Industrial Image Analysis Group Challenges, Contributions, & Roadmap Dr Matthew Thurley www.ltu.se/staff/m/mjt
Challenges • Which ones are worth doing? • The Long News: stories that might still matter in 50, 100, or ten thousand years from now (TED talk) • Can we think about our research this way? • Can we challenge accepted thinking and really innovate? • Ray Anderson – CEO Interface carpets – Listen to Ray outline how he lead his company to reduce greenhouse gas emission by 82% in tonnage and double profit in 12 years • Amory Lovins – Chief Scientist, Rocky Mountain Institute – Listen to Amory outline how capatilism and business for profit can eliminate America’s dependancy on foriegn oil
Challenges • How will my work be viewed in 20, 50, 100 years? • How can I make a difference for tomorrow’s child? • These are hard questions to answer • One way to start. Be inspired by the work of others • TED.com – Riveting talks by remarkable people • If you only take one thing from my presentation today, take TED
The Challenge that Inspired me • Between 1993 and 1997 I worked for an Australian research organisation, the CSIRO in their exploration and mining division • It was interesting enough but I was still trying to figure out my career • By 1997 it was time to for a change, it was time for a PhD. • Particle size distribution measurement of rock piles using machine vision • More efficient, greener mining - This could make a difference in a huge industry, this was an inspiring challenge
Automated Online Particle Size Measurement using 3D Range Data Dr Matthew Thurley www.ltu.se/staff/m/mjt
The Challenge – Increased Energy Efficiency and Product Quality through Feedback and Control of Particle Processes Dr Matthew Thurley www.ltu.se/staff/m/mjt
The objective is to facilitate this future ■ Facilitate this future by developing the necessary measurement systems to provide fast feedback ■ Fully automated online measurement of the particle size distribution ■ Particle size matters in many industries (cement, construction, steel, paper, agriculture, glass, chemical industry, laundry power, ...) Dr Matthew Thurley www.ltu.se/staff/m/mjt
Pilot Installations & Demonstrators ■ Pilot installation for automated online measurement of green pellets on conveyor belt – LKAB 2007 ■ Pilot installation for automated online measurement of limestone particles on conveyor belt – Nordkalk 2009/10 ■ Demonstration project for automated off-line measurement of rocks in excavator buckets – LKAB 2008 Dr Matthew Thurley www.ltu.se/staff/m/mjt
3D data is collected of the surface Advanced Algorithms determine the particle size distribution from 3D surface data Identifies invidual particles using image segmentation Identifies overlapped & non- overlapped particles preventing mis-sizing of overlapped particles as small particles Robust Image Analysis using 3D data overcomes limitations of 2D photographic imaging. Unaffected by variation in; particle color, shadows, ambient lighting Unaffected by scaling errors and perspective distortion Dr Matthew Thurley www.ltu.se/staff/m/mjt
Identifies non-overlapped particles Advanced Algorithms Automatic classification of overlapped and non-overlapped particles (Thurley & Ng, 2008) 82% classification accuracy using validation against a hold-out data set (Andersson & Thurley, 2008) M. J. Thurley and K. C. Ng. Identification and sizing of the entirely visible rocks from a 3d surface data segmentation of laboratory rock piles. Computer Vision and Image Understanding, 111(2):170 – 178, Aug. 2008. T. Andersson and M. J. Thurley. Visibility classification of rocks in piles. Proceedings of the Conference of the Australian Pattern Recognition Society on Digital Image Computing Techniques and Applications (DICTA 2008), pages 207 – 213, Dec. 2008. Dr Matthew Thurley www.ltu.se/staff/m/mjt
Automated sizing on conveyor – pellets LKAB - Malmberget Industrial Prototype Over two years of maintenance free operation Fully automated measurment Continuously estimates the sieve size distribution, once per minute 9 size classes between 5mm and 16mm (down to 0.5mm spacing between classes) Automatic Control Sizing results facilitate fully automatic control of the pelletizing process Thurley, M.J., Anderson, T. An industrial 3D vision system for size measurement of iron ore green pellets using morphological image segmentation , Minerals Engineering Journal, Vol 28, 5, pp 405-415, 2008 Dr Matthew Thurley www.ltu.se/staff/m/mjt
