Content uploaded by Dmitry G. Korzun
Author content
All content in this area was uploaded by Dmitry G. Korzun on Nov 30, 2021
Content may be subject to copyright.
Available via license: CC BY-ND
Content may be subject to copyright.
Trajectory Construction
for Autonomous Robot Movement
based on Sensed Physical Parameters and Video Data
Grigorij Rego, Nikita Bazhenov, Dmitry Korzun
Petrozavodsk State University (PetrSU)
Petrozavodsk, Russia
{rego, bazhenov, dkorzun}@cs.karelia.ru
Abstract—Positioning of a robot in its autonomous movement
is applied to the automatic control and trajectory planning.
There are two types of systems: sensed data come from physical
parameters sensors and video cameras. In this paper, we consider
the integration of both data sources, providing a novel way for
trajectory construction. The sensed data are collected from the
IMU-sensor, camera, and Lidar. We present an architecture for
collecting data from many sources and for using the data to
evaluate the position of a mobile robot. The results of our early
experiments confirm the feasibility of the data integration when
a robot moves between special points with different sequences
of actions. We also present an algorithm for the movement of
a mobile robot between given points without loss of positioning
accuracy at long distances. The accuracy is not only affected the
Euclidean coordinates, but the orientation angle is also improved.
As a result, a robot can be used in narrow indoor, where a special
point can be reached only at a certain orientation angle.
I. INTRODUCTION
The problem of controlling the movement of mobile robot
is important. One of the biggest challenges facing researchers
is the creation of unmanned vehicles. Trajectory control and
vehicle navigation are particular tasks of creating unmanned
vehicles. Humanity needs not only unmanned vehicles, but
also unmanned robots that could autonomously perform some
useful operations. It is a global task for researchers who engage
robot navigation.
In recent years, many studies have been published on this
issue. There are also many classifications of the task. The task
can be classified by environment: indoor or outdoor; by type
of solution: algorithms and sensors; by the object to which the
solution is applied: stationary or mobile; by sensors that are
used: portable or stationary. Currently, there is no universal
way to solve this problem for all types of applications.
This article discusses the problem of positioning a mobile
robot indoors. The task can be described as follows: mobile
robot moves indoor between some point. Robot position can be
denoted as vector q=[xc,yc,θ] where xcand ycare coordinates
of center of mass and θis an orientation angle of robot. We
have n points in the room and for each of them robot has some
restrictions for coordinates. For example, a robot must get to
some point and have the right orientation to carry out some
operations. To do this, we need to accurately determine the
current coordinates of the robot and its orientation.
There is a similar task for tracking a mobile robot. If the
trajectory of the robot can be specified with absolute accuracy,
then it would be possible to determine the location of the robot
at any time. There are many researches of this task. They
use stability analysis [1], [2], [3] and neural networks [4].
Trajectory stability analysis is an effective method for prob-
lems in which there is a periodically repeating trajectory.
However, there are many tasks, including ours, in which the
trajectory can be arbitrary. In such tasks, those methods do
not achieve the best results. Neural networks require large
computing resources. At the same time, mobile robots are
usually equipped with cheap single board computers with few
computing resources.
For the trajectory construction problem, we consider the
integration of both data sources—sensed data come from
physical parameters sensors and from video cameras. Such
integration enables a novel way for trajectory construction.
The sensed data are collected from the IMU-sensor, camera,
and Lidar. The proposed architecture supports collecting data
from many sources. The collected data are used to evaluate
the position of a mobile robot. The results of our early
experiments confirm the feasibility of the data integration
when a robot moves between special points with different
sequences of actions. We also present an algorithm for the
movement of a mobile robot between given points without
loss of positioning accuracy at long distances. The accuracy is
not only affected the Euclidean coordinates, but the orientation
angle is also improved. As a result, a robot can be used in
narrow indoor, where a special point can be reached only at a
certain orientation angle.
The rest of the paper is organized as follows. Section II
overviews existing approaches to solving the trajectory con-
struction problem. Section III introduces our architecture that
supports trajectory construction in autonomous robot move-
ment. Section IV presents our theoretical algorithm that aims
at more effective trajectory construction. We define criteria to
evaluate the algorithm and our early experiments that confirm
the feasibility. Finally, Section V concludes the paper.
