3D imaging for trait estimation of beef and sheep carcasses

Summary

The project builds upon initial proof of concept work reported under "Value based trading system: image analysis of sheep and beef carcasses- proof of concept" (MLA project no: B.SBP.0121) where the technique for estimation of objective traits was undertaken based on 3D shape curvature descriptors.

Developing a prototype scanning rig capable of capturing data with high fidelity 3D images in abbatoir settings was a crucial objective for this project. An approach to 3D carcass construction was developed using the rotating rig with three 3D cameras via a batch optimisation framework (g2o) to compute the best fit of all data into a common representation of the carcass. A sensor fusion pipeline with dense 3D information giving more coverage over the carcass and increased robustness to sensor noise was designed. This work has further allowed an evaluation of possibilities for speeding up the up the acquisition process of the 3D scanner. The system was tested for a total of 15 days in non-sray chillers at 3 commercial abbatoirs (beef and lamb).

Alterations to the fusion method for assembling 3D models of the carcass have been undertaken to allow more robustness in the estimation framework. The curvatures of the hindquarter produced a strong relationship between lean mean yield [%] of beef carcasses on a combined dataset (n=69) acquired over 3 slaughters with a 4.05 root mean square error (RMSE) and 0.7 R2 in estimating lean meat yield (LMY).

Porting of the 3D curvature approach for estimating LMY on beef carcasses could not be applied to lamb carcasses. Beef carcasses are halved in the abattoir chain, thus, the internal surface of musculoskeletal composition of the carcass is exposed. In contrast to the internal surfaces of the beef carcass, lamb carcasses have smooth curvatures and no correspondence of curvature to LMY was established. An alternative proposition to estimate subcutaneous fat cover using Hyperspectral Imaging has been preliminarly evaluated for improving accuracy of LMY estimation. Hyperspectral Image data was transformed to use reflectance as the feature that is invariant to the respective position of camera and light source. In a laboratory setting on a dataset of 16 lamb cuts the approach produced 0.92 R2 and 0.8mm root mean square error (RMSE) in estimating fat depths up to 12mm.

Precise delineation of muscle groups from 3D data is proposed as an important refinement to imrpove the accuracy of 3D imaging. Methods for determining carcass fat, combining Hyperspectral and 3D data on the sensing rig are recommended as future work with a view to evaluating the system on whole carcasses.