Systems, Man & Cybernetics - April 2016 - 37

detected stairway may be used as
reduce the accumulative pose error
a waypoint to access the next
of the dead reckoning approach.
As most of the
floor and a doorway a waypoint
Our recent study [16] shows that a
target objects are
to enter/exit a room. Detection of
PGO method has a better perfora hallway may assist the CRC
mance consistency than an EKF.
indoor structures,
user with moving direction. Five
Therefore, we use PGO to minimize
their determining
objects (doorway, hallway, stairthe pose error in this article.
way, ground, and wall) representLet x = ^ x 1, f, x N hT be a vector
factors are geometric
ing indoor structures and three
consisting of nodes x 1, f, x N,
features rather than
objects (monitor, table, and paralwhere x i for i =1, f, N, is the
lelepiped) related to office enviSR4000's pose at time step i (i.e.,
visual features.
ronment are considered in this
image frame i). Let z ij and X ij be
article. It is noted that the paralthe mean and information matrix of
lelepiped is not yet associated
a virtual measurement between
with an office object yet. But it can be a computer case
nodes i and j. Let zt ij be the expected value of z ij . The meaif the relevant data are used to train the detector. As
surement error e ij ^ x i, x j h = z ij - zt ij ^ x i, x j h and X ij are used
most of the target objects are indoor structures, their
to describe the edge connecting nodes i and j. Figure 4
determining factors are geometric features rather than
shows a pose graph with five nodes.
visual features. For example, a stairway contains a
A number of approaches are taken to ensure a quality
group of alternating treads and risers with a fixed size
graph for PGO. When adding node x i to the graph, z^i - 1hi
while its visual appearance can be various. Therefore, a
(i.e., pose change between i and j) may be unavailable if
geometric-feature-based object-recognition method
VRO-FICP fails. In this case, we assume a constant movebecomes appropriate.
ment and use z^i - 2h^i - 1h to create node x i, i.e., e^i - 2h^i - 1h and
The state-of-the-art graph-based object-recognition
X^i - 2h^i - 1h are used for edge E^i - 1hi . For nonconsecutive
method [19] is nondeterministic polynomial-time hard.
nodes, no edge is created if VRO-FICP fails or the pose
For computational efficiency, we propose a GMM-based
change uncertainties are large. We use the following bootobject-recognition method as depicted in Figure 5. It constrap method to compute pose change uncertainties: 1)
sists of five main procedures: range data acquisition,
compute a pose change using K = min ^0.75N, 40 h samples
plane extraction, feature extraction, GMM plane classifier
randomly drawn from the N correspondences given by
design and training, and plane clustering. Each of them is
VRO-FICP and 2) compute 50 pose changes by repeating
described in this section.
Step 1 and calculate the standard deviation. Finally, X ij is
computed as the inverse of the diagonal matrix of the pose
change uncertainties. Compared with the existing works
Range Data Acquisition and Plane Segmentation
that assume a constant uncertainty in graph construction,
Using the estimated poses, the 3-D range data of the camour method may result in a more accurate graph and thus
era are registered to form a large 3-D point cloud map. The
improve the pose estimation result.
point cloud data is then segmented into N planar patches,
The PGO process is to find the node-configuration x *
P1, P2, f, PN, by the NCC-RANSAC method [20].
that minimizes the nonlinear cost function
Features and Feature Vectors
(2)
F ^ x h = |e ij ^ x i, x j hT X ij e ij ^ x i, x j h .
Each of the N patches is then assigned a feature vector
ij
that describes its intrinsic attributes and geometric context, i.e., interplane relationships (IPRs) representing the
A numerical approach based on the Levenberg-Margeometric arrangement with reference to another planar
quardt algorithm is used to solve the optimization problem iteratively. Details on the PGO method are referred
to [17]. In this article, we use the GTSAM C++ library [18]
E13
for PGO. The VRO-FICP-based approach substantially
improves the pose estimation accuracy of the VROx2 E23
E12
based method and reduces the computational time of the
E35
E12
x1
VRO-ICP-based method. Readers are referred to [15] for
x4
x3
E24
more details.
E45

3-D Object Recognition
Object recognition is important to wayfinding. Objects
detected by the CRC can be used 1) as waypoints for
navigation, 2) for environment awareness (e.g. in an
office, hallway, etc.), and 3) for obstacle avoidance. A

E14
E15

Figure 4. a pose graph with five nodes; E ij =< e ij , X ij >
represents the edge between nodes i and j.

Ap ri l 2016

IEEE SyStEmS, man, & CybErnEtICS magazInE

Table of Contents for the Digital Edition of Systems, Man & Cybernetics - April 2016