An object classification framework based on unmeasurable area patterns found in 3D range images

This paper describes an object detection framework. Depth images obtained from 3D range camera are used, object detection with classification into three types, which are non-transparent, partly-transparent, and transparent, are performed. We focus on image region where measurement data does not obtained, and analyze the reason how such region is produced. It enables us to reduce uncertain region of an input depth image and to provide information with viewpoint changing to obtain more advanced object information. Using the proposed framework, we implemented an application to classify above three types of objects. Non-transparent objects and partly-transparent objects were classified from a single depth image, and multi-view measurements were used to reduce uncertain data and to narrow down the existing area of transparent objects.