In this paper, we demonstrate that computer vision pipelines utilising deep neural architectures are well-suited to perform automated Holstein Friesian cattle detection as well as individual identification in agriculturally relevantsetups. To the best of our knowledge, this work is the first to apply deep learning to the task of automated visual bovine identification. We show that off-the-shelf networks can perform end-to-end identification of individuals in top-down still imagery acquired from fixed cameras. We then introduce a video processing pipeline composed of standard components to efficiently process dynamic herd footage filmed by Unmanned Aerial Vehicles (UAVs). We report on these setups, as well as the context, training and evaluation of their components. We publish alongside new datasets: FriesianCattle2017 of in-barn top-down imagery, and AerialCattle2017 of outdoor cattle footage filmed bya DJI Inspire MkI UAV. We show that Friesian cattle detection and localisation can be performed robustly with an accuracy of 99.3% on this data. We evaluate individual identification exploiting coat uniqueness on 940 RGB stills taken after milking in-barn (89 individuals, accuracy = 86.1%). We also evaluate identification via a video processing pipeline on 46,430 frames originating from 34 clips (approx. 20 s length each) of UAV footage taken during grazing(23 individuals, accuracy = 98.1%). These tests suggest that, particularly when videoing small herds in uncluttered environments, an application of marker-less Friesian cattleidentification is not only feasible using standard deep learningcomponents – it appears robust enough to assist existing tagging methods.
展开▼
