Frontmatec BCC-3 beef classification system study and installation in Australia beef industry

Summary

Frontmatec have designed and built a 360° 3D classification system that was designed for European grading system (EUROP Score) for beef, with the potential of estimating primal weights. This project was to build on the work already undertaken by Frontmatec by designing and building an algorithm that will suit the Australian Beef processing industry and if successful, would be followed by a system installation in Australia. This project was discontinued due to program priorities.

The following provides a summary of project deliverables:

Milestone 1

Describes the optimization of the image analysis software required in Australia to achieve a throughput of 520 half carcasses/hour. The average time for image acquisition and analysis required to produce a classification result initially took approximately 213 seconds/carcass.

Four approaches were used to increase throughput;

1. Software optimization (Source code and analysis steps),
2. Software parallelization (multiple cores),
3. Hardware optimization (Processors and chipsets) and
4. Hardware parallelization (multiple computers). The initial benchmarks were performed on a Dell R730 server platform.

It became clear that the limited number of cores available would not support the massively parallelized structure needed to reduce processing time. Therefore, a Dell R740 equipped with state-of-the-art Xeon Gold Processors was chosen. The use of the R740 resulted in significant improvement of parallel processing, which reduced processing time substantially. The analysis time improved to 20.7 seconds on average, indicating that a total of three R740 analysis computers will be needed to fulfil the line speed requirements in Australia. In conclusion, a line speed of 520 half carcasses/hour was achieved by optimizing software and hardware, and introducing a topology supporting multiple analysis computers.

Milestone 2

Describes a new method of extrinsic calibration of BCC-3 geometry and the improvement in precision thereof. The calibration method is based on a double-sided chessboard placed at the centre of the BCC-3, which enables determination of individual camera position and rotation in a global coordinate system.

The calibration target also supports colour calibration, which is outside the scope of this report.

A number of tests were performed to assert the performance of the calibration software. Some of the tests were meant to determine the stability and response of the software when varying input parameters or when under imperfect conditions. We concluded that the software very reliably creates calibrations of high quality and reports on errors when certain parameters are registered to be out of specifications. The old and new calibration methods were benchmarked by performing multiple calibrations under varying conditions and evaluating the effect on distance to the mean centre of the scene.

Milestone 3

A half bovine carcass is an object with a complex shape that is not globally convex. A fundamental consequence of multi-view stereo 3D reconstruction on such an object is that there can be areas of the surface that are occluded by the object itself. A half bovine carcass is no exception. Additionally, the density of reconstructed points depends on the surface texture and the combined geometry of cameras and carcass. Thus, the BCC-3 must employ algorithms that fill in the missing information in order to calculate important properties such as the total volume and, by extension, the volumes and weights of primal cuts.

The performance of the method implemented is reported here and it is shown that it reliably and accurately solves the task. In this report it is shown how its performance scales in terms of magnitudes of various types of imperfections in the reconstructed point cloud.

By computing a watertight triangular mesh, a variety of global measures become computationally available. Quantities like the volume, surface area, volumetric centre of mass, or inertial axes become well-defined quantities. The evaluation of the success in meeting the milestone is based on computing the predicted volume on a PC that has artificially been made less than perfect. The volume is an excellent figure-of-merit because it relates in a very direct way to the weight - and hence to the value - of the carcass. Furthermore, it is necessary to predict the global volume accurately when computing more complicated measures such as the volumes/weights of primal cuts. Hence, accurate prediction of the carcass volume must be a hallmark of a good method for closing holes in the PC.

Milestone 4

The objective of milestone 4 was to build a BCC-3 image analysis tool that can identify beef carcass descriptors which can be used to predict the weight of primal cuts. BCC-3 images were acquired on warm carcasses. The weight of the primal pistol cut (rib + round) and 3 commercial cuts of the round (inside without cap, knuckle and rump) were recorded on 86 Danish beef carcasses during deboning.  

BCC-3 carcass descriptors were developed and include moments of inertia, area and circumference of largest round cross-section, volume and area of carcass. It was concluded that BCC-3 can predict the primal weight of the pistol cut and the three customised cuts in the round. Prediction accuracy and prediction error did not significantly improve by including the HCW in combination with the BCC-3 carcass descriptors. This indicates that the BCC-3 system successfully predicts the primal weight without external input.