Deep Learning Methods for Tracking the Locomotion of Individual Chickens

Abstract

Poultry locomotion is an important indicator of animal health, welfare, and productivity. Traditional methodologies such as manual observation or the use of wearable devices encounter significant challenges, including potential stress induction and behavioral alteration in animals. This research introduced an innovative approach that employs an enhanced track anything model (TAM) to track chickens in various experimental settings for locomotion analysis. Utilizing a dataset comprising both dyed and undyed broilers and layers, the TAM model was adapted and rigorously evaluated for its capability in non-intrusively tracking and analyzing poultry movement by intersection over union (mIoU) and the root mean square error (RMSE). The findings underscore TAM’s superior segmentation and tracking capabilities, particularly its exemplary performance against other state-of-the-art models, such as YOLO (you only look once) models of YOLOv5 and YOLOv8, and its high mIoU values (93.12%) across diverse chicken categories. Moreover, the model demonstrated notable accuracy in speed detection, as evidenced by an RMSE value of 0.02 m/s, offering a technologically advanced, consistent, and non-intrusive method for tracking and estimating the locomotion speed of chickens. This research not only substantiates TAM as a potent tool for detailed poultry behavior analysis and monitoring but also illuminates its potential applicability in broader livestock monitoring scenarios, thereby contributing to the enhancement of animal welfare and management in poultry farming through automated, non-intrusive monitoring and analysis.