From Detection to Identification and Phenotyping of Free-range Cattle with Aerial Data from UAVs using Deep Learning technologies

The main goal of this research, belonging to Precision Livestock Farming (PLF), is to improve the generation efficiency of essential information on herding free-range beef cattle on pasture for herders, therefore improving grazing management efficiency and then increasing individual meat output. The research purposes to employ automatic and non-intrusive 3D computer vision based on the high performance of 3D convolutional neural networks with the omniscient view angle from flying unmanned aerial vehicles (UAVs). The background is that more attention needs to pay to individuals accurately, efficiently and humanely in PLF. The reality is that conventional artificial markdowns such as branding and/or ear tags can cause physical injuries, while the electrical methods like harmful and expensive injectable transponders and short communication-distance wearable devices are hardly adaptable for wide-range applications, and the current state-of-the-art of 2D computer vision is susceptible to the environment such as illumination, background noise and intensity. So, this study aims to introduce innovative 3D computer vision, which extends the advantages of 2D computer vision and complements the disadvantages of that, for the sake of precision management in pasture-based freeranging systems from the number of cattle herds, to the identification and live weight estimation, to pursuance (individual spot searching). In summary, UAVs with high spatial-resolution RGB cameras and LiDAR will be employed to collect temporally raw data, multiple data pre-processing techniques such as programming to align training data with ground truth, segmenting point clouds plan to be applied before data are feeding into deep learning models, and elaborate recurrent 3D Recurrent Convolutional Neural Networks (RCNNs) with transfer learning will be iteratively trained. As a result, Graphics Processing Unit (GPU) being fitted into unmanned aerial vehicles can capitalize on these well-trained models for the prediction of animal traits in realtime. 7.