II. RELATED WORK
We analyzed articles by other authors and found several
works that are related to ours. In each of them, we found some-
thing important for our research and applied it. Currently, many
_
_____________________________________________________PROCEEDING OF THE 30TH CONFERENCE OF FRUCT ASSOCIATION
ISSN 2305-7254
researchers are experimenting with multisensory systems, as
they are more effective than single-sensor systems. However,
there are studies in which scientists are trying to improve the
performance of individual sensors.
The positioning problem has a number of related prob-
lems, such as: redundancy of data received from sensors [5];
selection of the most efficient data processing algorithm [6];
choosing a framework for data transfer [7], etc.
The task of mobile robot positioning also has many solu-
tions. In [8] researches used IMU-sensor with Lidar for better
positioning of mobile robot. Results of experiments showed
that complex system is more efficient for positioning than
IMU-sensor or Lidar separately.
In [9] researchers combined Lidar, odometer and IMU-
sensor with Kalman filter and used the idea of an external
and internal sensors with relative and absolute positioning.
In [10] researchers applied GPS, IMU, and visual odom-
etry. However, their study differs in that they did not work
indoors. They worked outdoor.
In [11] researchers also used multisensor approach for
mobile robot localization. They collected data from IMU,
Ultra-Wide Band (UWB) and Lidar technology. Sometimes it
leads to delays, which is unacceptable for the conditions of
our task.
In [12] researchers used data from IMU-sensor, optical
mouse sensor and wheel encoder and fuse it for input to
extended Kalman filter method. This allowed them to get an
error of less than 0.6% at a distance of 34 meters. IMU-sensors
are designed for accurate positioning, but they have one disad-
vantage: IMU-sensors accumulate measurement error. At the
same time, IMU-sensors have one important advantage. IMU-
sensors work with sufficient accuracy at short distances. Unlike
sensors, cameras work well over long distances. However, they
are not very effective at close range.
Video cameras in robotics often play a navigation role for
orienting a mobile robot in space and for recognizing various
obstacles on the path of movement. A low-resolution video
camera can be used to control a mobile smart robot in a
confined space. The authors [13] used images from an RGB-D
camera that were transmitted to the robot to detect and avoid
obstacles.
Machine vision for a static camera can be used to track the
position of the robot [14]. The most important tasks are robot
recognition, image for coordinate transformation, coordinates
of obstacles and other objects, background recognition and
subtraction (sky region, wall region, ground region, etc).
A video camera can be used on a mobile robot as an
additional component, for example, monitoring security on the
object [15]. In the most advanced version, the camera on the
robot can not only record video violations, but also identify
the offender by face, for example, in the case when it is an
employee working at the facility. Moreover, such solutions can
use a cheap video camera and inexpensive components of a
mobile robot, which will significantly save costs on security
at the facility.
In cases where global positioning is absent (for example,
when it is not possible to mount a static video camera), an
IMU-sensor or a mobile web camera can be added. There
are systems [16] that combine navigation data from sensors
and video cameras. Such approach can improve accuracy and
efficiency. In particular, data from a video camera should be
collected if the distance from the mobile robot to the nearest
object is too great, and the characteristics of the sensor or Lidar
are not sufficient to achieve this distance.
In [17] researches used new approach for calibration cam-
era. They used mirror and reflection of feature points. This
method gives high accuracy, but it can be applied only if we
have mirror surface indoor where robot works.
In [18] a new method of calibrating the camera and sensor
IMU is presented. The presented results make it possible to
reduce the error several times. However, the scope of this
method is very limited. You need to have an object directly in
front of the camera, to which the camera will be guided.
In [19] this study researches compare approaches for
navigation on mobile robots with IMU-sensors. In our work we
use them results from section “Odometry and Full IMU with
Motion Constraints”. They described math model for process
of moving mobile robot and show simulation results of them
methods. We plan to use some parts of that methods, but this
is not enough for our purposes and we want to modify it with
camera and LiDAR.
In [20] this study researchers compare different approaches
for mobile robot navigation. Difference between of approaches
is using the Kinect motion sensor. The data obtained with and
without it is compared. In our case, a different set of sensors is
used, so the methods of these researchers are not completely
suitable for us. However, we apply the general scheme of
research and the results of experiments with single sensors
and improved algorithms for our task.
