RGB-D Human Detection and Tracking for

Industrial Environments

Matteo Munaro1 , Christopher Lewis2 David Chambers2 , Paul Hvass2 , and

Emanuele Menegatti1
1
Department of Information Engineering, University of Padova, Via Gradenigo 6B,
35131 - Padova, Italy,
munaro@dei.unipd.it, emg@dei.unipd.it,
2
Southwest Research Institute, San Antonio, TX 78238, USA,
christopher.lewis@swri.org, david.chambers@swri.org, paul.hvass@swri.org

Abstract. Reliably detecting and tracking movements of nearby work-

ers on the factory floor is crucial to the safety of advanced manufactur-
ing automation in which humans and robots share the same workspace.
In this work, we address the problem of multiple people detection and
tracking in industrial environments by proposing algorithms which ex-
ploit both color and depth data to robustly track people in real-time. For
people detection, a cascade organization of these algorithms is proposed,
while tracking is performed based on a particle filter which can inter-
polate sparse detection results by exploiting color histograms of people.
Tracking results of different combinations of the proposed methods are
evaluated on a novel dataset collected with a consumer RGB-D sensor
in an industrial-like environment. Our techniques obtain good tracking
performances even in an industrial setting and reach more than 30 Hz
update rate. All these algorithms have been released as open source as
part of the ROS-Industrial project.

Keywords: human detection and tracking, ROS-Industrial, RGB-D,

open source

1 Introduction

The next generation of robots, both for service or cooperative work, is expected
to interact with people more directly than today. Industry is more and more
interested in exploiting both the dexterity and versatility of people and the
precision and repeatability of robots by enabling collaboration in dynamic and
reconfigurable manufacturing environments. Such collaborations, however, are
not yet possible because robots are still not capable of safely interacting coop-
eratively with their human coworkers in highly variable task scenarios. In most
cases, safety of the users interacting with industrial robot manipulators has been
addressed by using safety guards or other barriers to cordon robots off from peo-
ple. This assumption cannot be considered if humans and robots have to share
the physical environment or collaborate. In that case, people have to be detected
2

in order to prevent collisions with the robot. Moreover, if people are not only
detected, but their movements are tracked over time, their intentions and fu-
ture positions can be estimated. Based on this high-level information, the robot
can choose the strategy to avoid collision which could minimize the time needed
to resume operations, thus reducing productivity slowdowns. In this work, we
propose methods for detecting and tracking people by means of vision. This
approach allows to avoid the costs and hazards involved with perimeter fencing.
People detection and tracking have been deeply studied by the computer vi-
sion and robotics communities. However, unlike video surveillance or service
robotics scenarios, people tracking for human-robot interaction in industrial
workspaces requires to meet a certain number of constraints which should guar-
antee the safety of workers. Among these, all people within the workspace should
be detected, even if partially occluded, and their position should be sent to the
robot control system as soon as they enter the operating area. To meet the latter
constraint, the system update rate should be high (from 15 to 30 Hz) and the
latency should be minimum (< 0.2 s). It is worth noting that system reactivity is
not necessarily related to overall tracking accuracy. Finally, it would be desirable
if the system could also work when the camera is placed onboard of a moving
vehicle or horizontally translated on a track to follow a person while moving in a
wide area. This assumption requires to avoid people detection techniques based
on background subtraction in the color or depth image.
Algorithms which rely on 2D images ([8], [3]) are usually too slow and sensi-
tive to clutter and occlusion. Thus, depth-based approaches exploiting passive or
active sensors are usually preferred. Passive sensors, such as stereo cameras ([9],
[10]), have the need for finding correspondences between left and right image
points, which is a computationally expensive operation and it can fail for scenes
where texture is poor. For these reasons, an active sensor is usually preferable. In
the past, approaches based on active depth sensors, such as Laser Range Finders
([14], [22], [23], [4], [17]), have been limited by the fact that 3D sensors had low
resolution and high prices. With the advent of reliable and affordable RGB-D
sensors, such as Microsoft Kinect3 , cameras providing aligned color and depth
measurements of the scene became available, thus allowing for algorithms which
could exploit this combined information. Since a single sensor can provide a
dense 3D representation of the scene at 30 Hz, robust algorithms can be applied
and the system installation becomes straightforward. In this work, the proposed
algorithms are targeted to be used on RGB-D data from structured light, stereo
or time-of-flight sensors.
The remainder of the paper is organized as follows: in Section 2, the literature
on RGB-D people detection and tracking is reviewed, while the ROS-Industrial
project is presented in Section 3. Section 4 describes the dataset we use in
this paper, while our people detection and tracking algorithms are reported in
Section 5 and 6 and the full pipelines are described in Section 7. In Section 8, we
present the experiments we performed and conclusions are drawn in Section 9.

