Digital phenotyping of individual broilers in group-housing: detection, tracking and identification

For more than five decades, breeding companies have selected for leg health using subjective and laborious manual scoring systems. Automated phenotyping could make this process more efficient, objective, and suitable for evaluating a larger number of broilers, potentially improving the effectiveness of selection for better leg health. 
In this thesis, technologies to measure individual broiler locomotion using computer vision and sensor data fusion were developed and evaluated. Two video-based approaches were investigated: pose-estimation, which integrates with current manual walkway scoring, and multi-object tracking, which introduces a novel top-view approach to measure multiple individual broilers in group-housing.

For pose-estimation, multi-species models reduced the amount of manual annotation needed with a small decrease in performance for visually comparable species (broilers and turkeys). In addition, multi-object tracking combined with RFID sensor data allowed video tracks to be linked to individual broilers.

Overall, this thesis contributes to the video-based phenotyping of broiler locomotion and enables the phenotyping of individual broilers in group-housing.

Jan Erik Doornweerd,

The candidate of the PhD defence "Digital phenotyping of individual broilers in group-housing: detection, tracking and identification".