A probabilistic next best view planner for depth cameras based on deep learning

In this work we present a novel method to plan the next best view of a depth camera by leveraging on a Convolutional Neural Network (CNN), and on a probabilistic occupancy map of the environment for ray casting operations. In particular, a hybrid approach is introduced that exploits the convolutional encoder-decoder to perform object completion, and an algorithm based on ray casting to evaluate the information gain of possible sensor view poses. Automatic object completion consists of inferring the occupancy probability of the regions of space that have not been observed. A comparison against several methods, including City-CNN, was carried out in 2D and 3D environments on publicly available datasets. In particular, to enable comparison, the original City-CNN algorithm was modified to work with depth cameras. Experiments indicate that the proposed method achieves the best results in terms of exploration accuracy. Results on a real robot are also presented.