A View Planning Method for 3D Reconstruction with Unknown Feature Prediction

Abstract

Active reconstruction can be considered as an intelligent perception problem that seeks to optimize the efficiency of the modeling tasks by adjusting the configuration of vision system on robotic platform automatically. We consider a feature as several local surfaces that have the same kinds of characteristics. A standard geometry can be defined by attributes abstracted from a set of features. For the regular geometry of the same category, descriptions and topological relations about these objects’ features are known before modeling tasks. In order to maximize the acquisition of an object’s surface in a few views, knowledge of its features is employed as partial prior information that guides the view planning in active reconstruction. Since each reconstruction task may have a different type of object to be rebuilt, in order to make the view planning algorithm suitable for different tasks, we store the prior information of the object in a fixed-form database, so as to call the content of the library to perform verification and reasoning of all features, and make predictions on unknown local surface and rough shape of the object. The predicted unknown surface is discretized and then marked as predicted voxels in the voxel space, which are used to evaluate the priority of all candidate viewpoints, together with the detected voxels. A novel hybrid optimization function is proposed that takes the constraint on overlap, the accessibility of predicted surface, and the amount of unknown object surface into account. Simulated experiments are conducted and the results show that our method is efficient for reconstruction.