3
http://www.microsoft.com/en-us/kinectforwindows.
 3

2 Related Work
Kinect SDK4 performs people detection based on the distance of the subject
from the background, while NiTE middleware5 relies on motion detection. Both
these approaches work in real time with CPU computation, but they are thought
to be used with a static camera and for entertainment applications. Thus, they
are not suitable to work in cluttered environments and if other moving objects
(robots) are present because of the assumption that the background is static.
Moreover, these algorithms only detect people up to four meters of distance from
the camera.
In [21], a people detection algorithm for RGB-D data is proposed, which ex-
ploits a combination of Histogram of Oriented Gradients (HOG) and Histogram
of Oriented Depth (HOD) descriptors and is not limited to static sensors or to
a restricted distance range. However, each RGB-D frame is densely scanned to
search for people, thus requiring a GPU implementation for being executed in
real time. Also [6] and [7] rely on a dense GPU-based object detection, while
[13] investigates how the usage of the people detector can be reduced using a
depth-based tracking of some Regions Of Interest (ROIs). However, the obtained
ROIs are again densely scanned by a GPU-based people detector.
In [12], a tracking algorithm on RGB-D data is proposed, which exploits
the multi-cue people detection approach described in [21]. It adopts an on-line
detector that learns individual target models and a multi-hypothesis decisional
framework. No information is given about the computational time needed by
the algorithm and results are reported for some sequences acquired from a static
platform equipped with three RGB-D sensors.
Algorithms for tracking people in real time from a mobile platform have been
presented in [1], [15] and [16]. These people detection techniques estimate the
ground plane equation and exploit a depth-based clustering followed by color-
based classification. They obtain high accuracy and framerate while only relying
on CPU computation. A similar approach has been implemented in [25], where
the concept of depth of interest is introduced to identify candidates for detection.
In this work, we propose a people detection approach which combines the
techniques proposed in [1], [15] and [16] with the algorithms validated in [5]
for detecting people from stereo data. Moreover, we exploit a people tracking
algorithm which can recover from missed detections and identity switches by
means of a particle filter guided by color histograms of the tracked people.

3 ROS-Industrial
ROS-Industrial is an open-source project that extends the advanced capabilities
of the Robot Operating System (ROS) [19] software to new industrial applica-
tions. Among the goals of ROS-Industrial, there is the aim to develop robust and
reliable software that meets the needs of industrial applications. This is achieved
4
http://www.microsoft.com/en-us/kinectforwindows/develop.
5
http://www.primesense.com/solutions/nite-middleware.
4

by combining the relative strengths of ROS with existing industrial technologies

(i.e. combining ROS high-level functionality with the low-level reliability and
safety of industrial robot controllers). ROS-Industrial already provides libraries,
tools and drivers for interfacing with a number of industrial manipulators and it
is creating standard interfaces to stimulate hardware-agnostic software develop-
ment. All these aspects concur to create a generation of robots which could be
versatile in the tasks they execute and collaborative with humans. Thus, ROS-I
is also concerned with the problem of detecting people in shared robot-human
workspaces and of defining standard requirements which should be satisfied by
such algorithms in order to ensure people safety in industrial environments.

