Page - 121 - in Proceedings of the OAGM&ARW Joint Workshop - Vision, Automation and Robotics

B. Classification systems The current application employs two different classifica- tion subsystems. Both systems provide a satisfactory solution for the non-ferrous metal classification. Currently linear classifiers with several features are used, but in the future neuronal networks will be trained. The LIBS system measures the chemical composition of the particlesandseparates themintocastandwroughtaluminium categories and into selected aluminium and magnesium alloys.[1] The EMTS system on the other hand measures the electrical conductivity of the particles to separate the fragments into aluminium, copper and brass categories.[2] III. RESULTS A. Timing Performance Due to the described settings, the maximum computation time is 50ms for one subframe. Two types of evaluations were realized with separated aluminium particles on a 1000mm x 400mm belt area. The first test (small covering, SC) simulated the real coverage with 139g particles. In the second test the belt was fully covered (full covering, FC). Three test sets of particles were used. The particles were placed on the belt and processed three times in the same arrangement. The arrangement itself was varied three times, such that nine timing tests for every particle size were done. Table I shows the test conditions and the calculation time for one subframe. As can be seen, the maximum computation time is 24.5ms for a full covered belt with 55 large samples. TABLE I MAXIMUM COMPUTATION TIME FOR FEATURE CALCULATION ON ONE SUBFRAME (200 3D PROFILES) Test set Small Medium Large Sample size [mm] 9x9 20x20 30x30 Test size SC FC SC FC SC FC Sample count 71 355 29 145 11 55 Max. time [ms] 20.91 23.33 21.26 24.1 11.78 24.5 B. Accuracy To verify the accuracy of the Image Analysis System objects with defined dimensions were used (e.g. eurocents and washers) as well as real particles. Due to the complex real particle shapes no ground truth for heights, areas and diameters were available. Therefore, just the positions and recognition rates of real particles were tested. The feature calculation accuracy is higher than 95% and almost every single particle can be detected (see Table II). Nearly all coins were detected correctly. Only one misdetec- tion was observed since two 2 eurocents were not separated on the belt. The small deviation of the area can be explained by the fact that reflections on the edge lead to overestimate the real object size. Thus, the height is measured also on edge regions with height values produced by reflections. The height is furthermore influenced by the shape of the coins. Only the edge has full height, whereas the rest of the surface TABLE II ACCURACY OF THE IMAGE ANALYSIS SYSTEM IN % Sample Height Area Diameter Found 1 eurocent 96.74 98.13 99.89 100 2 eurocent 98.91 96.68 99.87 99 5 eurocent 99.14 99.29 99.29 100 10 eurocent 98.25 99.10 99.10 100 20 eurocent 99.67 98.33 99.27 100 50 eurocent 98.97 98.97 99.58 100 Washer 16 98.94 95.18 97.12 100 Washer 20 95.38 98.50 98.84 100 Washer 22.5 97.38 98.49 95.66 100 Shredder - - - 100 is below this level caused by different motives. As can be seen in Fig. 4 the height and area could be used to derived a simple image based classifier for coins. Fig. 4. Simple coin classification and comparison of the calculated values for the mean height and area with the nominal values. All washers were detected correctly. The accuracy of the washer analysis is similar to the coins, only the area is a little less accurate, due to the hole in the middle of the washers. The real particles were all detected by the system and the position on the belt were calculated correctly. IV. CONCLUSIONS In this work we have shown that the Image Analysis System is capable of detecting particles and calculating all requiredfeaturesatveryhighaccuracyover95%.Thiscanbe done in less than24.5msata full coveredbeltwithsubframes of 400mm width and 100mm length. The system exceeds all requirements and has enough processing capabilities for several extensions. Its simplicity and independency of other systems enable its usage for other applications as well. ACKNOWLEDGMENT This research was funded by the European Commis- sion under FP7, project ShredderSort, Grant Agreement Nr. 603676. REFERENCES [1] E. Grifoni, S. Legnaioli, G. Lorenzetti, S. Pagnotta, and V. Palleschi, “Applying libs to metals processing,” Spectroscopy, pp. 20–31, 2015. [2] Y. Tao, W. Yin, W. Zhang, Y. Zhao, C. Ktistis, and A. Peyton, “A very- low-frequency electromagnetic inductive sensor system for work-piece recognition using the magnetic polarizability tensor,” IEEE Sensors Journal, 2017. 121