AN ALGORITHM AND SYSTEM FOR FINDING THE NEXT BEST VIEW IN A 3-D OBJECT MODELING TASK A thesis

Sensor placement for 3-D modeling is a growing area of computer vision and robotics. The objective of a sensor placement system is to make task-directed decisions for optimal pose selection. This thesis proposes a Next Best View (NBV) solution to the sensor placement problem. Our algorithm computes the next best view by optimizing an objective function that measures the quantity of unknown information in each of a group of potential viewpoints. The potential views are either placed uniformly around the object or are calculated from the surface normals of the occupancy grid model. For each iteration, the optimal pose from the objective function calculation is selected to initiate the collection of new data. The model is incrementally updated from the information acquired in each new view. This process terminates when the number of recovered voxels ceases to increase, yielding the nal model. We tested two di erent algorithms on 8 objects of various complexity, including objects with simple concave, simple hole, and complex hole self-occlusions. The rst algorithm chooses new views optimally but is slow to compute. The second algorithm is fast but not as e ective as the rst algorithm. The two NBV algorithms successfully model all 8 of the tested objects. The models compare well visually with the original objects within the constraints of occupancy grid resolution. Objects of complexity greater than mentioned above were not tested due to the time required for modeling. A mathematical comparison was not made between the objects and their corresponding models since we are concerned only with the acquisition of complete models, not the accuracy of the models.