Augmenting Cattle Tracking Efficiency Through Monocular Depth Estimation

We present a method for 3D cattle tracking and inter-camera pose transformation using depth information from monocular depth estimation with deep networks. Camerabased animal monitoring offers a minimally invasive and easily adaptable solution for tracking and welfare monitoring, relying solely on commercial RGB camera systems. However, environmental factors and inter-animal occlusion often hinder tracking efficacy and consistency. To address these challenges, we developed a pipeline to extract 3D point cloud data of individual cows in a straw-bedded calving yard environment, generating quasi-3D bounding boxes (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$x, y, z$</tex>, height, width, <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\theta$</tex>), where <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\theta$</tex> is the polar angle. We then estimate the camera system extrinsic parameters by minimising the rotation, translation, and scale discrepancies between the apparent motion of animals across different frames of reference. This approach demonstrates a strong agreement between the 3D centroids of tracked animals in motion. Our work advances the development of algorithmic occlusion handling and object handover techniques in multicamera systems, particularly pertinent to the high-occlusion, low-locomotion scenario of animals within barn environments.