3 Deep Frame-based tracking on nursery pigs

Abstract

Abstract Rapid identification of morbid/injured pigs is essential for swine producers to ensure the health and well-being of each individual pig and ensure production efficiency. As such, there is a need to develop advanced technology as a means to further ensure the health and well-being of pigs. The objective of this study was to evaluate a novel computer vision, Deep-Frame – Detection and Tracking Platform (DF–DTP), for the ability to automatically identify/maintain identity and continuously track the activities of group-housed pigs. Utilizing a depth-enabled camera and multi-ellipsoid expectation maximization technology, 28 nursery pigs were continuously evaluated during the first 42 d of the nursery phase. Over the 42-d nursery period, the DF–DTP was capable of achieving a 93.7% accuracy for identifying and maintaining the identity of individual pigs. Through visual validation (10,544 observations), 642 identification errors were detected. Of the identification errors, 82% occurred when pigs were lying, 7.8% standing, 2.1% walking, 7.2% at the feeder (at the feeder), and 1.0% at the waterer. The DF–DTP was capable of a 96.2% accuracy rate for classification of an individual pig’s activity. Accuracy for classification of activities was 96.3% for walking, 96.3 for standing, 99.1 for lying, 86.4% for at the feeder, and 73.6% for at the waterer. Across the 42-d trial, the average time pigs spent 77.8% (+ 0.02) lying, 8.6% (+ 0.32) standing, 2.9% (+ 0.09) walking and traveled 943.1 meters/d (+ 195.9). Over time (d 1 – d 42), the time pigs spent lying, standing, walking, and meters/d decreased (P 0.001). As d within the nursery phase increased, time at the feeder increased (P < 0.001), there was no change (P = 0.11) for time at the waterer. Results of the study indicate that the DF–DTP is capable of accurately identifying, maintaining the identity of and continuously tracking the activity of group-housed nursery pigs.