RÉPUBLIQUE ALGÉRIENNE DÉMOCRATIQUE ET

POPULAIRE
MINISTÈRE DE L'ENSEIGNEMENT SUPÉRIEUR ET DE LA
RECHERCHE SCIENTIFIQUE

Université SAAD
DAHLEB Blida1

Département
D’électronique
Faculté de
Technologie
Master2
Electronique Des
Systèmes Embarqués

Following Methods in Robotics

Prepared by:

 Guesmia Houssam
 Menad Meriem
 Academic Year: 2025/2026
1. Introduction:
Human-following mobile robots are designed to autonomously track, follow and assist a
designated individual in a variety of environments. These systems integrate advanced
sensor technologies, such as cameras, LiDAR, and range finders, with robust algorithms to
ensure safety, adaptability, and seamless human–robot interaction. By harnessing sensor
fusion techniques and machine learning, these robots can operate in both structured and
unstructured settings, ranging from industrial production floors to care environments for
the elderly. Recent developments have focused on enhancing the precision of target
tracking, obstacle avoidance and the ability to function in mixed environments where prior
mapping may be partially available.
A wide range of human-robot collaborative applications in diverse domains such as
manufacturing, health care, the entertainment industry, and social interactions, requires an
autonomous robot to follow its human companion. that what makes human following
robots a highly requested need from helping and assisting humans to Accomplishing tasks
that are hard for humans to accomplish

2. Classification of Following Methods:

Robots use multiple ways to provide data to the robot itself to treat it and start following
the target, multiple sensors and tools could be used, we saw that we can classify it in 4 main
categories based on the type of sensor used to estimate the relative position of the target:

 Vision-based following:
 Beacon/tag-based following:
 Range-based following:
 Cooperative multi-robot following:

3. Following Methods:
3.1. Vision-Based Following:
3.1.1. Fiducial marker:
Fiducials are artificial visual features designed for automatic detection, and often carry a
unique payload to make them distinguishable from each other. Although these types of
fiducials were first developed and popularized by augmented reality applications, they have
since been widely adopted by the robotics community. Their uses range from ground
truthing to object detection and tracking.
The Ubiquity Robotics localization system uses a number of fiducial markers of known size.
Detection of the markers is done by using the robot’s camera. When fiducial markers are
attached to a human or robot, their detected pose can be used as a following target. The
characteristics of the images of the fiducial markers enable the robot to compute its
location.
 In our launch files we will have a node to detect Aruco fiducials from the camera images.
The output of this autodetect node goes to two topics where one topic gives the verticies of
each detected fiducial and another topic publishes the transforms of each fiducial relative
to the camera itself. We have other nodes that subscribe to the aruco_detect topics to then
do the math to sort out where each fiducial is relative to the robot itself using 3D math and
form the map or use the map to navigate.

3.1.2. Deep Learning Detection:

AI empowers robots to process data from sensors, analyze their surroundings, and make
autonomous decisions. This is particularly critical in autonomous mobile robots, which must
navigate dynamic indoor and outdoor environments.
Machine learning enables robots to analyze specific data and develop customized solutions.
For example: Learning traffic patterns and optimizing warehouse processes for better
efficiency or using historical data to predict and prevent equipment failures.
the process starts when A neural network detects a person or robot in each camera frame
and outputs a bounding box. The detected box is tracked over time which make the robot
can follows the same human/robot. and the Distance and direction are estimated from
bounding-box size. using this information the robot adjusts its speed and direction to keep
the target centered and at a fixed distance.
HuskyLens: HuskyLens is designed to be smart. Built-in algorithms allow HuskyLens to
learn new objects, faces, and colors through a single click. Moreover, by long pressing,
HuskyLens can continually learn them even from different perspectives and in various
ranges. The more it learns, the more accurate it can be. Those AI algorithms outfit the
camera with facial recognition, object recognition/tracking, color detection, line following,
and tag detection. Users can also select different algorithms by pressing some buttons, so
even people with limited understanding of AI can easily use the camera.

