We found a match
Your institution may have access to this item. Find your institution then sign in to continue.
- Title
Multitarget Tracking Using Orientation Estimation for Optical Belt Sorting.
- Authors
Pfaff, Florian
- Abstract
Bulk material is a frequently encountered type of product in industrial applications. Characteristic of bulk materials, which include mining products, grains, and recyclable materials, is that they are loosely transported in large quantities. The processing of bulk materials is estimated to consume up to 10% of all energy produced worldwide. Optical belt sorters are a versatile technology for reliably separating streams of bulk material even when classical separators such as sieves cannot be applied. In optical belt sorters, a stream of bulk material is separated into two streams using bursts of compressed air that alter the motion of the particles that are hit. The bursts are emitted using digitally controlled valves. In sorters currently used in industry, the valves are controlled based on a single observation of each particle obtained using a line scan camera. Due to processing delays, the particles must be observed before they are in the reach of the bursts of compressed air. Therefore, a prediction of each particle's motion is required. We consider the advantages and challenges of replacing the usual line scan camera with an area scan camera, which allows us to obtain multiple observations of each particle. To make use of all observations of a particle, a multitarget tracking algorithm tailored to the bulk material sorting task is derived. The tracking is based on the particles' centroids extracted using image processing algorithms, which are employed without modifications in this thesis. For our novel approach, we adjust and extend a classical multitarget tracking algorithm. Further, we discuss potential challenges specific to the application and suggest solutions. A particularly important component of thetracking algorithm is the motion model. In this thesis, we propose novel motion models that take knowledge about previously observed particles into account. Thus, we are able to outperform classical motion models given in the target tracking literature. Reliably determining which observations in consecutive frames stem from the same particle is key to the success of a multitarget tracking algorithm. To achieve a high reliability, we not only rely on the predicted position of each particle but also estimate and integrate the orientation as an additional feature. Estimating the orientation of a particle around the vertical axis is an estimation problem on a circular manifold. Estimation on circular or, more generally, periodic manifolds is an active field of research to which this thesis also contributes. For circular manifolds, two filters based on Fourier series are introduced that allow representing asymmetric and multimodal densities. The first filter is based on approximating prior and posterior densities directly using Fourier coefficients. In contrast, the square roots of the densities are approximated in the second filter. The two essential steps for recursive Bayesian estimators, the update and the prediction step, are explicated for both filters. As we explain how to implement Bayes' formula for the update step and the Chapman-Kolmogorov equation for the prediction step, the filters are applicable to almost arbitrary estimation problems on the unit circle. The presented filters are generalized for more challenging estimation problems on higher-dimensional manifolds. While the generalization to hypertoroidal manifolds only requires the use of multidimensional Fourier series and minor modifications, more extensive changes are required for adapting the concepts to estimation problems on the unit sphere. For the unit sphere, two filters based on spherical harmonics are presented. The explained filters for arbitrary-dimensional hypertori are the first that provide continuous approximations of the probability densities. For the unit sphere, the filters are the first that allow for improvements to the approximations of the densities by increasing the number of parameters used. The filters are evaluated on a variety of scenarios with different challenges. In all scenarios, our filters achieve high-quality estimation results at lower run times than a particle filter adapted to periodic domains.
- Subjects
INDUSTRIAL applications; BULK solids; RECYCLABLE material; ENERGY consumption; TRACKING algorithms
- Publication
Karlsruhe Series on Intelligent Sensor-Actuator-Systems, 2019, Vol 22, pi
- ISSN
1867-3813
- Publication type
Article
- DOI
10.5445/KSP/1000094502