Self-supervised Learning Through Scene Observation for Selective Item Identification in Conveyor Belt Systems

Conveyor belts are core components in several industries. Often, workers need to operate on conveyor belts, selectively picking objects based on visual features following inspection. In such cases it is desirable to be able to selectively identify such items automatically. The need of tailored systems discourages solutions based on common data collection and labeling, as well as vision systems based on known object features. We device a novel framework to learn from existing human labor, without need of explicit data gathering or labeling. The framework autonomously detects, tracks and learns the visual features of salient objects in conveyor belt-based systems. The system is comprised of two cameras, a Convolutional Neural Network and a novel training regime, devised to support learning on on-line visual feeds. We show the framework is capable of autonomously learning the visual features of the objects picked by human operators. The system trains entirely through visual observation of human labor and achieves detection accuracy of over 97% on a set of 7 different objects after only 10 min of operation, without any knowledge of the objects at priori.