Feature-Aided Multitarget Tracking for Optical Belt Sorters

Industrial optical belt sorters are highly versatile in sorting bulk material or food, especially if mechanical properties are not sufficient for an adequate sorting quality. In previous works, we could show that the sorting quality can be enhanced by replacing the line scan camera, which is normally used, with an area scan camera. By performing multitarget tracking within the field of view, the precision of the utilized separation mechanism can be enhanced. The employed kinematics-based multitarget tracking crucially depends on the ability to associate detection hypotheses of the same particle across multiple frames. In this work, we propose a procedure to incorporate the visual similarity of the detected particles into the kinematics based multitarget tracking that is generic and evaluates the visual similarity independent of the kinematics. For evaluating the visual similarity, we use the Kernelized Correlation Filter, the Large Margin Nearest Neighbor method and the Normalized Cross Correlation. Although no clear superiority for any of the visual similarity measures mentioned above could be determined, an improvement of all considered error metrics was attained. Index Terms-feature-aided multitarget tracking, industrial optical belt sorters, metric learning, visual tracking.