Human-aware Robot Navigation in Logistics Warehouses

Mourad A. Kenk

1,2 a

, M. Hassaballah

3 b

and Jean-Franc¸ois Breth

1 c

Electrotechnics and Automatics Research Group (GREAH), Normandy University, Le Havre, France

Department of Mathematics, South Valley University, Qena, Egypt

Department of Computer Science, South Valley University, Luxor, Egypt

Keywords:

Human-aware Navigation, Collision Avoidance, Human Detection in RGB-D Camera, 2D-LIDAR Sensor,

Logistics Warehouses, Mobile Robot Autonomous Navigation, Robotics.

Abstract:

Industrial and mobile robots demand reliable and safe navigation capabilities to operate in human populated

environments such as advanced manufacturing industries and logistics warehouses. Currently mobile robot

platforms can navigate through their environment avoiding coworkers in the shared workspace, considering

them as static or dynamic obstacles. This strategy is efﬁcient for safety, strictly speaking, but is not sufﬁcient

to provide humans integrity and comfortable working conditions. To this end, this paper proposes a human-

aware navigation framework for comfortable, reliable and safely navigation designed to run in real-time on a

mobile robot platform in logistics warehouses. This is accomplished by estimating human localization using

RGB-D detector, then generating a virtual circular obstacle enclosing human pose. This virtual obstacle is

then fused with the 2D laser range scan and used in ROS navigation stack local costmap for human-aware

navigation. This strategy guarantees a different approach distance to obstacles depending on the human or

non-human nature of the obstacle. Hence the mobile robot can approach closely to pallet to pick up objects

while maintaining an integrity distance to humans. The reliability of the proposed framework is demonstrated

in a workbench of experiments using simulated mobile robot navigation in logistics warehouses environment.

1 INTRODUCTION

In logistics warehouses, pick and place mobile robots

are designed to navigate in an environment often

structured in long and narrow aisles. Robots need to

approach very closely the racks located on the side

of the aisle (Wahrmann et al., 2017). In the same

time, robots are likely to cross human coworkers. Hu-

mans need to feel safe in their personal space when

sharing workspace with robots. In vision-based au-

tonomous navigation, several typical obstacle avoid-

ance methods have been proposed recently. Krajnik

et al. (Krajn

ık et al., 2017) proposed to use hetero-

geneous visual features such as points, line segments,

lines, planes, and vanishing points to process a visual

simultaneous localization and mapping (SLAM) or to

use depth information from a low-cost sensor (RGB-

D) to localize the obstacles. In (Yang et al., 2016),

a nonlinear controller is designed to achieve target

tracking and obstacle avoidance in complex environ-

https://orcid.org/0000-0002-2374-4032

https://orcid.org/0000-0001-5655-8511

https://orcid.org/0000-0002-6962-954X

Figure 1: Example of an output frame of ROS based human

perception process, showing detector and localizer results

in RViz.

ments. In (Malone et al., 2017), stochastic reachable

sets are used to generate accurate artiﬁcial potential

ﬁeld for dynamic obstacles for online path planning.

In (Nardi and Stachniss, 2017), a probabilistic ap-

proach is developed for modeling uncertain trajecto-

ries of the moving entities that share workspace with

robot.

Kenk, M., Hassaballah, M. and Brethé, J.

Human-aware Robot Navigation in Logistics Warehouses.

DOI: 10.5220/0007920903710378

In Proceedings of the 16th International Conference on Informatics in Control, Automation and Robotics (ICINCO 2019), pages 371-378

ISBN: 978-989-758-380-3

371

Most of the previous research works consider hu-

mans as dynamic obstacles using different techniques

and sensing modalities. Nevertheless, considering hu-

mans as dynamic obstacles to simply avoid them is

certainly sufﬁcient from the safety point of view, but

can cause annoyances for humans. Indeed, proxemics

theory, one of the most popular principles in Human

Robot Spatial Interaction (HRSI), advises to select

appropriate interaction distances between robots and

humans (Lindner and Eschenbach, 2017). The robot

should keep a certain distance from a human while

navigating, preferably greater than 1.2m not to vio-

late the Personal Space (PS).

