DNA-SLAM: Dense noise aware SLAM for ToF RGB-D cameras

Abstract

SLAM with RGB-D cameras is a very active field in Computer Vision as well as Robotics. Dense methods using all depth and intensity information showed best results in the past. However, usually they were developed and evaluated with RGB-D cameras using Pattern Projection like the Kinect v1 or Xtion Pro. Recently, Time-of-Flight (ToF) cameras like the Kinect v2 or Google Tango were released promising higher quality. While the overall accuracy increases for these ToF cameras, noisy pixels are introduced close to discontinuities, in the image corners and on dark/glossy surfaces. These inaccuracies need to be specially addressed for dense SLAM. Thus, we present a new Dense Noise Aware SLAM (DNA-SLAM), which considers explicitly the noise characteristics of ToF RGB-D cameras with a sophisticated weighting scheme. In a rigorous evaluation on public benchmarks we show the superior accuracy of our algorithm compared to the state-of-the-art.