In [21] the researchers made a series of experiments to
improve vehicle positioning based on MIMU. They used one
type of sensor because their goal was to create the most reliable
system that would work in almost any environment. In our
study, we use similar algorithms: pseudo-acceleration removal
procedure and TVU corrections. However, instead of MIMU,
we use IMU, so the obtained accuracy is not enough to be
limited to these methods.
In [16] researchers used the fusion data from camera and
onboard sensors of mobile robot. The researchers concluded
that no single sensor can provide sufficient accuracy. We also
use this idea in our research. However, this study has one
significant disadvantage. All experiments were carried out at
very small distances less than one meter. At such distances,
the error of the IMU-sensor does not have time to accumulate
and this significantly improves the result. Our study requires
a mobile robot to move long distances (up to several hundred
meters), so this method is not enough for this purpose.
In this study, we consider a comprehensive solution with
a new architecture in which the IMU-sensor, Lidar and the
camera complement each other. Existing studies are inves-
tigating positioning by location coordinates without paying
much attention to the orientation coordinate. However, in real
applications, when a robotic arm is installed on a mobile robot,
it is often necessary that the robot is not only at the desired
point, but also at the desired angle of rotation. Such robots are
_
_____________________________________________________PROCEEDING OF THE 30TH CONFERENCE OF FRUCT ASSOCIATION
---------------------------------------------------------------------------- 201 ----------------------------------------------------------------------------
often used at warehouses and sorting points to solve logistics
problems.
III. ARCHITECTURE
In this section we propose an event-driven approach based
on the generation of events at the periphery based on data
received from various sensors [22]. This approach is the most
optimal, given the large amount of information supplied to the
board of the mobile robot. Furthermore, this approach allows
the use of everyday mobile devices and sensors, for example,
web cameras [23].
First we need to define the functional roles involved in the
calculations. Such roles of handlers are described in Fig. 1.
Entities that are mediators of information are at different levels
of processing. The developer is encouraged to select data
handlers from the above example and highlight the functional
roles that these handlers should perform.
The data publication level consists of data collectors:
“collectors”—processors that receive data from the outside
(for example, video sensors or sensors), data extractors—
software modules that extract data from sensors, nodes - data
transmitters between other software objects, gateways - data
senders (for example, transmission over the Internet to another
communication site). In our example, the data collectors video
Fig. 1. Layered data architecture
Fig. 2. Edge computing architecture
sensor is represented as a Raspberry Pi video camera, and the
sensors are represented as IMU-sensor and LIDAR sensor. In
the simplest case, the Raspberry Pi can perform several tasks
at once, for example, extract, filter and formalize data coming
from heterogeneous sensors. Such operations do not require
complex processor calculations and can be implemented on a
low-performance computer (even on older versions of Rasp-
berry). As a Gateway, a router that operates at a frequency of
2.4 or 5 GHz can be used.
The processing layer consists of a data sink (it can be a
database or storage in the file system, both local and remote)
and a data interpreter (a server or broker that analyzes raw
data from the database, selecting only the necessary data, for
example, for a period or at regular intervals). In our example,
Recipient and Interpreter are represented by a mid-range
laptop. Basic calculations can be done on this laptop. However,
in the case of using algorithms that are complex in terms of
computations, an FPGA board or a neural accelerator can be
used. This is also recommended in the case of processing data
received from 2-3 mobile robots (using more Raspberry Pi).
Due to their compactness, neuroprocessors can be attached to
the mobile robot itself or connected via USB.
The semantic layer consists of a monitor that monitors
incoming (processed by the server) data and converts it into
simple events. A generator based on many simple events
creates complex events. A representative creates an interface
from complex events that are valuable from the user’s point of
view. In our example, Monitor is presented on both a Raspberry
and a laptop. Simple events that do not require additional
processing can be monitored on the on-board computer. In
this case, the processing layer (which usually requires complex
calculations) can be ”skipped” for some simple tasks. However,
complex events that are generated based on basic events
usually require the most powerful computing resources. Event
visualization can be deployed on a device that has an output
to the screen. Usually this is the screen of the same laptop,
however, in special cases it can be an on-board Raspberry Pi
display.
