Efficient and Precise Multi-Camera Reconstruction

Abstract

Human-robot cooperation requires efficient and precise geometric reconstruction of objects in any shared workspace to meet safety requirements. State-of-art solutions accept a trade-off to satisfy these requirements. Usually, this implies limited robot speed, expensive hardware, or low-precision input and output. In contrast, we present a novel multi-camera reconstruction approach that is both efficient and precise: We apply incremental and hierarchical algorithms to image preprocessing and visual hull reconstruction. Subsequent knowledge-based postprocessing refines the initial reconstruction. Given ground-truth input, we furthermore maintain conservative bounds of objects in the workspace. In total, our contribution satisfies real-time and anytime requirements even on modest hardware and for high-precision input and output. We validate this by synthetic and real-world experiments. Thus, our approach is suited for safety-, time-, and precision-critical reconstruction with applications in home robotics, surveillance, and industrial automation.