Perception planning for an exploration task of a 3D environment

Incremental recovery of the 3D model of a closed environment is a complex task as it involves several functions: 3D acquisition, data registration and fusion, construction of a 3D triangular mesh or other representations. The sensor must be placed in several positions so that the resulting model is complete (no unseen areas) and has the required resolution. This paper presents a method devoted to the selection of the next best view for a sensor used for 3D modelling of an environment; the objective is to minimise the number of acquisitions and optimise the quality of the final model. From the current sensor position, an optimal next view is selected by using utility functions computed for possible solutions; the method combines exhaustive search and hill climbing optimisation. Thanks to a simulation, results are presented for a synthetic environment, using a virtual 3D sensor on a 5 degrees of freedom robot.