I. INTRODUCTION
Today, laser scanners are still widely used in robotics and
autonomous vehicles, mainly because they directly provide
3d measurements in real-time. However, compared to traditional
camera systems, 3d laser scanners are often more
expensive and more difficult to seamlessly integrate into
existing hardware designs (e.g., cars or trains). Moreover,
they easily interfere with other sensors of the same type as
they are based active sensing principles. Also, their vertical
resolution is limited (e.g., 64 laser beams in the Velodyne
HDL-64E). Classical computer vision techniques such as
appearance-based object detection and tracking are hindered
by the large amount of noise in the reflectance measurements.
Motivated by those facts and the emergent availability
of high-resolution video sensors, this paper proposes a
novel system enabling accurate 3d reconstructions of static
scenes, solely from stereo sequences1. To the best of our
knowledge, ours is the first system which is able to process
images of approximately one Megapixel resolution online
on a single CPU. Our contributions are threefold: First, we
demonstrate real-time scene flow computation with several
thousand feature matches. Second, a simple but robust visual
odometry algorithm is proposed, which reaches significant
speed-ups compared to current state-of-the-art. Finally, using
the obtained ego-motion, we integrate dense stereo measurements
from LIBELAS [12] at a lower frame rate and
solve the associated correspondence problem in a greedy
fashion, thereby increasing accuracy while still maintaining
efficiency. Fig. 1 illustrates the input to our system and
resulting live 3d reconstructions on a toy example.