Real-Time RGB-D Mapping and 3-D Modeling on the GPU Using the Random Ball Cover

Abstract In this chapter, we present a system for real-time point cloud mapping and scene reconstruction based on an efficient implementation of the iterative closest point (ICP) algorithm on the graphics processing unit (GPU). Compared to state-of-the-art approaches that achieve real-time performance using projective data associ-ation schemes which operate on the 3-D scene geometry solely, our method allows to incorporate additional complementary information to guide the registration pro-cess. In this work, the ICP’s nearest neighbor search evaluates both geometric and photometric information in a direct manner, achieving robust mappings in real-time. In order to overcome the performance bottleneck in nearest neighbor search space traversal, we exploit the inherent computation parallelism of GPUs. In particular, we have adapted the random ball cover (RBC) data structure and search algorithm, orig-inally proposed for high-dimensional problems, to low-dimensional RGB-D data. The system is validated on scene and object reconstruction scenarios. Our imple-mentation achieves frame-to-frame registration runtimes of less than 20 ms on an off-the-shelf consumer GPU.