New traits in pig and cattle based on 3D imaging technology

In this project, the Norwegian breeding organizations, Norsvin, Geno and Tyr have, together with NTNU and The University of Auckland (NZ) developed traits for animal breeding, based on image technology. The project has partly aimed to replace manual recordings of traditional traits within livestock breeding with camera-based solutions, and partly worked on developing novel traits from images and videos. The project has been divided in four work packages (WPs). WP1: Develop 3D camera units with RFID connection. In this work package we have developed prototypes for hand-held camera unit for acquisition of 3D images of cow udders, and dataset of udders and bull’s scrotum were collected, using these cameras. In addition, similar cameras were mounted to automatically record both size and activity phenotypes of group housed animals. However, with the rapid development of machine learning (ML) algorithms for analyzing color images, 3D cameras were partly replaced by normal surveillance cameras. WP2: Tracking of animals in the pen. To be able to breed for a healthy, well behaving pig, we like to know the position of each individual pig continuously, in addition to monitoring their activity. Thus, the pigs need to be uniquely identifiable, and all the pigs wear ear tags with RFID. These ear tags are detected in the feeder station when an animal is entering for eating. However, the pigs are very similar, and occluding scenes occur, which might lead to losing the tracks of animals. To be able to reidentify the pigs outside the feeder, we have also added unique colored ear tags. In addition to assuring quality and connecting different data sources together, this task has involved development of datasets for training of ML algorithms for detection of pigs in the pen. Further, we have tested different algorithms for detection, identification, and tracking. However, some research is still to be done, and another IPN-project (Genes2Behave, project number 321409) will continue these R&D activities. WP3: Modelling the shape based on 3D images. Images and videos are complex data sources and extracting the same relevant object across different images is challenging. Animals are moving and this require robust algorithms that could manage moving, rotating, and missing objects. In this work package, we have been using different techniques to classify, segment and align objects to generate datasets independent of the positioning of the animal in the scene. WP4: Develop new traits from complex measures. To be able to create genetic gain, individual data needs to be summarized into one or a few scalar values. For some traits, like weight, this is trivial, but for other traits, like the shape of a complex surface e.g. a cow udder, this process is challenging. We have tested methods to extract multi-dimensional shape variation into a few interpretable/relevant traits for breeding. Further, we have developed a computational pipeline to automatize the process of building genetic/statistical models for new traits, and we have organized new data into quality-controlled databases for routine genomic evaluations. Some of the results from the project are already implemented in the breeding program, and they will give improved genetic gain, with positive implications for both breeding organizations, farmers, and consumers. Dissemination of project results has been performed widely, both on international conferences, national magazines, podcasts and several peer reviewed research publications. In addition, we have gathered knowledge about technology and animals, and this will be used to create more sustainable breeding goals in the future.

During the project, the project group have gathered knowledge about how to use camere technology to automatize measurements of animals and how to develop computational pipelines for storage and processing image data for breeding purposes. Thus, the project has given the researchers new tools and knowledge that would speed up future research and built some strong collaboration between teams. This has founded other research proposals, whereof 4 are funded (“3D shape analysis of cattle to improve data collection”, NFR project number 310239, “Genes2Behave”, NFR project number 321409, Breeding for better cardiovascular and respiratory function of pigs - “A heart for pigs”, NFR project number 332120 and Meat Quality Evaluation using Non-Invasive Advanced Imaging Technologies and Computer Vision, RFFI project number 337814. Further, some traits and models developed through the project has strongly contributed to improved selection accuracy within the breeding programmes of both cattle and pigs. For example, by implementating the new conformation models on Norwegian red cattle, the genetic gain increased by about 25%. This gives a huge value for the farmers and society as whole. In addition, it improves the competitiveness of GENO on the international market. Similarly, by implementing the new carcass composition AI model, Norsvin increase the genetic gain for carcass traits, which again lead to better economics for swine farmers and improved competitiveness of the company. Being in the front when it comes to utilization of new technology and developing new traits related to health and animal welfare, makes also the companies more visible in the community. This makes the breeding organizations attractive, both as research partners and when hiring people for permanent positions.

In this project, we want to improve the production, functionality and longevity of livestock animals, by developing objective registrations of new traits from images and videos. Increased precision and extent of data recordings improves the genetic gain of pigs and cattle and will enable an increment of sustainable food production based on Norwegian resources - throughout the country As new tools within genomics has led to an increase in the precision of breeding, there is in parallel a high need for more precise registrations of the traits we want to improve on our livestock animals. Historically, scoring of an animals locomotion and conformation have been scored subjectively by trained technicians in both pigs and cattle. This is labour intensive and leads to high costs as well as insufficient quality of data. Weighing animals in the field is expensive and associated with a lot of hard manual work. Increased precision and more data, due to more simple registrations, will highly improve the accuracy of selection of the best animals in the breeding program. In this project, we are going to develop 3D camera prototypes for measuring animals and utilize computer vision to enable cost-effective precise measurements of livestock animals. The images/videos will be directly linked to the ID of the animal, using radio frequency identification (RFID). Further, we will develop effective automatic algorithms to get out registrations on single animals from videos containing many animals in a pen and model the surfaces, describing the shape of the animals. This will found the basis for development of new traits that can be implemented in the breeding programme for pigs, dairy cattle and beef cattle.