4 ROS-Industrial People Dataset

In order to evaluate the performance of the proposed tracking pipelines in a

real industrial setting, the ROS-Industrial People Dataset has been collected in
a warehouse very similar to an industrial scenario. The dataset contains about
11000 RGB images and their corresponding disparity images acquired with an
Asus Xtion Pro Live at VGA resolution. One to five people were moving in
front of the camera. Other than people, also robots, objects and industrial tools
feature the dataset. About one fifth of the images do not contain people, in order
to test robustness to false positives, and some images were collected while the
camera was moving parallel to the ground plane. Some sample RGB, disparity
images and annotations from this dataset are reported in Fig. 1. About one tenth

Fig. 1. Sample RGB, corresponding disparity images and ground truth annotations
from the ROS-Industrial People Dataset.

of the dataset frames are annotated, that is the corresponding ground truth is
provided is files which contain the bounding boxes of the people present in the
images.
 5

5 People Detection

In this section, we present the people detection approach we used in this work.
Since we cannot rely on the assumption that the background is static, we de-
veloped a detection technique which could work from a single frame without
knowledge inherited from the previous ones.

5.1 Cascade classifier modularity

For limiting the computational onerosity while preserving accuracy, we orga-

nized the detection process as a cascade of different algorithms. In this cascade,
detection methods are ordered in decreasing order of speed and increasing order
of classification accuracy. In this way, we applied a very simple and fast verifica-
tion process when many detection windows have to be processed, while we keep
more accurate but slower detection methods for the last stages of the cascade,
when only a small number of detection windows have to be analyzed.
In the next paragraphs, we will describe the algorithms we developed for
composing the single stages of the cascade. They can be divided in consistency,
depth-based and color-based algorithms.
In Fig. 2, the detections produced at every stage of a detection cascade that
we will see in Section 7 are shown.

Fig. 2. Visualization of the output detections at every stage of the detection cascade
used in the pipeline of Figure 4a for a frame of the ROS-Industrial People Dataset.

5.2 Consistency constraints

The first module we present implements two constraints which allow us to check
if a detection window is compatible with people being on the ground plane. The
first constraint is a height in image vs. disparity constraint. The idea behind this
constraint is that a pedestrian of a certain height in an image, measured in pixels,
should have a disparity that places the pedestrian at a reasonable distance.
The second constraint forces objects to be located on the ground plane. For
doing this, we estimate the ground plane equation with the Hough-based method
proposed in [5] and then limit search windows by constraining object height and
y location.
6

The thresholds for these contraints have been set empirically, after an evalu-
ation performed on some training videos with the camera placed at 180 cm from
the ground.
In addition to these constraints, we remove high overlapping detection win-
dows resulting from the consistency module. In particular, we remove those
windows which overlap more than 80% with another one. This operation allows
to considerably reduce the number of detection windows which have to be ana-
lyzed by the following stages of the detection cascade, thus saving computational
time and reducing the false positive rate. Later on in the paper, we will use the
Cons abbreviation for the consistency module, while its version which removes
highly overlapping detection windows will be called CROW (Consistency with
Removal of Overlapping Windows). In Figure 3, the output of the Cons and
CROW modules can be compared for an image of the ROS-Industrial People
Dataset.

(a) Cons (b) CROW

Fig. 3. Effect of removing highly overlapping windows in the consistency module.

5.3 Color-based algorithms

Haar-like features extraction on color images We utilize Haar-like features
in the same manner as [18] but with the key differences proposed in [5].
These features are used with two classifiers. The first of these is a preliminary
AdaBoost classifier which utilizes the Haar features and simple decision trees
as the weak classifiers. The classifier is trained with preliminary weights which
favor correct detections over limiting the false positive rate, the purpose being to
eliminate windows which are obviously not persons. The preliminary AdaBoost
classifier used in our final implementation has a detection rate of about 97% and
a false positive rate of about 5% on the INRIA Person Dataset [8].
We also tried these Haar-like features with a Support Vector Machine (SVM)
classifier. Using SVM is advantageous because it allows to obtain the same per-
formance obtained by Viola and Jones [24] with Adaboost, but the training
process is faster.
We will refer to the approach using Haar features extracted on the color
image and the Adaboost classifier as HaarAda, while we will call HaarSvm the
method which uses the Support Vector Machine.
 7

