Enhancing Consistency and Efficiency of Tracking Multiple Individual Broiler Chickens in Group Settings
Funded by the Cobb Research Initiative, researchers at the University of Georgia published their development of a digital phenotyping system aimed at tracking broiler chickens. Digital phenotyping systems use a variety of technology to capture observable and measurable characteristics. Technology including cameras, thermal imaging, scales, and radio frequency tracking devices have been employed in digital phenotyping. The digitally captured characteristics can be converted into data points, which are used to develop models and create base lines.
In broiler production systems, digital phenotyping systems are being developed for functions such as real-time monitoring and integrating with other systems for comprehensive flock surveillance. The technology is still in early stages of development and researchers point out tracking individual birds is very challenging. Its difficult to track individuals in a house setting due to obstructions such as feeders and drinking lines, overlapping individuals, similar appearances among broiler chickens, motion blur from fast movement, inconsistent illuminations and backgrounds, and low resolutions.
With these challenges in mind, the researchers set out to create a system that could track individual broiler chickens in a flock and maintain their IDs consistently and efficiently. First they had to overcome a few optical limitations of the camera system. The team noted that there were both radial and tangential distortions. Radial distortions gave a “fish bowl” view where birds become smaller than actual size as the image progresses from the center. Tangential distortion occurs because the image-taking lens is not aligned perfectly parallel to the imaging plane, making images appear closer than they actually are. To correct these distortions, the researchers used a checker board that was strategically placed with known distances and sizes for calibration and corrections. In this way, the errors were reduced to less than ½ a pixel which produced a high-quality calibration.
Camera image illustrating distortion (left) and after calibration (right).
The You Only Look Once (YOLO) algorithm, a widely adopted real-time object detection model, was employed to detect an track chickens. The “out-of-the-box” version of YOLO performed poorly on its pre-trained data sets. So, the researchers created a custom trained model from more than 9,000 frames captured from videos of chickens 2 to 7 weeks old. In total, the training set included nearly 340,000 images of chickens in their house setting. After training, the model went from <40% accurate to nearly 97% accuracy. Moreover, with a little “pruning” to remove background in the image that wasn’t needed, they were able to substantially increase the processing time and enhance computing efficiency.
As noted, obstructions can block the cameras from seeing chickens causing a chicken to disappear in one frame and reappear in subsequent frames. The absence can create issues and cause the tracker to misidentify the chicken. To solve this issue, the researchers constructed a second algorithm. This re-identification and tracking algorithm used machine learning classifiers, which identified chickens using defined information. The researchers trained the classifiers with detailed images about the birds and how they moved. These classifiers are computationally expensive. So this second algorithm was only used if an individual chicken was missed between frames to validate the identity of the chicken.
By putting all components together, the integrated model system could track individual broiler chickens in feeder, drinker, and open areas with high accuracy and fast processing times. The researchers are continuing to work on this system with the goal of achieving a robust real-time system that can be scaled to use in broiler houses industry wide.
Read the full article here:
https://www.sciencedirect.com/science/article/pii/S2772375525006392
Article citation:
Li, Guoming, Sai Akshitha Reddy Kota, Tongshuai Liu, Oluwadamilola Moyin Oso, Venkat Umesh Chandra Bodempudi, Mahtab Saeidifar, Ehsan Asali et al. “Enhancing Consistency and Efficiency of Tracking Multiple Individual Broiler Chickens in Group Settings.” Smart Agricultural Technology (2025): 101408.