The presentation layer is made up of clients who view and
interact with events. Such a scheme is extensible and can be
supplemented with other functional roles, for example, in the
case when events consist of several levels of data and there
is a need to create more complex analyzing structures. In our
example, Client (the web interface) was deployed on a laptop.
For the web part, remote servers can be used with the ability
to connect globally from the outside (if the system is designed
for remote monitoring of events). However, it is not profitable
to deploy the web on a Raspberry Pi due to limited computing
resources.
In a view of the huge amounts of data and growing comput-
_
_____________________________________________________PROCEEDING OF THE 30TH CONFERENCE OF FRUCT ASSOCIATION
---------------------------------------------------------------------------- 202 ----------------------------------------------------------------------------
Fig. 3. The architecture of data processing on Raspberry Pi
ing power [24], it is necessary to imagine how the considered
data layers will be processed and stored on various computing
devices. Examples of configurations for such devices are as
follows.
•Edge computing: Sensor (= Video Sensor) - LAN
(Local Area Network) - all basic calculations are per-
formed on end devices, that is, “next to” the sensors.
•Fog computing: Sensor - LAN - DPC (Data Processing
Center) - calculations are performed using a decentral-
ized system (data center) that processes data “near” the
sensors.
•Cloud computing: Sensor - LAN - DPC - CP (Cloud
Platform) - calculations are performed in the cloud,
on the cloud platform, “far” from the sensors.
Video processing and image analysis are computationally
intensive. Small microcomputers like the Raspberry Pi or
Arduino do not have enough computing power to read and
process large amounts of sensor data. One of the solutions is
to transfer video analytics computations to the data processing
center (Fig. 2), but this algorithm has a large delay in the
mobile robot’s response to obstacles. In our solution based on
Edge-devices we use the local neuro-accelerator Google Coral
to reduce the load on the Raspberry Pi.
The architecture of data processing on Raspberry Pi is
shown in Fig. 3. Raspberry Pi camera, Lidar and IMU-
sensor send data to the Raspberry Pi microcomputer (which
is installed on the mobile robot). It can recognize patterns and
calculate distance to the near objects in the current working
area. Raspberry Pi then sends data using the ZeroMQ broker
that generates video streams on the ports and transmits it
to a more powerful computing device - the local PC. Such
PC implements all the necessary analytics. Depending on the
states, the event monitor, based on the processed data, gener-
ates events that an object was detected or current distance to
object and sends it back to Raspberry Pi and in the MongoDB
database for storage and subsequent retrieval as needed, and to
the RabbitMQ message broker to receive notifications on the
browser. Depending on the current statuses, a regular event or
a critical one will be sent. A critical event happened if the cart
going on the wrong direction or if insurmountable obstacle
was found.
IV. EXPERIMENTS
A. Mathematical model
Let us describe the original problem and the upcoming
experiments. We are faced with the task of testing an algorithm
that allows positioning the robot with high accuracy.
We have some kind of rectangular indoor in which there
are no thresholds. Also in this indoor it is always possible to
move in a straight line, there are no curvilinear trajectories. An
example of such a indoor is a warehouse with rows of shelving.
The mobile robot only moves between predetermined points.
Let’s denote it as special points. For each special point (SP),
the x, y and ?-rotation coordinates are known, with which the
robot should be at the point. For each special point i, deviations
of ?ix,?iy and ?iθ are allowed. The deviations could be either
the same or different for each special point. It depends on the
_
_____________________________________________________PROCEEDING OF THE 30TH CONFERENCE OF FRUCT ASSOCIATION
---------------------------------------------------------------------------- 203 ----------------------------------------------------------------------------
conditions of the specific problem for which the algorithm is
applied.
At the initial moment of time, the mobile robot is at its
parking lot. For definiteness, let’s denote this point as SP0with
coordinates x=0, y=0, ?=0. The input to the robot is a sequence
of special points that the robot must visit. It looks like array C
of coordinates where variables are numbers. For example, x1,
y1,θ1. Another array E contains variables of allowed deviation.
For example, ?1x,?1y,?1θ. When the robot arrives at a special
point, it compares its current coordinates xcur,ycur ,θcur with
the specified coordinate range xi±?ix,yi±?iy ,θi±?iθ , where i
is current special point. If the current coordinates are within the
specified range, then the robot continues to move in accordance
with the algorithm, otherwise it corrects its position.
