The resulting vision system mainly relies on monocular vision algorithms, while being able to track metric scale by
stereo triangulation. Since the two cameras are configured in a stereo setup, the field of view of the vision system is
not expanded. The improved work in Shen et al. (2013a) enables a quadrotor to autonomously travel at speed up to 4
m/s, and allows roll and pitch angles exceeding 20 degrees in 3D indoor environments.
In Huang et al. (2011), autonomous flight of an MAV is enabled by the proposed SLAM system using an RGB-
D camera. A visual odometry is used for real-time local state estimation, and integrated with the RGBD-Mapping
(described in Henry et al. (2014)) to form the SLAM system. An efficient RGB-D SLAM system is described in
Scherer and Zell (2013), which enables an MAV to autonomously fly in an unknown environment and create a map of
its surroundings. In this work, sparse optical flow is used for feature matching, which is of advantage when motion
blur may result in very limited number of local features being detected. The work in Valenti et al. (2014) perform
RGB-D visual odometry on an MAV to enable its autonomous flight. 3D occupancy grid map is then built for path
planning of the MAV.
Previous work on developing multi-camera systems can mainly be found in applications of surveillance and object
tracking (Collins et al., 2000; Krumm et al., 2000). More relevant related work appears in the context of structure from
motion (SFM). The work in Pless (2003) presents a theoretical treatment of multi-camera systems in SFM deriving
the generalized epipolar constraint. The work in Frahm et al. (2004) proposes a virtual camera as a representation of a
multi-camera system for pose estimation. A structure-from-motion scheme is achieved using multiple cameras in this
work.
A number of multi-camera systems for pose estimation of mobile robots can be found in the literature. In Ragab
(2008), pose estimation of a mobile robot is solved by placing two back-to-back stereo pairs on the robot using the
Extended Kalman Filter (EKF). The work in Lee et al. (2013a) adopts a generalized camera model described in
Pless (2003) for a multi-camera system, to estimate the ego-motion of a self-driving car using a 2-Point RANSAC
algorithm. This system allows point correspondences among different cameras. In Lee et al. (2013b), pose-graph
loop-closure constraints are computed. The relative pose between two loop-closing pose-graph vertices is obtained
from the epipolar geometry of the multi-camera system. Kaess and Dellaert (2006) presented a visual SLAM system
with a multi-camera rig using Harris corner detector (Harris and Stephens, 1988). In their further work in Kaess and
Dellaert (2010), a Bayesian approach to data association is presented taking into account moving features which can be
observed by cameras under robot motion. The work in Sol
`
a et al. (2008) provides solutions to two different problems
in multi-camera visual SLAM: automatic self-calibration of a stereo rig while performing SLAM and cooperative
monocular SLAM.
In Harmat et al. (2012), PTAM with multiple cameras mounted on a buoyant spherical airship is reported in a manual
flight experiment. It employs a ground-facing stereo camera pair which can provide metric scale, together with another
camera mounted pointing to the opposite direction using a wide-angle lens. In Tribou et al. (2015), multi-camera visual
SLAM is achieved based on a modified version of PTAM. This SLAM system allows convergence in pose tracking
and mapping with the absence of accurate metric scale, taking advantages of the Taylor omnidirectional camera model
and a spherical coordinate update method.
In order to use multiple cameras for pose estimation, the extrinsic parameters of those cameras need to be calibrated.
Here we are intereted in the previous work solving this calibration problem for multiple cameras with non-overlapping
fields of view. Carrera et al. (2011) proposed a SLAM-based automatic calibration scheme for multiple cameras. The
scheme uses global bundle adjustment to optimize the alignment of maps built by different visual SLAM instances,
each processing images from one corresponding camera. The proposed solution computes the relative 3D poses among
cameras up to scale. A more recent work can be found in Heng et al. (2014a), which uses a computationally expensive
SLAM system to accurately map an infrastructure before the extrinsic calibrations of multi-camera systems. During a
calibration process, 2D-3D correspondences among visual features in the current scene and in the previously known
map are used for tracking the camera poses based on a Perspective-n-point (PnP) method. Then camera extrinsics are
optimized via non-linear refinement. Self-calibration of multiple stereo cameras and IMU is further achieved in Heng
et al. (2014b), which also achieved multi-camera visual SLAM based on the generalized camera model. At lease one
set of stereo cameras is required in this work. Real-time loop closing runs on-board of MAVs is achieved in both Yang
et al. (2016) and Heng et al. (2014b)