3.1.3. RGB-D camera + Color-Based Segmentation:

We can reach human or robot following using vision and without ai using RGB depth
camera, which is a type of camera that provides both depth (D) and color (RGB) data as the
output in real-time. Depth information is retrievable through a depth map/image which is
created by a 3D depth sensor such as a stereo sensor or time of flight sensor. RGBD
cameras are able to do a pixel-to-pixel merging of RGB data and depth information to
deliver both in a single frame. it exists some types of RGBD camera that has a built-in
algorithm like The Microsoft Kinect that provides the auto detection of human skeleton and
the movement which is a very useful argument in robotics for human following.
The target is extracted using color thresholding, shape/size constraints and motion
detection, Once the target region is segmented, the depth map gives the direct distance to
the target and the robot follows by regulating distance and orientation.

3.1.4. Thermal Person Tracking:

Thermal cameras coupled with deep neural networks are a much more robust strategy to
actually detect the presence of people. Unlike motion sensors, they will detect the presence
of people even when they aren’t moving. And, unlike optical cameras, they detect bodies by
measuring the heat that they emit in the form of infrared radiation, and are therefore much
more robust — their sensitivity doesn’t depend on lighting conditions, on the position of
the target, or the color. it is mostly used with drones and very developed projects.
 the issue with this method is that it is very expensive hardware and not available much.
this method should also combine a thermal camera output with an AI model to treat the
images and use mathematical and other algorithms to give specific commands for
movement.

3.1.5. LiDAR-based following:

LiDAR sensors are efficient for the applications of following, especially for applications that
detect and follow people while maintaining a suitable distance, at a close range, using
accurate range measurements.
Laser range finder (LRF)-based human tracking has several advantages over methods that
use other sensors. Measurements using LRFs are more robust to lighting conditions, more
accurate than camera-based methods, and require less computation than camera-based
measurements.
A 3D LiDAR observes the environment in multiple planes. Since a 3D LiDAR obtains more
information than an LRF, it is expected to enable more robust and accurate human tracking.
To perform person following, a single human track is selected as input to an Object
Following Controller (OFC). The track can be manually selected by the robot operator, or
automatically selected using a predefined decision criterion, such as picking the human
closest to the robot. The OFC uses the track’s latest state estimate to compute the
difference (position error) from the desired position of the human with respect to the robot
(position goal). The position error is then used to compute the robot’s velocity setpoints for
low level actuator controllers. The motion generated decreases the position error over time
and results in a trajectory that causes the robot to follow the human.
To minimize the position error of the tracked human, the OFC modulates the robot’s
angular and linear velocity setpoints independently. Two vectors are used for these
calculations: a vector from the robot’s center to the goal position, and a vector from the
robot’s center to the human position. The first vector is expected to be constant (as long as
the goal position does not change), while the second vector changes as the human and
robot move with respect to each other. The control actions aim to equalize the length of
these vectors, and drive the angle between them to zero. A Proportional-Integral-Derivative
(PID) controller was implemented to calculate the angular velocity setpoint using the angle
between the vectors, while a second PID controller calculates the linear velocity setpoint
using the difference in lengths of the vectors. Both controllers were tuned using classical
Ziegler-Nichols method. A dead-band zone was defined to address vibrations in the control
actions when the position error is too small.

3.2. Beacon-Based Following:

3.2.1. UWB Tag + Anchors (RTLS):
Ultra-wideband systems have emerged as a promising technology in the field of position
monitoring, especially for moving objects within a closed indoor environment. The quality of
the pulses issued by ultra-wideband transmitters is characterized by high temporal accuracy,
which makes it possible to measure the transmission and reception time (time of flight) easily,
enabling high precision in determining the location. The accuracy can vary from a few
decimeters to centimeters in optimal conditions. Most position detection systems usually
consist of two main components, which may be identical in terms of design and physical
components but different in terms of the method of use within the system: the anchor, and the
tag.
In most of the current systems, anchors are used as receivers for the signal issued by the tag,
and to determine the distance between the anchor and tags and then forward the
measurement to the IPS controller. Numerous anchors are used, depending on the area of the
place to be covered.
Several techniques can be used with ultra-wideband waves to determine the indoor position
of movable objects. In general, the UWB location tracking of the unknown movable tag can be
determined using one of the following basic techniques:

