Tracking individual broilers on video in terms of time and distance

Area of Science:

• Animal Science
• Agricultural Engineering
• Computer Vision

Background:

• Individual broiler tracking via video offers insights into health, welfare, and performance for genetic improvement.
• Similar broiler appearances in group settings present significant tracking challenges.

Purpose of the Study:

• To analyze broiler tracking performance on video, including ID-switches, tracking time, and distance.
• To examine tracking errors such as ID-losses related to location, proximity, and kinematics.
• To establish a baseline for broiler locomotion phenotyping in group-housed settings.

Main Methods:

• Utilized a YOLOv7-tiny model for broiler detection, achieving high precision (0.99) on a test set.
• Implemented a SORT multi-object tracker evaluated on ground-truth trajectories of 13 white broilers.
• Analyzed 136 minutes of video data from broilers at 18 days of age in an experimental pen.

Main Results:

• Tracking errors were primarily associated with broiler occlusion by drinkers and close proximity (within 10 cm).
• Velocity and acceleration showed a lesser impact on tracking errors compared to spatial factors.
• ID-switches averaged 9.92 per trajectory, with tracking times and distances varying significantly.

Conclusions:

• Video tracking of group-housed broilers has potential for locomotion phenotyping but faces challenges like ID-switches.
• Addressing occlusion and proximity issues in tracking algorithms is crucial for improving accuracy.
• An external animal identification system is recommended for objective, simultaneous, and semi-continuous phenotyping.