A global generalized maximum coverage-based solution to the non-model-based view planning problem for object reconstruction

Abstract

A robotic active vision system can automatically perform the reconstruction task, but the core problem is the non-model-based view planning problem (NMVPP). Unlike the model-based one that requires a priori known 3D models of objects or environments, the NMVPP faces an unknown environment, making it more challenging. This is usually solved by an iterative greedy algorithm based on information gain (i.e., quantitative statistics of uncertainty in the environment). For better surface coverage, we seek a global optimization solution to overcome the bad performance of greedy selections. We first naturally generalized the model-based set covering optimization (SCO) problem, but it performs poorly because of uncertainty in the environment. Therefore, we present a generalized maximum coverage (GMC)-based solution to the NMVPP, which generates an optimal sequence of views capable of sensing all valuable areas of an unknown object placed in a 3D environment. Given the uncertainties in the environment, the goal is to maximize the total information of the optimal view subset, making NMVPP a class of GMC problems. For the sake of reconstruction efficiency, we introduce an additional subjection of the GMC model to balance the movement cost and surface coverage. We compared the results of our methods with state-of-the-art algorithms, both on a set of 3D objects and in the real world, and the results showed that our methods could outperform the others in most cases. Thus, the proposed active vision system can automatically complete an unknown object reconstruction task with higher coverage and competitive running time.