HOG-like descriptors We implemented two methods exploiting two different

versions of the HOG descriptor which is then used together with a Support
Vector Machine classifier. The first method (called HogSvm), computes HOG
features only in chosen parts of the detection window, where the outside edges
of the person should be. This method is motivated by the attempt to extract
only relevant information and discard gradients on the background.
In this work, we also use a HOG implementation which extracts features from
the whole detection window, thus being similar to the original version of Dalal
and Triggs [8]. This second version is also optimized to use SSE2 instructions,
which allow to speed up descriptor computation up to four times. Since this
algorithm has been contributed to the Point Cloud Library [20], we will refer to
this method as HogSvmPCL. We trained the Support Vector Machine classifier in
two different ways: on whole person bodies (HogSvmPCLWholebody) and on half
person bodies containing only persons’ head and torso (HogSvmPCLHalfbody).
The latter mode is computationally faster because the descriptor size is a half
and it should behave better when only the upper part of a person is visible.
The HOG-based classifiers, though expensive in terms of both feature com-
putation and classification, serve as a final verifier and are usually placed at the
end of the detection cascade, when only a small number of detection windows
remain.

5.4 Depth-based algorithms

Depth-based clustering Similarly to the consistency method proposed in

Section 5.2, this module is intended to act as the first stage of a detection cascade
in order to find a small number of detection windows which are then better
analyzed by the following stages of the cascade. However, unlike the consistency
method, this algorithm requires a point cloud of the scene as input, instead of a
disparity image. Given the point cloud, this method removes the ground plane
points and performs a clustering of the remaining points based on their Euclidean
distance. The obtained clusters are then sub-clustered in order to center them on
peaks of a height map which contains the distance of the points from the ground
plane. These peaks are intended to represent the position of people’s head. The
abbreviated name for this method will be DC (Depth Clustering).

Haar-like features extraction on disparity images With some slight mod-

ifications, the same preliminary Haar-like features classifier presented in Section
5.3 can be used to locate people in disparity images. The key difference in imple-
mentation is that the scaling for computing the features must take pixels with
unknown disparity into account. In our system, pixels for which disparity cannot
be determined are assigned a value of zero. We apply a special scaling function
in computing the Haar-like feature maps which does not include zero pixels in
its area averages. The abbreviated name for this method will be HaarDispAda.
It is worth noting that the features from the disparity map are quite in-
dependent from the features of the color image. Thus, the combination of the
8

disparity map classifier and preliminary Haar-like features classifier leads to very
high detection rates and low false positive rates.

6 Tracking Algorithm
The tracking algorithm we propose exploits the output of the people detection
cascade in order to perform data association between existing tracks and new
detections. This matching is based on computing color histograms of persons
from the color image. If a detection is not associated with any existing track, a
new track is initialized in a new thread. This thread is dedicated to a particle
filter in the x-y plane which exploits detections and the corresponding color
histograms to search for the target person within the image. Such an approach
allows to detect people even when the detection cascade returns no valid output.
Moreover, the use of a particle filter allows to recover from errors of the tracker,
such as identity switches. In this data association scheme, multiple detections
can be associated to the same track, in order to avoid multiple tracks generated
from the same person. It should be noted that this algorithm requires a separated
thread for every track, thus the number of tracks which can be handled is limited
by the multi-threading capabilities of the computer.

7 Detection and Tracking Pipelines

In this work, we compare different pipelines for people detection and tracking.
These pipelines differ in the detection cascade, which is composed by a com-
bination of the detection modules described in Section 5. In Fig. 4, we report
the main pipelines we tested. It should be noted that the more are the stages of
the detection cascade and higher is the system latency, thus a shorter cascade
is preferable. However, a too short cascade could result to be inefficient if the
single modules are not fast enough. As we mentioned in Section 5, we tried both
the CROW and the DC modules at the first stage of the cascade. As we will see
in Section 8.3, the HogSvmPCL module resulted to be the best option for the
second stage in terms of both framerate and accuracy.

