Computer vision systems for monitoring hutch-housed dairy calves

Computer vision systems (CVS) have emerged as a powerful technology for animal monitoring. However, there is limited research on CVS for behavior monitoring of hutch-housed dairy calves, which account for >50% of all calf housing in the United States. The objectives of this study were to (1) develop a CVS to monitor animal location and posture of hutch-housed dairy calves; (2) compare the predictive performance of 2 deep learning algorithms (large vs. small model) for object detection that can potentially be used in edge computing systems; (3) quantify lying bouts; and (4) investigate the relationship between image-based behavior, temperature-humidity index (THI), and respiration rate (RR) of outdoor hutch-housed dairy calves. A total of 27,704 images were collected from 3 cameras every 5 min for 24 h over 20 d during the preweaning phase. Images were leveraged from a previous experiment comparing 3 hutch ventilation conditions designed in a 3 × 3 Latin square replicated 4 times (n = 12 preweaning heifer calves) housed in individual outdoor hutches. For each image, calves were spatially located (inside or outside), and their postures when outside were classified (lying or standing), resulting in 3 location and posture classes: inside, standing outside, or lying outside. We used 297 randomly selected images for training and 128 randomly selected images for testing 2 deep neural networks (YOLOv3: large and YOLOv3-tiny: small). The precision of predicting calves as inside, lying outside, or standing outside the hutch was 94.7%, 97.3%, and 95.1% for YOLOv3 and 90.1%, 86.7%, and 90.0% for YOLOv3-tiny, respectively. The recall was 96.9%, 98.3%, and 100% for YOLOv3 and 94.4%, 97.7%, and 90.0% for YOLOv3-tiny, respectively. With THI ≥69, calves showed an increased RR (56.9 vs. 64.9 breaths/min) and an increased lying interbout interval (3.48 vs. 2.75 h/interbout interval). When regressing the change in RR between ≥69 and <69 THI, calves with greater changes in RR tended to decrease total time inside (slope = -0.10) and increase total time lying outside (slope = 0.09). Overall, both small (YOLOv3-tiny) and large (YOLOv3) deep learning models performed well in tracking the location and posture of hutch-housed dairy calves during 24-h periods. Deep learning models with fewer parameters, such as YOLOv3-tiny, offer a promising solution for implementing automated edge computing applications. Our findings highlight the feasibility of CVS to monitor the position and posture of dairy calves in outdoor hutches. In turn, this CVS provides valuable insights to detect changes in calf behavior that may serve as early indicators of health and welfare concerns, particularly during periods of heat stress.