Automated sizing on conveyor – pellets LKAB - Malmberget Industrial Prototype Two years of operation Fully automated measurment Continuously estimates the sieve size distribution, once per minute 9 size classes between 5mm and 16mm (down to 0.5mm spacing between classes) Automatic Control Sizing results allow fully automatic control of the pelletizing process Dr Matthew Thurley matthew.thurley@ltu.se
Estimating the Sieve-Size Distribution – Sources of Error • Segregation error (brazil-nut-effect): vibration causes a separation effect where large fragments move to the surface • Surface bias due to size: larger fragments are more likely to be visible due to their size • These errors limit the range of sizes we can see (only the larger sizes) • Overlapped particle error: overlapped particles will look like smaller particles • Profile error: Is the visible profile of a particle indicative of its size • Weight transformation: How to convert number of rocks recorded by imaging, to a weight of rocks equivalent to sieving Dr Matthew Thurley www.ltu.se/staff/m/mjt
Automated sizing on conveyor – rocks Nordkalk – Gottland Limestone Quarry Requirements Fully automated measurment Quality control of material size during ship loading. Capability to report loading of the wrong size class Ability to report deviations in desired material size during loading of the ship in order to detect mechanical failure in the screen decks. Thurley, M.J. Automated online measurement of the particle size distribution using 3D range data, IFAC MMM Workshop, Vina del Mar, Chile, October 2009 Dr Matthew Thurley www.ltu.se/staff/m/mjt
1 Automated 2 Segmentation 3 1. Raw 3D surface data 2. Fully automated segmentation 3. Fully automated exclusion of overlapped rocks Dr Matthew Thurley www.ltu.se/staff/m/mjt
Automated sizing on conveyor – Preliminary Online Results Nordkalk – Gottland Limestone Quarry: Measurement trends in the right direction ■ By the end of 2009 we demonstrated that the online fully automated results trend in the right direction ■ Research is ongoing to improve accuracy of the absolute size results ■ The following three images were taken from the online measurement system during the 14th and 15th of December 2009 Dr Matthew Thurley www.ltu.se/staff/m/mjt
20-40mm being loaded 14.12.2009 17:23 Dr Matthew Thurley www.ltu.se/staff/m/mjt
10-90mm being loaded 14.12.2009 20:56 Dr Matthew Thurley www.ltu.se/staff/m/mjt
10-90mm being loaded 15.12.2009 9:02 Dr Matthew Thurley www.ltu.se/staff/m/mjt
Automated sizing on conveyor – rocks Nordkalk – limestone quarry: Measurement trends in the right direction ■ The raw size measurement results trend in the right direction because they are based a physically observable property, the Best-Fit-Rectangle (BFR) area of the non-overlapped rocks ■ 20-40mm product ■ Median value 882mm 2 ■ 40- 70mm product ■ Median value 1901mm 2 Dr Matthew Thurley www.ltu.se/staff/m/mjt
Identifying the Product Nordkalk – limestone quarry ■ Using this physically observable property, the BFR area of the non- overlapped rocks we can identify the product being loaded ■ For each measurement data set (surface data for approx 1m length of the belt) we calculate the median and IQR of the sample Anderson, T., Thurley M., Carlson, J.E. Online Product Identification during ship loading of limestone based on machine vision. Machine Vision & Applications 2010 (in submission) Dr Matthew Thurley www.ltu.se/staff/m/mjt
Identifying the Product Nordkalk – limestone quarry Product identification probabilities: 98.8 % accurate classification Andersson, T. and Thurley, M.J. and Carlson, J. Online Product Identification during Ship Loading of Limestone using Machine Vision, In submission, 2010 Dr Matthew Thurley www.ltu.se/staff/m/mjt
How was this overlapped/non-overlapped classification achieved Collected good experimental data ■ Collected a rock sample ■ Sieved into 3 size classes, and painted each rock to identify it by size class ■ Blended the rocks into mixed pile in a cylindrical bucket ■ Captured 3D surface data of the pile ■ Manually characterised the surface ■ Size of each visible rock ■ How visible is each rock? ■ Entirely visible = non-overlapped ■ Predominantly visible = only a minor corner or edge overlap ■ Overlapped = everything else ■ Develop algorithms ■ Segment the rocks ■ Identify overlapped vs non-overlapped
Key Algorithm ■ Detecting overlapped vs non-overlapped particles
Recommend
More recommend