8 Experiments
In this section, the pipelines proposed in Section 7 are evaluated in terms of
accuracy and framerate. These pipelines differ in the detection cascade, while
the tracking node is the same for all, thus the name chosen to represent them is
the description of the detection cascade.

8.1 Tracking evaluation

We evaluate tracking results in terms of false positives and missed detections.
In order to compute these quantitative indices, we compare ground truth and
 9

(a) (b) (c) (d)

Fig. 4. People detection and tracking pipelines.

tracking bounding boxes in the image by means of the PASCAL rule usually
exploited for evaluating object detectors [11]. Given that tracking results can
be considered good even if the computed bounding box is a bit off with respect
to the person center, a threshold of 0.3 has been used in the PASCAL rule,
instead of the standard 0.5 threshold. Associations between results ROIs and
ground truth ROIs are computed with a Global Nearest Neighbor approach, by
means of the Hungarian algorithm6 . Then, a number of quantitative indices are
computed. In this work, these two indices have been used:

– False Rejection Rate (%) (FRR): 100*miss/(TP + miss) = % miss

– False Positives Per Frames (FPPF): FP/frames,

where FP is the number of false positives, TP is the number of true positives

and miss is the number of false negatives.

8.2 Results on ROS-Industrial People Dataset

In Fig. 5, the pipelines described in Section 7 are compared by means of Detection

Error Trade-Off (DET) curves reporting FRR in the x-axis and FPPF in the
y-axis. These curves have been obtained while performing detection and tracking
on the whole ROS-Industrial People Dataset and by varying the threshold on
6
http://en.wikipedia.org/wiki/Hungarian_algorithm.
10

Fig. 5. DET curves comparing the main approaches proposed in this work.

the SVM score associated to the HOG descriptor. The ideal working point for
these curves is located at the bottom-left corner (with FRR = 0% and FPPF =
0). For visualization purposes, the curves are reported in logarithmic scale.
It can be noticed that the removal of overlapping windows in the CROW
module leads to a very small increase in the percentage of missed people. This
means that some good detection windows are sometimes removed because highly
overlapping with other detection windows. The algorithm classifying Haar fea-
tures from the disparity image (HaarDispAda) proved to be very effective in
reducing the number of false positives which are produced by the methods oper-
ating on the color image. Moreover, the HogSvmPCLHalfbody technique, which
is trained on half person bodies, seems to be the essential module for obtaining
the minimum number of false positives, while tracking people in about 80% of
the frames. This proved to be true both in combination with the CROW and
with the DC methods at the first stage of the detection cascade. In fact, for
the pipeline CROW+HogSvmPCLHalfbody+HaarDispAda, with a threshold of -
0.925 on the HOG threshold for people detection, we obtain FRR = 21.95% and
FPPF = 0.17. Some qualitative tracking results obtained with this approach are
shown in Fig. 6.
From a qualitative point of view, the detections produced by the cascades
which exploit the DC module at the first stage are more persistent and well cen-
tered on people, thus the people detector obtains superior performance. However,
 11

Fig. 6. Tracking results with the CROW+HogSvmPCLHalfbody+HaarDispAda

pipeline on some sample images of the ROS-Industrial People Dataset.

it seems not to make the difference at the tracking level with respect to pipelines
based on the CROW module because the particle filter is able to interpolate
when some detections are missing.
Since our detection methods independently process every frame, we obtained
good tracking results also for the part of the dataset where the camera is moving.

