Point Cloud Computing for Rigid and Deformable 3D Object Recognition

Machine vision is a technologically and economically important field of computer vision. It eases automatization of inspection and manipulation tasks, which in turn enables cost savings and quality improvement in industrial processes. Usually, 2D or intensity images are used for such applications. However, thanks to several technological advances, nowadays there are sensors available that allow depth or 3D measurements with high resolution, frequency and accuracy at a reasonable cost. Such 3D data enables new applications that are difficult or impossible to implement with 2D images only. This work develops several performant, robust and accurate algorithms for processing such 3D data. The algorithms were developed with the requirements of industrial image processing in mind. They are, however, applicable to other areas such as robotics and reverse engineering as well. Two fundamental challenges are solved in this work: The fast localization of 3D points that neighbor a given query point and the detection of rigid and deformable objects in 3D point clouds or in multimodal data. Additionally, a fast and robust method for refining the position of two point clouds is presented and some fundamental algorithms regarding rotations are discussed. For the detection of nearest neighbors in 3D point clouds, a voxel-based method is introduced that allows almost constant lookup times of O(log(log(N))). Additionally, in contrast to prior art, lookup times are almost independent of the distribution of query and data points, allowing the use of this method in real-time systems. A variant of the method that allows for approximate nearest neighbor lookups further improves the required time for creating the underlying data structure. For the detection of rigid and deformable objects in 3D and multimodal data, a local variant of the Hough transform is introduced that circumvents the usual problems of voting schemes over high-dimensional parameter spaces. The very robust, fast, and generic baseline method detects rigid objects in 3D point clouds. It works for arbitrary free-form objects and different 3D sensors alike and can cope with large amounts of clutter, noise, occlusion, sparse data and multiple object instances. Three variants of the baseline method are introduced that detect rigid objects in multimodal data, geometric primitives in 3D point clouds, and deformable objects in 3D point clouds. For this, the parameter space, the used feature, the model representation and the number of voting rounds of the original method are modified according to the type of data and model. All methods are evaluated using up-to-date datasets.