 Triangulation Technique:
The triangulation technique is based on calculating angles. The location of a moving tag
can be identified if the location of two fixed anchors is known, by estimating the angles
between the movable tag and the antennas of the fixed anchors and intersecting the
extension of the moving tag with the known line between the two fixed anchors, which
enables the accurate determination of the location of the moving tag.
  Trilateration Technique:
The trilateration technique depends on calculating the distances between the moving tag
and the fixed anchors. The distance is calculated by multiplying the time of wave
propagation by the speed of the UWB wave, which is approximately (3 × 10 8 m/s). If the
coordinates of the used anchors (which are predefined by the user) and the distances
between the movable objects and anchors can be provided by the used technology, then
we can find and track the object location using trilateration techniques.
The distance d between any two points can be calculated with the known coordinates

𝑑=√(𝑥2−𝑥1)2+(𝑦2−𝑦1)2
(x, y) of the two points; for example, the distance between (x1, y1) and (x2, y2) is:

3.2.2. Inter-Robot UWB Ranging:

To reduce localization error, we could complement a traditional feature-based image
tracking with ultra-wideband (UWB) distance measurements between the robots. The
robots use an advanced mesh-ranging protocol that allows them to continuously share
distance measurements amongst each other rather than relying on the common "anchor"
and "tag" UWB architecture.
Unlike RTLS systems inter-robot UWB ranging focuses on relative localization. The distance
between robots is obtained using UWB time-of-flight measurements, exploiting the high
temporal resolution of UWB signals. With this method we can achieve a high-level accuracy
that makes the approach suitable for robot–robot following.

3.2.3. BLE/ Bluetooth beacon:

The Bluetooth component of the system was designed to offer an infrastructure-free
method of localization using Angle of Arrival (AoA) and Received Signal Strength Indicator
(RSSI) analysis. The setup featured a Silicon Labs BG22 Bluetooth antenna array mounted on
the robot and a BLE-emitting tag carried by the user. By measuring the phase of incoming
signals from a Constant Tone Extension (CTE) in Bluetooth packets, the system estimated
the direction of the tag relative to the robot. This directional information was combined
with signal strength data to estimate the distance to the tag, effectively approximate
relative direction and distance.
However, this Bluetooth-based tracking system proved unreliable in practice. The AoA
method, though theoretically capable of sub-degree resolution, suffered from high noise
levels and poor accuracy in real-world conditions.

3.2.4. RFID /WI-FI beacon:

WIFI and RFID tags following methods are not the best following methods in robotics, like
the Bluetooth case they are coarse, proximity-based, not continuous tracking like UWB or
LiDAR.
The human or leader robot carries the tag and the follower robot carries an RFID/WIFI
reader and antennas.
When the tag enters the reader’s range, the robot knows if the target is nearby, in front,
left or right if multiple antennas are used.
Signal strength is used to approximate distance. and the commands are given to the robot
in order to maximize the RSSI (Received Signal Strength Indicator) By switching between
antennas, the robot estimates which direction produces the strongest response.

3.2.5. Light Beacon:

The most common indoor positioning systems are based on radio frequency technologies
like Wi-Fi, BLE, ultra-wide band (UWB), and radio frequency identification (RFID). However,
their relatively low localization accuracy means that they are more suited as a secondary
source of information for sensor fusion-based navigation. In contrast technologies which
offer centimeter-range indoor positioning accuracy such as acoustic signals and visible light
signals hold more promise for robotic navigation. Visible light positioning (VLP) has the
comparative advantage of being able to leverage pre-existing lighting infrastructure. Light-
emitting diodes (LEDs), which are energy efficient, long lasting, and emit low heat, are
rapidly becoming the preferred technology for modern indoor lighting infrastructure. The
abundance of pre-installed LED lightings presents opportunity to simultaneously use them
for visible light Positioning.
In VLP systems, LED luminaires are used as the transmitting beacons with either a photo
diode (PD) or camera as the common receiver sensor attached to the tracked object.
The LED luminaires act as optical beacons by transmitting uniquely identifiable light signals
with high-frequency, each light source can encode an identifier or positional information.
A photodiode or camera mounted on the robot receives the emitted light signals and
detects their intensity, angle, or image position. Using this information, the robot estimates
the relative position or direction of the light beacon.

