Linking video tracks and RFID identification to monitor individual broilers in group housing

Monitoring an individual broiler’s movement, interactions, and environment using video and RFID allows breeders to identify novel or indicator traits of their health, welfare, and performance, potentially improving the effectiveness of selection. However, video monitoring and identifying multiple, similar-looking, group-housed broilers is difficult. This study developed and investigated a data-fusion approach integrating video tracking with RFID identification to reconstruct individual broiler trajectories in group housing. Thirty-nine RFID-tagged broilers were recorded in a pen with thirty subfloor RFID antennas and a video camera at 18 days of age. Tracking was achieved with a YOLOv7-tiny model and a modified version of SORT. Tracks were linked to RFID-tags either through a direct approach or through the developed two-step approach consisting of location estimation through a Hidden Markov Model (HMM), with or without incorporating additional video information, and trajectory reconstruction using a greedy algorithm. Four antenna configurations were assessed: full-grid (#30), two-lines (#12), cross-pattern (#10), and two-antenna (#2). In all antenna configurations, the developed approach outperformed the direct approach, with lower average positions errors and less missing positions when evaluated against ground-truth positions [mean (SD)]: 0.021(0.10) vs. 0.48(0.54), 0.14(0.30) vs. 0.71(0.61), 0.30(0.45) vs. 0.69(0.50) and 0.85(0.59) vs. 1.03(0.60) m, respectively. The developed sensor data-fusion approach provides individual trajectories of group-housed broilers for subsequent locomotion phenotyping (such as distance walked, speed, and acceleration) by combining the spatial precision of video with the identification reliability of RFID.