Visual classification of single waste items in roadside application scenarios for waste separation

In order to automate green space maintenance at roadsides, mobile robots need specialized perception systems for waste discovery and classification enabling them for direct waste separation during collection. This paper presents an approach to classify single waste items which typically can be found at roadside. It is based on the concept of transfer learning which reuses features of pre-trained neural networks and needs little data. These features are evaluated with different classifiers. For the evaluation, we extended an existing database with an ample collection of new waste images and introduced an alternative division into waste classes which better matches the waste separation task.