Extracting the three-dimensional shape of live pigs using stereo photogrammetry

Abstract

A stereo imaging system with six high-resolution cameras (3032 × 2028 pixels) and three flash units was developed to capture the 3D shapes of live pigs. The cameras were arranged in three stereo pods, which captured the side, top and rear views of each pig. The image resolution was 0.4 mm per pixel on the surface of the pig. The system was used to capture images of 32 pigs over 14 weeks as they grew from approximately 30-80 kg. They were divided into two groups on different diets, in order to induce shape differences. Each stereo view was processed to produce a range image of the surface, and for each pig the three views were integrated to produce a complete 3D mesh. The pigs were introduced singly into a standard 3 m × 4 m pen, which was large enough to provide an unobstructed view of the pigs. Each pig was initially manoeuvred by hand into a pig-sized area in one corner where the image capture took place. The majority of pigs learned to do this with little or no human intervention. The 3D imaging system reconstructed a flat test object with an rms deviation from flatness of ±0.1 mm, and an residual curvature of no more than -3.5 × 10-5 mm-1. Depth discontinuities of 1.2 and 2.3 mm were reconstructed with an accuracy of ±0.1 mm, and a measured 451.5 mm distance parallel to the image plane was reconstructed with an accuracy of ±0.6 mm. The stereo imaging system worked successfully on the pigs by triangulating on the natural skin texture, and did not require any additional, artificial, texture to be projected. The 3D models of the pigs were qualitatively good in appearance, and locally smooth, with an rms deviation of ±0.6 mm. © 2004 Elsevier B.V. All rights reserved.