High-throughput sensor-based sorting via approximate computing

Sensor-based sorting provides solutions for separating
cohesive, granular materials. In order to reliably locate the position
of material objects with deviating velocity, perception and
separation shall be close together. This in turn poses challenges
on the data analysis systems, since available processing time depends
on this distance and the velocity of the object. Whenever
the sorting decision for an object cannot be derived in time, no
information about this object is taken into account, potentially
leading to a sorting error. In this paper, we present an analysis
of the impact of this distance and an approach which allows utilizing
information about an object before the final classification
result is available. Therefore, we apply the concept of anytime
algorithms to a decision tree-based classifier. First results suggest
that the approach can indeed increase the sorting quality for
complex objects for which the deadline would else not be met.