Multi-camera tracking of turkeys in large groups using instance segmentation

Aggressive feather pecking in poultry can result in poor plumage, skin damage and mortality. Feather pecking is a multidimensional problem, relating to e.g. housing conditions, feed, management and genetics. Although the trait is heritable, breeding against feather pecking is still limited due to a lack of automated phenotyping methods for large groups of birds. To develop an automated phenotyping method, there are two main challenges: (1) detecting pecking events, and (2) identifying the birds involved. Here, we focus on the latter. We set up a system based on computer vision and radio-frequency identification (RFID) to track over 300 turkeys in a pen of 120 m2. We installed 35 cameras, such that each part of the pen was covered by two cameras. We equipped each bird with a RFID tag and placed RFID antennas in the feeders for (re-)identification. Between a feather pecking event and re-identification at a RFID antenna, birds need to be tracked through computer vision. For this purpose, we here present a multi-camera tracking algorithm. First, the 35 fields of view were calibrated and transformed to a global coordinate system. Second, a Yolov7 instance segmentation model was trained to segment the turkeys in each field of view. Third, to track birds across cameras, the overlap of instance segments from neighbouring cameras was calculated. If two segments overlapped more than 60%, they were considered to be the same bird. Fourth, to track birds over time, the overlap of segments between subsequent frames was calculated. Again, if overlap was more than 60%, the segments were considered to be the same bird. Here we present results of the tracking accuracy. The current tracking time is not enough yet to bridge the time gap between a pecking event and re-identification at the feeder. However, we believe that improvement of this model will enable tracking of animals for longer time periods in the near future.