8.3 Framerate analysis

In Fig. 7, the framerates measured for all the tested pipelines are reported.
The measurements have been done every second while performing tracking on
the ROS-Industrial People Dataset. In order to detect the maximum framerate
achievable by each pipeline, tracking has been performed when streaming the
dataset images at 50 Hz, while the sensor framerate was of 30 Hz. Only for the
pipeline using the DC approach, we could not use this measurement method
because it was not possible to stream point clouds at 50 Hz. Thus, the framerate
we report has been measured while processing a live Kinect stream (at 30 Hz).
For comparing our two implementations of the HOG descriptor, we also re-
port the framerate of the CROW+HogSvm pipeline, which resulted three times
slower than what obtained when using the HogSvmPCL module.
All our people detection and tracking algorithms are implemented in C++ and
exploit the Robot Operating System (ROS) [19] and open source libraries for
2D [2] and 3D [20] computer vision.
The detection cascades are composed of ROS nodelets which share the same
process, thus avoiding data copying between them and gaining in terms of fram-
erate.
The best approches in terms of framerate are the CROW+HogSvmPCLWhole
body+HaarDispAda and CROW+HogSvmPCLHalfbody+HaarDispAda, which track
people at framerates between 33 and 35 fps. It is worth noting that these pipelines
resulted to be the best also in terms of accuracy.

9 Conclusions

In this work, we proposed algorithms for detecting and tracking people from
aligned color and depth data in industrial environments. Several pipelines have
12

Fig. 7. Framerates measurements for the main approaches proposed in this work.

been considered and evaluated on a dataset featured by levels of clutter and oc-
clusion typical of an industrial setting. Our detection approach combines depth-
based and color-based techniques in a cascade and our tracking algorithm allows
to track people even if detections are missing for many frames and to recover
from identity switches. The best pipelines we proposed allow to track people
for about 80% of the frames with a very low number of false tracks and at a
framerate higher than 30 fps.
More than one pipeline showed very good performance and all the code has
been publically released in ROS-Industrial Human Tracker repository7 , so that
a wide choice of algorithms is available to users and developers.
Even if the obtained results are good, safety measures impose that no person
should be missed in industrial applications. Thus, as a future work, we envision
to further improve the detection of people while retaining the same false posi-
tive rate. For this purpose, we will add a validation phase within the tracking
algorithm, so that false tracks could be immediately deleted if the detection
confidence does not pass a proper multi-frame check. Moreover, we will work on
making the tracking algorithm more scalable in the number of tracked people.

10 Acknowledgements

We would like to thank Open Perception Inc., the Southwest Research Institute
(SwRI) and the National Institute of Standards and Technology (NIST) for
7
At the moment of writing, the source code is in the develop branch of the
human tracker GitHub repository: https://github.com/ros-industrial/human_
tracker/tree/develop.
 13

funding this research and Radu Bogdan Rusu for the fruitful discussions on this
topic.