3.2.6. Infrared (IR) Beacon:

The Pololu IR beacon is a compact board that is used in pairs to allow robots to locate each
other. Each board has infrared emitters that shine in all directions and four IR receivers for
detecting the other beacon. The IR beacons have a range of about fifteen feet indoors. They
meant to be used in pairs to give autonomous robots simple means for detecting each
other.
The beacons work by transmitting and detecting infrared light, much like a television
remote control. Each beacon has six IR emitters and four IR detectors. The beacons
alternate between transmitting and receiving so that they never get confused by reflections
of their own transmissions. The transmit and detect cycle is carried out more than one
thousand times per second, and a small microcontroller monitors all four detectors to
decide the direction to the other beacon. The beacons have four red LEDs that indicate the
direction to the other beacon; if you take two beacons and rotate them, the LEDs will
always keep lighting up in the direction of the other beacon.

3.2.7. Ultrasound Beacon:

Ultrasound beacons are mounted on the ceiling to provide consistent coverage and
minimize obstructions. The mobile robot, equipped with ultrasound receivers, moves within
the area covered by the beacons. This configuration allows for accurate positioning of
robots with varying heights, as the system can account for the vertical offset between the
beacons and the robot-mounted receivers.
The system operates by firing the beacons in pairs according to a pre-determined
sequence. Each pair of beacons emits an ultrasound signal simultaneously. This paired
emission strategy allows for relative time measurements, reducing the complexity and
potential errors associated with absolute time synchronization. By using pairs of beacons,
the system can create distinct time differences that are crucial for the localization process.

differences (Δ⁢𝑡) between the arrivals of these signals at each receiver. These time
As the signals travel from the beacons to the receivers, the system measures the time

differences are then used to calculate the relative distances between the beacons and the
robot, based on the known speed of sound in the medium. By using time differences rather
than absolute times, the need for precise clock synchronization between beacons and
receivers is eliminated.
By analyzing the time differences from multiple pairs of beacons, the system can
triangulate the precise position of the robot. The calculation utilizes hyperbolic
multilateration techniques, where each time difference measurement defines a hyperbolic
curve on which the robot must lie. The intersection of these curves, derived from multiple
beacon pairs, pinpoints the robot’s position with high accuracy.
This approach is mainly designed for indoor positioning. Its application to robot following
is limited to slow-moving targets and controlled environments due to latency and reflection
sensitivity.

3.3. Audio/ Sound Following:

A robot can record sound in stereo using two microphones. This works because sound
waves travel radially outward from the sound source. As the waves expand outward, they
get quieter. This means that the sound will seem louder to the microphone closer to the
sound source. The sound will also reach that microphone slightly before it reaches the
farther microphone.
 There are different ways to process the sound recorded by multiple microphones in order
to determine the direction of the source. One method is to compare the amplitudes of the
two sound waves and assume that the louder wave was detected by the microphone closer
to the source. If you can record data fast enough, you can also check to see which
microphone recorded a sound earlier.

3.4. Cooperative Multi-Robot Following:

Follower approach one of the preferred methods for coordination of robots as it is a
centralized control method which provide stability and tracking performance and known for
its simplicity. the leader robot is equipped with needed sensors to detect obstacles and to
avoid collisions were the follower robot get all the information from the leader robot via
Wi-Fi using DDS publish–subscribe communication in ROS2.
In the first step of the communication protocol, the leader agent has to establish a reliable
communication with its followers in the system in order to control their actions and
movements. Then, the leader should be aware of any arising problems in the environment
such as communication loss or obstacle detection and avoidance techniques. The leader can
initiate the communication with the followers by broadcasting a request message over
wireless transmission. The leader message has many purposes: Send a command to a
specific follower or broadcast a command to all followers in range. Implemented commands
so far are: move forward, backward, turn right, left and stop.

