Machine learning-based multiobject tracking for sensor-based sorting

Sensor-based sorting provides state-of-the-art solutions for sorting of granular materials. Current systems use line-scanning sensors, which yields a single observation of each object only and no information about their movement. Recent works show that using an area-scan camera bears the potential to decrease both the error in characterization and separation. Using a multiobject tracking system, this enables an estimate of the followed paths as well as the parametrization of an individual motion model per object. While previous works focus on physically-motivated motion models, it has been shown that state-of-the-art machine learning methods achieve an increased prediction accuracy. In this paper, we present the development  of a neural network-based multiobject tracking system and its integration into a laboratory-scale sorting system. Preliminary results show that the novel system achieves results comparable to a highly optimized Kalman filter-based one. A benefit lies in avoiding tiresome manual tuning of parameters of the motion model, as the novel approach allows learning its parameters by provided examples due to its data-driven nature.