Date Full Report Received


Date Abstract Report Received



Primary Investigator:

Technology has been progressing at a rapid pace throughout the world in a multitude of industries. While technologies in agriculture are advancing, many of those advancements are more targeted to the crop sector than livestock as evidenced by the rapid adoption of GPS, the use of drones, and the targeted application of fertilizer within a field. However, the use of technologies for animals has been significantly lagging behind our counterparts in crops. Computer vision and machine learning have been the advancements in other industries that offer pork producers with the most promise to be adapted for precision management of pigs. To that end, we aimed to further the advancement in accuracy, precision, and speed of the NUtrack System used to monitor pigs in group housing for individual pig activity. Our objectives were 1) Determine correlation and repeatability of pig activities, behavior, and locomotion of pigs from weaning to harvest on production efficiencies and well-being, 2) Enhance algorithm and computational efficiency of the current system in efforts to reach near real-time data analysis vs retrospective analysis, and 3) Utilize visual cues and computational deep learning of activity, behavior and locomotion output to determine management deficiencies and strategies to optimize swine well-being and efficiency. To this end, 192 pigs were placed on trial at weaning and continuously videoed utilizing Lorex Security Cameras that were centered over pig pens in the nursery and finisher to capture wean to finish video for the complete 132 days post-weaning to harvest. Pigs were weighed at weaning, nursery exit, mid finisher, and final finisher weight. Ultrasound was also performed on pigs at final finisher weights for back fat and loineye area.

Video captured on-farm was exported via FTP back to the computer farm on campus for further processing using MATLAB. Briefly, a single fully-convolutional neural network was utilized to detect the location and orientation of each animal in the pen. The proposed method achieves over 99% precision and over 96% recall when detecting pigs in environments previously seen by the network during training. To evaluate the robustness of the trained network, it is also tested on environments and lighting conditions unseen in the training set, where it achieves 91% precision and 67% recall. To ensure the highest accuracy, it is suggested to train the NUtrack System on the needed environment prior to data analysis. A probabilistic tracking-by-detection method was employed to maintain tracking of individual pigs for long term identification and tracking. The long-term tracking method uses, as input, visible key points (left ear, right ear, point of shoulder and tailhead) of individual animals provided by the detector. Individual animals are also equipped with ear tags that are used by a classification network to assign unique identification to instances. The fixed points of the pigs are leveraged to create a continuous set of tracks and a forward-backward algorithm is used to assign ear-tag identification probabilities to each detected instance. Tracking
achieves real-time performance on consumer-grade hardware. Results demonstrate that the proposed method achieves an average precision and recall greater than 95% across the entire dataset. Small newly weaned pigs are more challenging to track than larger finisher pigs due to a combination of size, increased activity level, and huddling tendencies of weaned pigs. Datasets for detection and tracking were made publicly available for download at http://psrg.unl.edu/Projects/Details/12-Animal-Tracking for use by the scientific community for the advancement in swine tracking programs.

Data on individual pig activities were generated using the NUtrack System for traits such as time spent laying (sternal and lateral), standing, eating, rotations, distance traveled, and velocity when in motion. Data were analyzed for associations with production traits such as average daily gain, final weight, back fat and loineye area. Correlation across time was also analyzed. For brevity, distance traveled was the main trait examined. Distance traveled was significantly associated with average daily gain, final weight, and back fat and trended towards significant for loineye area. The greater the distance traveled by a pig reduced averaged daily gain, final weight, and backfat. Once the pigs fully transitioned to the nursery (approximately the third week in the nursery) the weekly distance traveled by a pig was correlated with the following week’s distance with correlations ranging from 0.75 to 0.92 with the exception of week 6 to week 7 when the pigs were moved and mixed in finisher pens. This indicates that activity from pigs from one week can be used as a baseline indicator for monitoring “normal” activity for each pig. Individual pig data were analyzed on pigs removed prior to the scheduled end of the trial. For pigs that were removed for lameness, removed pigs tended to show clear patterns, such as a notable reduction in distance traveled several days prior to treatment or removal by farm staff. A single pig that was otherwise noted as sound, healthy, and fast-growing exhibited sudden death. All activity traits of this pig showed no discernable deviation from normal except that it spent considerable more time sitting than any other pig in the trial. Future data similar to this should be examined as a possible indicator of twisted gut.

The NUtrack System significantly advanced due to this grant funding oppurtunity. NUtrack advanced from a system needing 7 weeks to process one week of video data on a single computer to being able to process 6 video feeds simultaneously in near real-time. At the same time, the accuracy of the system to locate, identify, and maintain identification advanced significantly as well. Data from the system is also highly meaningful for production analysis with traits identified by the system showing promise to be used by genetic companies to improve production efficienes. Furthermore, the advancements in tracking individual pig activities such that it holds enourmous promise for continued development into systems that can alert farm staff when pigs are comprimised and would benefit from intervention. Combined the program developed and data generated offers great promise in the use of computer vision and machine learning to aid farmers in improving production and well-being of swine.

Contact Benny Mote at 402-472-6033 or benny.mote@unl.edu for questions.