4. Supporting Techniques for Robot Following:

4.1. Kalman filters:
The KF (Kalman filter) developed by Rudolf E. Kalman in 1960 is an algorithm for the
estimation of non-observable state variables based on observable variables that may have
some measurement error.
The EKF extension is achieved using Taylor’s approximation, through which a nonlinear
system can be linearized employing estimation techniques. This algorithm is a very powerful
tool for positioning: it can accept different types of data, solve a large number of
parameters, and produce reliable and accurate results. Both the KF and the EKF are used in
robotics, where they are applied in trajectory tracking, position estimation for manipulator
robots, SLAM, and object detection, among others, depending on the linearity or
nonlinearity of the model. The flexibility of this algorithm has enabled the integration of
information from different types of sensors and techniques such as odometry, GNSS (Global
Navigation Satellite System), laser and ultrasonic sensors, and recently, artificial vision,
 making it possible to answer the fundamental questions of autonomous navigation: Where
am I? Where am I going? And how do I reach my destination?

4.2. VIO (Visual + IMU Odometry):

Visual Inertial Odometry (VIO) is an algorithm that integrates camera and Inertial
Measurement Unit (IMU) data to achieve robot localization. The VIO positioning algorithm
has the advantages of low cost and wide applicability, and can effectively compensate for
failure scenarios such as occlusion and multipath interference in satellite positioning in
outdoor environments. An excellent and robust VIO algorithm is key to achieving high-
precision outdoor navigation positioning.
Distinct features like corners or edges are detected in camera images and tracked to
estimate relative motion. The IMU measurements (acceleration and angular velocity) are
integrated to predict short-term motion between images or when vision stop.
VIO is mainly used for improving motion estimation and stability when combined with
target detection sensors.

5. Comparative Analysis of Following Techniques:

This table represent all the mentioned methods and the Strengths and weaknesses for
each method in different conditions and contexts:

Method Accuracy Range Robustness Line-of- Cost Indoor/ Real-World Human/Robot

Sight Outdoor Adoption
Required
Fiducial Marker high mid mid yes low In/out high both

DL-Based Visual mid mid mid yes mid In/out high both

RGB-D Camera high low mid yes mid/ Indoor high both
high
Thermal camera mid mid mid yes very In/out mid Human
high
LiDAR high high high no mid/ In/out Very high both
high
UWB (RTLS) Very high Very high no mid/ In/out Very high both
high high
Inter-Robot UWB high mid high no high In/out high Robot

BLE / Bluetooth low mid low no low In/out low both

RFID / Wi-Fi low mid low no low In/out low both

Light Beacon low mid Very low yes low Indoor Very low both
IR low low Very low yes low indoor Very low both
Audio / Sound low low low no low indoor low Human
Cooperative multi- mid mid mid no mid In/out high Robot
robot

Most of the mentioned methods are used in the real life and we could rely on it in the
purpose of following, but some other methods are not really recommended to use
according to the lack of efficiency and most of the researches are for small projects.

6. Conclusion:
in robotics, following robots are a fundamental skill in the student level projects and also
in some high-level projects in robotics in industrial or agriculture and many others. In this
work, a wide review of the main following methods was presented, covering all the main
categories and strategies used in real life and some other technics that are not ideal for
these tasks but still could be used with more research and deep understanding of the used
material.
Each method relies on different sensing principles and offers distinct advantages and
limitations in terms of accuracy, robustness and others.
at the end we could say that no single method is optimal. the choice of a following
approach depends on the application context.

References:
-[Link]
mobile-robots-micro-194101
-[Link]
-[Link]
vision-sensor-89916c
-[Link]
-[Link]
-[Link]
-[Link]
-[Link]
-[Link]
-[Link]
[Link]
%20for%20a%20single%20unit
-[Link]
-[Link]
camera-and-machine-learning-376d3bbcd45c
-[Link]
-[Link]
navigation