Mapping is a crucial task for robot navigation. Especially, to develop a fully autonomous robot that can interact well with humans, the map is required to contain not only geometrical contents but also the semantic contents. Moreover, the mapping algorithm needs to be lightweight for real-time robot applications. In this paper, we propose a new method for environmental detail mapping and annotation. The proposed method integrates feedforward neural network for mapping the depth image into real-world coordination and convolutional neural network for environment segmentation. The segmented objects are modelled and placed on the three-dimensional map of the environment. Experiments with real robot show a good performance of the proposed method.
展开▼
