Page - 120 - in Proceedings of the OAGM&ARW Joint Workshop - Vision, Automation and Robotics

An Image Analysis System for Selective Recovery of Non-ferrousMetal Malte Jaschik1, Alfred Rinnhofer2, Martina Uray3 and Gerhard Jakob4 Abstract—To increase the recycling rate for non-ferrous metal, a high speed sorting line with a high throughput rate of up to 1ton per hour was built. The system comprises an Image Analysis System to detect shredder particles and calculate their position on the belt as well as several 2D and 3D shape features. ElectroMagnetic Tensor Spectroscopy (EMTS) or Laser-Induced Breakdown Spectroscopy (LIBS) characterize each particle based on its metal components. Tests were conducted under hard conditions in an industrial environment. For a full covered 400mm x 100mm belt area the ImageAnalysis Systemneeds less than 24.5ms at a feature calculation accuracy up to 95%. The developed system can easily be adapted to other scenarios. I. INTRODUCTION The requirements of the sorting line described in this work are to detect and classify non-ferrous metal particles. The load speed of the sorting line is given by 1ton/hour. A vibrating feeding system is used to load the belt (width of 400mm) and for fragment separation. Due to the limitation of the vibrating feeding system, about 28g/s of particles can be loaded with a belt speed of 2m/s. The subsystems need an accuracy of 0.5mm/px in length and width. Therefore, a line rate of 4kHz is required. Fig. 1 shows a schematic of the components and Fig. 2 a sample image of such a particle. The system is developed to work even under heavy industrial conditions (like dust, vibrations of machines, etc.). Fig. 1. Schematic of the developed sorting line. The developed sorting line consists of two independent subsystems (detector and classifier). An Image Analysis System identifies every single particle on the conveyor and calculates its exact position on the belt as well as 2D and Joanneum Research Forschungsgesellschaft mbH DIGITAL - Institute for Information and Communication Technologies 1malte.jaschik@joanneum.at 2alfred.rinnhofer@joanneum.at 3martina.uray@joanneum.at 4gerhard.jakob@joanneum.at 3D shape features. The analysis and classification of the material of the particles is provided by EMTS or LIBS. The EMTS measures the electrical conductivity while the LIBS characterizes the chemical composition. To achieve optimum results of the classification systems, it is essential that the position and specified shape features of every single particle is derived to utmost precision by the Image Analysis System. Therefore, all subsystemsaresynchronisedbyan incremental encoder. II. DISCRIPTION OF SUB SYSTEMS A. Image Analysis System for position and shape calculation The Image Analysis System is based on laser triangulation and comprises a line laser, an Automation Technology C4- 1280-GigE 3D camera and a computer for calculation. Using subpixel algorithms a height resolution, defined by the opti- cal resolution and the angle between laser plane and camera, of 0.15mm can be achieved. Due to the belt movement the laser line migrates along the surface of the fragments. The camera acquires a 2D image of each laser line and calculates a 3D profile that is sent to the computer via GigE. The analysis algorithm stores the 3D profile and builds an image, called subframe (currently consisting of 200 3D profiles), a small part is illustrated in Fig. 3. The camera can acquire a grayscale image (Fig. 2) as well, but it is not used in the current application. Fig. 2. Grey-value image of one metal particle Fig. 3. 3D image of one metal particle (same as Fig. 2) To remove noise due to non-flatness of the conveyor a background model is calculated and subtracted from the image (for each 3D profile). By binarizing the subframe with an experimentally determined height threshold areas of interest are selected. If the size of an area is big enough, it defines a particle hypothesis and features are calculated. Features reach from simple positions to more complex ones, like feret diameters or maximum cross-sectional area (overall 25 different features in 2D and 3D respectively). 120