The robot knows the coordinates of all the special points,
based on this information, the robot builds its path. For
definiteness, we will assume that the robot can move forward
or turn one of the angles: 90 degrees to the right, 180 degrees,
90 degrees to the left. We will also assume that the robot
spends 10 seconds at each particular point on its path, after
which it continues to move. This is necessary in order to
compare the absolute values of coordinates that are known
to researchers with the relative ones that the robot has.
B. Path Control and Positioning Algorithm
The main sensor, the readings from which will be the main
ones when orienting at a particular point, is the IMU sensor.
The main problem of such sensors is the accumulation of
errors, however, using the data of the camera and rangefinder,
one can reset the accumulated error. At short distances (less
than one meter), we plan to use the extended Kalman filter.
Over long distances, calibration will be carried out using a
camera and Lidar.
The URG-04LX-UG01 laser rangefinder allows one to
determine the distance to objects with a range of up to 4
meters. The measurement error is 3 percent of the distance
when the distance to the object is more than 1 meter and
0.03 meters when the distance is less than one meter. The
rangefinder allows you to measure distances to objects at a
speed of 10 measurements per second. Rangefinder data can
help at medium distances, when a feature point falls into the
camera’s dead zone when approaching it.
Before starting the movement, after receiving a sequence
of points, the algorithm calculates an approximate route of
movement. The route contains a sequence of movements and
turns. Each speed corresponds to the length of time during
which the mobile robot must move. Each turn corresponds to
the direction of rotation and the number of degrees that the
mobile robot must turn. Also, each action contains a motion
stage parameter. For example, moving from the base to the first
singular point is the first stage, from the first singular point to
the second is the second, and so on.
The main idea of the algorithm is as follows. We have some
marks that can be recognized by the camera image. The marks
are located on the walls at a certain height and distance from
the point on the floor that the mobile robot should occupy. If
the image from the camera contains the image of the mark,
then at large distances (more than 4 meters) it is possible to
Fig. 4. The movement of a mobile robot around its axis
analyze the video sequence and, on its basis, determine the
distance to the special point. If the mark is not visible on
the camera image, then the mobile robot moves along a pre-
calculated route, in accordance with the stage at which the
robot’s route is.
When the robot approaches a special point at a distance
of about 10-20 centimeters, it is necessary to calibrate the
IMU sensor. We need to define the parameters for position
and orientation. In order to determine the orientation, one can
take an image from the camera and analyze the position of
the robot relative to the mark. When the robot is directly in
front of the mark, the image must be exactly rectangular. If
the mobile robot is rotated with respect to the mark by a non-
zero angle, then the image of the mark will be curved. In
addition, one can get data on the distance to different ends of
the mark from the Lidar. By filtering camera and Lidar data
with extended Kalman filter, one can calibrate the IMU sensor
by orientation. Coordinate parameters are calibrated using the
same mechanism.
We used Raspberry Pi Camera v2 that allows to read
images with a Sony IMX 219 PQ sensor type, a resolution
of 1280x720 (sufficient for our experiment) and a light weight
of 3 grams. We used camera at an average distance (from
4 meters to 10 centimeters), a laser rangefinder becomes the
main sensor, which corrects the trolley on the approach to a
special point.
At the final stage of movement, the IMU-10dof becomes
the main sensor. From its work depends the correct position
of the mobile robot more than from other sensors, therefore
its accurate calibration is very important.
More formally, the algorithm can be described as follows:
1) Turning on the robot.
2) Built-in sensor calibration.
3) The robot receives a sequence of special points.
4) If the robot sees the next special point, then the path
is built based on the indicators of the sensors
5) Otherwise, the robot does not see the next special
point, the path is built on the basis of odometry.
6) The movement begins to the next special point
7) If at step 4 the robot did not see the special point,
then it checks every 3 seconds if the special point has
appeared in the visibility zone. If it appears, step 4.
8) At the entrance, the robot calibrates the data from the
sensors and checks the coordinates. If the coordinates
are in the allowable range, then the next special point
is sent to the robot, Step 4.
9) Otherwise Step 4 with the current special point.
10) If the path is complete, robot returns to base.
