concept machine vision technologies are currently being explored, and further research should perform extended on-farm evaluations.