Tuning and Comparison of Optimization Algorithms for the Next Best View Problematic

The aim of this paper is to tune and compare different optimization algorithms on the Next Best View (NBV) problem, which consists in finding the next position that the sensor or camera needs to take to scan an object or scenery in its totality. A simulated 5 Degree-of-Freedom (DOF) mobile robot with a mounted simulated range sensor was used on a Virtual Reality Modeling Language (VRML) environment, and the space discretization was made using a voxel map. For the objective function, two main factors were included: an area factor to make sure that the image taken by the sensor provides the best possible information, and a motion factor made up of distance and energy sub-factors to reduce the resources used by the robot. Tasks such as a backstepping technique to escape local minima and a dynamic change in the objective function were implemented. The retrievement of the scene was made on an iterative process, and three different optimization methods were tuned and tested: Nelder-Mead, an Evolution Strategy, and Simulated Annealing. A set of experiments comparing the three methods in computational time and retrievement efficiency were made on three different environments with increasing difficulty to test their repeatability, with them being a laboratory model, a room with a cube and a pyramid inside it, and a study room with multiple furniture and windows.