_
_____________________________________________________PROCEEDING OF THE 30TH CONFERENCE OF FRUCT ASSOCIATION
---------------------------------------------------------------------------- 204 ----------------------------------------------------------------------------
Fig. 5. Detecting walls and objects via camera on a mobile robot
During the operation of the algorithm, all data is transmitted
and processed as described in the Architecture section.
C. Early and Future Experiments
Let us describe the conditions for applying any algorithm,
as well as the practical needs in such an algorithm. In the early
experiments, we studied a mobile robot on which a manipulator
was installed (Fig. 4). The experiment consisted in making the
cart rotate 180 degrees around its axis. Since only the base
link was rigidly fixed to the manipulator, the inertia generated
by the manipulator was different all the time, since the rest of
the links began to move.
This problem can be solved by holding a certain position of
the links with the help of servomotors. However, this leads to
a huge increase in battery consumption as well as overheating
of the servos. Because of this, in practice, such a solution is
not feasible, at least for cheap robots.
The second possible solution to the problem of changing
inertia from the movement of the manipulator links is the
development of a mathematical model of inertia change using
dynamic manipulation methods. If it is possible to build a
model of inertia change, then it will be possible to find a
control action that will compensate for the resulting inertia.
A similar solution is planned to be investigated in the future.
In particular, at the first steps of the experiment, using
machine vision technologies and OpenCV algorithms, we
tried to recognize objects (doors, walls) that arise during the
movement of the cart in the corridor of an office building
shown in Fig. 5. The approximate recognition algorithm was:
•Image translation to GrayScale with Canny border
detector, noise smoothing;
•Calibration of the camera (some of it require mea-
surement of objects in ideal conditions: doors, door
handles, scale ratio);
•Detection of boundaries of space in real time, building
a diagram of the space (walls / doors / ceiling);
•Determining the distance to the wall;
•Determination of the trajectory of movement (where
we can or cannot go).
Another problem found in our early experiments is the
friction of the wheels of the mobile robot against the floor
surface. A mobile robot, especially if it contains additional
equipment, has a minimum value of the force that must be
applied for the robot to start moving. However, if the friction
with the surface is large, then the minimum force becomes
too large to accurately control the robot. This leads to the fact
that it is impossible to predict the behavior of the robot in a
particular situation.
In the first version of the research, we used IMU-10DOF
and 2 cameras: one was external and static, the second was
fixed on a mobile robot and moved with it. During the
experiments, we tried to achieve two goals: 1) turn the mobile
robot with high accuracy; 2) correct the movement of the
mobile robot so that it moves in a straight line.
The rotation of the mobile robot could not be standardized
due to the reasons described above. However, we tried to use
another mobile robot with the most robust design and a fixed
Raspberry Pi camera on board.
To achieve the second goal, we conducted a series of ex-
periments, during which two types of tuning were performed:
mechanical and software. Mechanical adjustment included
changing the position of the camera and battery. This made
it possible to change the center of mass of the mobile robot
and smooth out the resulting inertia.
In addition, we received data from the IMU sensor and
camera. As a result of the analysis of the data obtained, using
the Kalman filter, it was possible to achieve a deflection of
the mobile robot by no more than 3 degrees with a rectilinear
movement at a distance of 10 meters.
At the next stage of experiments, we plan to modify
several technological aspects: modification of a mobile robot
(the most efficient arrangement of components - a battery, a
camera, a microcomputer), an increase in the performance
of the Raspberry Pi - since the computational capabilities
of a microcomputer processor are not enough for object
recognition, it is planned to integrate part of the calculations
on a portable neuro accelerator (for example, Google Coral);
improvement of the image segmentation algorithm (the best
separation of the wall and the floor), the best noise filtering
algorithms for the IMU-sensor.
Based on our early experiments, we put forward the fol-
lowing hypotheses.
1) If the deviation from the required coordinates is
small (within a few tens of centimeters), then the
mobile robot will not be able to make the necessary
movement for correction. It is necessary to drive a
couple of meters back/forward and repeat the algo-
rithm again.
2) The given set of sensors is capable of effectively
solving problems of autonomous movement.
3) The given set of sensors is capable of solving prob-
lems in rooms where there are dynamic obstacles.
Corresponding experiments are also planned:
1) Comparison of attempts to correct the position of the
robot at short distances locally and using a new run
of the algorithm.
