Model-based active viewpoint transfer for purposive perception

摘要

In vision-based manipulation tasks, it is a fundamental procedure to transfer the sensor to a target viewpoint for better observation or manipulation. However, the limitations of field of view (FOV) and obscure issues, which are common on industrial occasions, impede the application of visual perception in unstructured environment. In this paper, an online navigation planning strategy of viewpoint transfer for purposive perception is developed, then a 2D viewpoint estimation method using CNN-based spherical viewpoint node classification is presented. To satisfy the requirements of large scale annotated training samples for deep learning, a sample synthesizing method by applying virtual camera in CAD environments is also proposed. We evaluate our method on specific datasets and a UR10 robot, with the experimental results shown in high efficiency and feasibility.