Real-Time Stairs Geometric Parameters Estimation
for Lower Limb Rehabilitation Exoskeleton
Xiaoming Zhao
1
, Weihai Chen
1
*, Xing Yan
2
, Jianhua Wang
1
, Xingming Wu
1
1. School of Automation Science and Electrical Engineering, Beihang University, Beijing, 100191
E-mail: *whchenbuaa@126.com
2. Science and Technology on Space Intelligent Control Laboratory, Beijing Institute of Control Engineering, Beijing, 100094
E-mail: 13717990916@163.com
Abstract: Stairs are common structures that hinder the rehabilitation exoskeleton applications in the artificial environ-
ment. It is impossible for the exoskeleton to save all the geometric parameters such as height and depth of each stair
in real-world. By detection and modeling the stairs with computer vision, this paper provides the possibility of wearing
rehabilitation exoskeleton for training hemiplegia patients on stairs. Based on the point cloud reconstructed from RGB-D
data, normal of each point are computed firstly. We subsequently apply over-segmentation then re-aggregation on the
point cloud and normals to extract planes exhaustively. Finally, a stairs graph are modeled based on this planes and
the geometric parameters are computed based on the stairs model. Our algorithm is designed as simple as possible to
reach the real-time requirement for practical situations. We evaluated this stairs modeling algorithm on two situations.
The results indicate that it achieves equivalent precision to state-of-art approaches and even works on partial occlusion
conditions.
Key Words: Stair Detection, Robot Vision Systems, Obstacle Detection, Real-Time
1 INTRODUCTION
As for the improvement of medical treatment, rehabilita-
tion exoskeleton for patients with hemiplegia is gradually
aroused widespread concern. A variety of lower limb reha-
bilitation exoskeletons such as [1–5] have been presented.
In order to reach an effective rehabilitation, it is necessary
that patients can walk as a norm people in the daily environ-
ment. This requirement puts forward environmental per-
ception for lower limb rehabilitation exoskeletons. Stairs
are one of the most common building structure in artificial
environments hindering the patient from walking. This pa-
per focuses on stair estimation problem for the use of lower
limb rehabilitation exoskeletons.
Some stair estimation methods [6] measure the distance to
the step using range sensors, but a distinct drawback is that
sensors must be strictly perpendicular to the step plane.
This shortcoming limits the practicability. Therefore, in
this work, we consider using vision technique to solve stair
estimation and modeling problem.
In some early works [7–9], stair detection form single im-
age using Gabor filters has been proposed. And other re-
cent approach tried to perfect this method by using three
connected point and triangular similarity [10]. However,
these image-based methods just recognized the stair area
in images and cannot obtain accuracy geometric parame-
ters of stairs.
Since depth images can provide more environmental pa-
rameters that 2D images cannot provide, many superior
This work is supported by National Nature Science Foundation under
Grant 61620106012, 61773042 and 61573048.
stairway modeling methods based on depth image have
been proposed. The most convenient way to obtain depth
image is using stereo cameras in previous researches [11,
12, 14, 15]. Nevertheless, dense stereo matching is com-
putationally expensive and may fail in low-texture areas.
Other LIDAR based stair modeling approaches either too
oversimplified [16] or needed to take a long time to scan the
three-dimensional environment [17–19]. Therefore, they
are not suitable for real-time applications.
With the popularity of commodity depth cameras, such
as Microsoft Kinect, researchers began to use the three-
dimensional information to detect the stairs. There are
mainly two kinds of approaches in stair detection or mod-
eling simultaneously using color and depth images cap-
tured from RGB-D sensors: line-based and plane-based ap-
proach.
The former approaches such as [20–22] usually extract
lines from RGB or depth image, and the stairs are de-
fined as a parallel set of lines. And the second main cat-
egory of stair detection algorithms using RGB-D data is
based on plane fitting methods, where stairs are expressed
as a collection of horizontal and vertical planes such as
[23, 24]. Tang et al. [23] segmented horizontal planes from
point cloud and fit these planes by iteratively Preemptive
RANSAC [13]. Inertial measurement unit was embedded
in [24] to correct the orientation of point cloud, horizon-
tal planes were extracted from point cloud and the former
edges of these planes were used to model the stair subse-
quently. In [25], where ground was found by identity the
lowest and largest horizontal plane for orientation correc-
tion, then planes were extracted based on normals of point
5025
978-1-5386-1243-9/18/$31.00
c
2018 IEEE