2) A series of runs of the algorithm for a different set
of singular points and input deviation parameters to
find out the accuracy of the solution.
3) The same as in 2, but on condition that there are
people in the room.
_
_____________________________________________________PROCEEDING OF THE 30TH CONFERENCE OF FRUCT ASSOCIATION
---------------------------------------------------------------------------- 205 ----------------------------------------------------------------------------
V. C ONCLUSION
This paper considered the problem of trajectory construc-
tion based on positioning a mobile robot. We considered the
tasks of navigating a mobile robot when moving between
special points. Our early experiments used a camera built into
a mobile robot, an inertial sensor and a Lidar. Our goal is
that the sensed data are used in integration. At long distances
(over 4 meters) the camera works best, at medium distances
(from 10 centimeters to 4 meters) the Lidar works best, at
short distances it is necessary to use an IMU sensor.
Our primary contribution is the proposed architecture for
trajectory construction from the integrated sensed data. Our
early experiments to evaluate the architecture were performed
without a lidar. The results show the possibility to set up the
linear motion of a mobile robot in various conditions.
We formulated further requirements for the design of
mobile robots. In particular, robots should not have a large
mass and friction of the wheel against the surface at the same
time. In addition, the equipment that is on the robot must either
be stationary or not affect the center of mass of the mobile
robot. Data received from IMU, camera, and Lidar. Our plans
are to carry out experiments with sensed data from all three
sensors, processing them with an extended Kalman filter, and
calibrating the values relative to each other.
ACKNOWLEDGMENT
This development was implemented in Petrozavodsk State
University (PetrSU) with financial support of the Ministry of
Science and Higher Education of Russia within Agreement
no. 075-11-2019-088 of 20.12.2019 on the topic “Creating
the high-tech production of mobile microprocessor computing
modules based on SiP and PoP technology for smart data
collection, mining, and interaction with surrounding sources”.
This research study is supported by RFBR (research project
# 19-07-01027) in part of edge-centric computing models for
data processing. The work is in collaboration with the Artificial
Intelligence Center of PetrSU.
REFERENCES
[1] R. Anushree and B. S. Prasad, “Design and development of novel
control strategy for trajectory tracking of mobile robot: Featured with
tracking error minimization,” in 2016 IEEE Annual India Conference
(INDICON), 2016, pp. 1–6.
[2] A. Y. Lee, G. Jang, and Y. Choi, “Infinitely differentiable and continuous
trajectory planning for mobile robot control,” in 2013 10th International
Conference on Ubiquitous Robots and Ambient Intelligence (URAI),
2013, pp. 357–361.
[3] X.-Z. Jin, J.-Z. Yu, L. Zhou, and Y.-Y. Zheng, “Robust adaptive
trajectory tracking control of mobile robots with actuator faults,” in
2019 Chinese Control And Decision Conference (CCDC), 2019, pp.
2691–2695.
[4] M. Asai, G. Chen, and I. Takami, “Neural network trajectory tracking
of tracked mobile robot,” in 2019 16th International Multi-Conference
on Systems, Signals Devices (SSD), 2019, pp. 225–230.
[5] K. Krinkin and A. Filatov, “Correlation filter of 2d laser scans for
indoor environment,” Robotics and Autonomous Systems, vol. 142,
p. 103809, 2021. [Online]. Available: https://www.sciencedirect.com/
science/article/pii/S0921889021000944
[6] A. Huletski, D. Kartashov, and K. Krinkin, “Vinyslam: An indoor slam
method for low-cost platforms based on the transferable belief model,”
in 2017 IEEE/RSJ International Conference on Intelligent Robots and
Systems (IROS), 2017, pp. 6770–6776.
[7] A. Gavrilov, M. Bergaliyev, S. Tinyakov, and K. Krinkin, “Analysis of
robotic platforms: Data transfer performance evaluation,” pp. 437–443,
2021.
[8] D. R.-Y. Phang, W.-K. Lee, N. Matsuhira, and P. Michail, “Enhanced
mobile robot localization with lidar and imu sensor,” in 2019 IEEE In-
ternational Meeting for Future of Electron Devices, Kansai (IMFEDK),
2019, pp. 71–72.