References
1. F. Basso, M. Munaro, S. Michieletto, E. Pagello, and E. Menegatti. Fast and
robust multi-people tracking from rgb-d data for a mobile robot. In 12th Intelligent
Autonomous Systems Conference (IAS-12), pages 265–276, Jeju Island, Korea,
June 2012.
2. G. Bradski. The OpenCV Library. Dr. Dobb’s Journal of Software Tools, 2000.
3. Michael D. Breitenstein, Fabian Reichlin, Bastian Leibe, Esther Koller-Meier, and
Luc Van Gool. Robust tracking-by-detection using a detector confidence particle
filter. In International Conference on Computer Vision (ICCV) 2009, volume 1,
pages 1515–1522, October 2009.
4. Alexander Carballo, Akihisa Ohya, and Shin’ichi Yuta. Reliable people detection
using range and intensity data from multiple layers of laser range finders on a
mobile robot. International Journal of Social Robotics, 3(2):167–186, 2011.
5. David R. Chambers, Clay Flannigan, and Benjamin Wheeler. High-accuracy
real-time pedestrian detection system using 2d and 3d features. Proc. SPIE,
8384:83840G–83840G–11, 2012.
6. W. Choi, C. Pantofaru, and S. Savarese. Detecting and tracking people using
an rgb-d camera via multiple detector fusion. In International Conference on
Computer Vision (ICCV) Workshops 2011, pages 1076–1083, 2011.
7. W. Choi, C. Pantofaru, and S. Savarese. A general framework for tracking multiple
people from a moving camera. Pattern Analysis and Machine Intelligence (PAMI),
35(7):1577–1591, 2012.
8. N. Dalal and B. Triggs. Histograms of oriented gradients for human detection. In
Computer Vision and Pattern Recognition (CVPR) 2005, volume 1, pages 886–893,
June 2005.
9. A. Ess, B. Leibe, K. Schindler, and L. Van Gool. A mobile vision system for robust
multi-person tracking. In Computer Vision and Pattern Recognition (CVPR) 2008,
pages 1–8, 2008.
10. A. Ess, B. Leibe, K. Schindler, and L. Van Gool. Moving obstacle detection in
highly dynamic scenes. In International Conference on Robotics and Automation
(ICRA) 2009, pages 4451–4458, 2009.
11. Mark Everingham, Luc Gool, Christopher K. Williams, John Winn, and Andrew
Zisserman. The pascal visual object classes (voc) challenge. International Journal
of Computer Vision, 88:303–338, June 2010.
12. Matthias Luber, Luciano Spinello, and Kai O. Arras. People tracking in rgb-d data
with on-line boosted target models. In International Conference On Intelligent
Robots and Systems (IROS) 2011, pages 3844–3849, 2011.
13. D. Mitzel and B. Leibe. Real-time multi-person tracking with detector assisted
structure propagation. In International Conference on Computer Vision (ICCV)
Workshops 2011, pages 974–981. IEEE, 2011.
14. Oscar Mozos, Ryo Kurazume, and Tsutomu Hasegawa. Multi-part people detection
using 2d range data. International Journal of Social Robotics, 2:31–40, 2010.
15. M. Munaro, F. Basso, and E. Menegatti. Tracking people within groups with rgb-d
data. In Proc. of the International Conference on Intelligent Robots and Systems
(IROS), pages 2101–2107, Algarve, Portugal, October 2012.
14

16. M. Munaro and E. Menegatti. Fast rgb-d people tracking for service robots. Au-
tonomous Robots Journal, 2014.
17. Luis E. Navarro-Serment, Christoph Mertz, and Martial Hebert. Pedestrian detec-
tion and tracking using three-dimensional ladar data. In The International Jour-
nal of Robotics Research, Special Issue on the Seventh International Conference on
Field and Service Robots, pages 103–112, 2009.
18. C. Papageorgiou, T. Evgeniou, and T. Poggio. A trainable pedestrian detection
system. In In Proceedings of IEEE Intelligent Vehicles Symposium ’98, pages 241–
246, 1998.
19. Morgan Quigley, Brian Gerkey, Ken Conley, Josh Faust, Tully Foote, Jeremy Leibs,
Eric Berger, Rob Wheeler, and Andrew Ng. Ros: an open-source robot operating
system. In International Conference on Robotics and Automation (ICRA), 2009.
20. Radu Bogdan Rusu and Steve Cousins. 3D is here: Point Cloud Library (PCL).
In International Conference on Robotics and Automation (ICRA) 2011, pages 1–4,
Shanghai, China, May 9-13 2011.
21. Luciano Spinello and Kai O. Arras. People detection in rgb-d data. In International
Conference On Intelligent Robots and Systems (IROS) 2011, pages 3838–3843,
2011.
22. Luciano Spinello, Kai O. Arras, Rudolph Triebel, and Roland Siegwart. A lay-
ered approach to people detection in 3d range data. In Conference on Artificial
Intelligence AAAI’10, PGAI Track, Atlanta, USA, 2010.
23. Luciano Spinello, Matthias Luber, and Kai O. Arras. Tracking people in 3d using
a bottom-up top-down people detector. In International Conference on Robotics
and Automation (ICRA) 2011, pages 1304–1310, Shanghai, 2011.
24. Paul Viola and Michael Jones. Rapid object detection using a boosted cascade
of simple features. In Computer Vision and Pattern Recognition (CVPR) 2001,
volume 1, pages 511–518, 2001.
25. Hao Zhang, C. Reardon, and L.E. Parker. Real-time multiple human perception
with color-depth cameras on a mobile robot. IEEE Transactions on Cybernetics -
Part B, 43(5):1429–1441, Oct 2013.