Phenomics

Automated Tracking of Broiler Chicken Activity with Deep Learning and Object Tracking Algorithms

---

Abstract

Activity indicates animals' health and welfare and influences growth and efficiency due to its high energetic cost. In this study, a computer vision system was developed to automatically quantify chicken activity through instance segmentation and tracking. The system employed a MaskR-CNN deep learning library to perform instance segmentation and detect all visible chickens in each frame, generating a table with the coordinates of individual birds. Subsequently, detected chickens were tracked across frames using the Simple Online and Real-Time Tracking (SORT) algorithm. Recordings were obtained at 7, 14, 21 and 28 days old from overhead cameras covering the entire surface of 12 pens, each housing 20 Ross broiler chickens. The instance segmentation model achieved a mean Average Precision (mAP) of 0.95 for pixel assignment to the chicken class in the validation dataset, indicating high segmentation accuracy. Heatmaps were generated by aggregating the number of center positions of each detection within spatial grid squares for one-hour videos. To reduce noise caused by small positional shifts when animals were stationary, two correction methods were applied: (1) filtering out displacements shorter than half the average body length (10-pixel threshold), and (2) applying Nadaraya-Watson (NW) regression smoothing with a bandwidth of 10. Animal activity was quantified as the sum of Euclidean distances between consecutive center positions of the same individual. Instantaneous speed and acceleration were calculated using positions separated by 25 frames (equivalent to 1 second). Switchbacks in movement patterns were identified by comparing the NW regression outputs with raw positional data and calculating Euclidean distances, which were then averaged per animal. Additionally, feeder and drinker visits and their durations for each video were quantified. The method was based on defining specific regions containing them and verifying whether the Intersection over Union (IoU) between the detection area and the corresponding region exceeded 95% for the drinker and the feeder. Two filters were applied to avoid overrepresented results: only visits longer than 5 seconds were considered as drinking or eating events, and if two consecutive visits occurred less than 10 seconds apart, they were counted as a single visit. Daily activity patterns showed morning peaks corresponding to increased visits to feeders and drinkers, followed by a midday resting period and a secondary increase in activity in the afternoon. Early activity consisted of longer, faster, and straighter movements, while older birds displayed more frequent switchbacks (0.31 cm at 7 days vs 0.51 cm at 28 days old), potentially indicating locomotor issues. The NW correction method effectively smoothed tracking data and provided an alternative to fixed-distance filtering. This automated computer vision approach enables reliable monitoring of broiler activity and supports behavioral studies through automatic assessment of activity budgets such as eating, drinking, and resting behaviors.

Keywords: 2026

How to Cite:

Tarres, J., Robledillo, D. & Piles, M., (2026) “Automated Tracking of Broiler Chicken Activity with Deep Learning and Object Tracking Algorithms”, World Congress on Genetics Applied to Livestock Production Digital Archive 2026(1): 2284580. doi: https://doi.org/10.31274/wcgalp.23597

Rights: 1