[9] Z. Li, Z. Su, and T. Yang, “Design of intelligent mobile robot posi-
tioning algorithm based on imu/odometer/lidar,” in 2019 International
Conference on Sensing, Diagnostics, Prognostics, and Control (SDPC),
2019, pp. 627–631.
[10] G.-S. Cai, H.-Y. Lin, and S.-F. Kao, “Mobile robot localization using
gps, imu and visual odometry,” pp. 1–6, 2019.
[11] Y. Gao, F. Wang, J. Li, and Y. Liu, “Localization of mobile robot based
on multi-sensor fusion,” pp. 4367–4372, 2020.
[12] Y. Liu, Y. Ou, and W. Han, “Mobile robot localization based on
optical sensor,” in 2019 IEEE International Conference on Real-time
Computing and Robotics (RCAR), 2019, pp. 874–879.
[13] S. Kebir, H. Kheddar, M. Maazouz, S. Mekaoui, A. Ferrah, and
R. Mazari, “Smart robot navigation using rgb-d camera,” in 2018
International Conference on Applied Smart Systems (ICASS), 2018, pp.
1–6.
[14] R. J. C. Saclolo, J. B. Delos Reyes, N. N. R. B. Formeloza, J. I. O.
Abaca, K. K. Serrano, A. Dela Cruz, E. Roxas, and R. R. Vicerra,
“Machine vision system, robot hardware design and fuzzy controller
design for autonomous multi-agent mobile robot platform,” in 2015
International Conference on Humanoid, Nanotechnology, Information
Technology,Communication and Control, Environment and Management
(HNICEM), 2015, pp. 1–6.
[15] F. Khalid, I. H. Albab, D. Roy, A. P. Asif, and K. Shikder, “Night
patrolling robot,” in 2021 2nd International Conference on Robotics,
Electrical and Signal Processing Techniques (ICREST), 2021, pp. 377–
382.
[16] A. Novitsky and D. Yukhimets, “The navigation method of wheeled
mobile robot based on data fusion obtained from onboard sensors and
camera,” in 2015 15th International Conference on Control, Automation
and Systems (ICCAS), 2015, pp. 574–579.
[17] G. Panahandeh and M. Jansson, “Imu-camera self-calibration using
planar mirror reflection,” in 2011 International Conference on Indoor
Positioning and Indoor Navigation, 2011, pp. 1–7.
[18] X. Ouyang, Y. Shi, You, and Zhao, “Extrinsic parameter calibration
method for a visual/inertial integrated system with a predefined me-
chanical interface,” Sensors, vol. 19, p. 3086, 07 2019.
[19] T. Fauser, S. Bruder, and A. El-Osery, “A comparison of inertial-based
navigation algorithms for a low-cost indoor mobile robot,” in 2017 12th
International Conference on Computer Science and Education (ICCSE),
2017, pp. 101–106.
[20] H. G. Y. X. Mingjing Gao, Min Yu, “Mobile robot indoor positioning
based on a combination of visual and inertial sensors,” Information
Systems Frontiers,vol. 19, Apr. 2019.
[21] A. Mikov, A. Moschevikin, and R. Voronov, “Vehicle dead-reckoning
autonomous algorithm based on turn velocity updates in kalman filter,”
in 2020 27th Saint Petersburg International Conference on Integrated
Navigation Systems (ICINS), 2020, pp. 1–5.
[22] N. Bazhenov and D. Korzun, “Event-driven video services for moni-
toring in edge-centric internet of things environments,” in Proc. 25th
Conf. Open Innovations Association FRUCT, Nov. 2019, pp. 47–56.
[23] N. A. Bazhenov and D. G. Korzun, “Use of everyday mobile video
cameras in IoT applications,” in Proc. 22nd Conf. Open Innovations
Association FRUCT, May 2018, pp. 344–350.
[24] X. Wang, Y. Han, V. C. M. Leung, D. Niyato, X. Yan, and X. Chen,
“Convergence of edge computing and deep learning: A comprehensive
survey,” IEEE Communications Surveys Tutorials, vol. 22, no. 2, pp.
869–904, 2020.
_
_____________________________________________________PROCEEDING OF THE 30TH CONFERENCE OF FRUCT ASSOCIATION
---------------------------------------------------------------------------- 206 ----------------------------------------------------------------------------