Currently, many researches are devoted to design

navigation approaches with safety control in com-

plex environments. For instance, Choi et al. (Choi

et al., 2017) proposed Gaussian process motion re-

gression based robot navigation, which predicts fu-

ture trajectory of human with two Microsoft Kinect

sensors in dynamic environments. In (Song et al.,

2017), a shared-control scheme based navigation is

proposed via combining active obstacle avoidance

and passive-compliant motion behavior prediction for

human, where walking-assistant robot avoids colli-

sion and allows safe guidance for human using leg

detector by 2D laser sensor. In (Zimmermann et al.,

2018), the robot trajectory planning is done by esti-

mating 3D human pose in RGB-D images while exe-

cuting tasks in cluttered environments occupied with

different features of the obstacles.

Numerous solutions have been presented for hu-

man detection and tracking in a human populated

environment based on 2D Laser range sensors (Lin-

der et al., 2016; Song et al., 2017). The core idea

of human detection is a binary classiﬁer trained on

leg appearance features that is reﬂected by beams of

2D laser sensor. First, the reﬂected 2D laser data

(points) is segmented according to Euclidian distance

(jump distance) between continuously neighboring

2D points. These segments are then classiﬁed as a

human leg if it scores approximation values of 2D ge-

ometrical structure features, which are characterized

by pre-deﬁned fourteen parameters such as number of

reﬂected points, width, linearity, circularity, radius as

described by (Arras et al., 2007). Nevertheless, even

if these methods achieve great performances in air-

port and hospital environment, they are not applicable

in logistics warehouses environment due to the high

number of false detection as reported in this study.

On the other hand, human detectors based on 3D vi-

sion show high performances in industrial environ-

ment (Munaro et al., 2016) and intralogistics ware-

houses (Linder et al., 2018).

For these reasons, the main goal of this research is

to design a human-aware navigation framework that

makes the robot behaviour more sociable (as shown in

Figure 1). It implies being able to distinguish between

human and non-human obstacles, which is currently

the case in the state-of-the-art. We investigate the

constraints of human detection algorithm implemen-

tation taking into account different point of view, from

the computational costs, to the safety and proxemics

theory requirements. The proposed framework is de-

signed to be optimal in logistics warehouses environ-

ments with their speciﬁc properties and constraints.

Then, the functionality of the framework is proved

in an experiment based on autonomous Summit XL

mobile robot in a logistics warehouses simulated en-

vironment relying on the well known Robot Operat-

ing System (ROS) (Quigley et al., 2009). To illus-

trate, one of the powerful ROS toolbox is Navigation

Stack

which detects obstacles through 2D laser sen-

sor in real time. Then, inﬂation layer is created with

around obstacles by inﬂation radius value to avoid the

collision. The idea is to use a small inﬂation radius

value that allow pick and place mobile robot to close

enough to its targets that are installed in selective

racks. Additionally, coworkers safety area is manip-

ulated as a circular obstacle fused in laser scan data.

The ﬁnal achievement of this research is a global ar-

chitecture keeping a balance between real-time con-

straints leading to investment and operational costs

and societal issues to make the human working con-

ditions more comfortable.

This paper is structured as follows: Section 2

describes the proposed sensor interaction algorithm

for human-aware mobile robot navigation in logistics

warehouses. Section 3 presents experimental results

demonstrating the advantages of the proposed frame-

work. Finally, conclusions are drawn in Section 4.

2 HUMAN-AWARE NAVIGATION

FRAMEWORK

2.1 Mobile Robot Navigation

ROS is a compilation of tools, libraries, and utili-

ties which facilitate the design of complex robot tasks

through powerful algorithms implementation such as

Navigation Stack which is composed of several algo-

rithms (ROS nodes). These nodes are communicating

via publish/subscribe message-oriented middleware.

For each ROS node, information is sent/received via

a given topic as a structured data message. Conse-

quently, a variety of information patterns can be ex-

ROS navigation, http://wiki.ros.org/navigation.

ICINCO 2019 - 16th International Conference on Informatics in Control, Automation and Robotics

372

ROS Navigation Stack

move_base

Localization

Safety area

Human pose

Transforms

RGB-D frame

2D-Lidar frame

Sensors

RGB-D Camera

ORBBEC Astra

2D-Lidar

HOKUYO

UST-20XL

Human_Safety_Fusion

2D-Lidar

TEB Local planner

Local costmap

Figure 2: Conceptual overview of the system architecture

based on ROS Navigation Stack. Safety fusion is discussed

in Algorithm 1.

changed between different information sources and

manipulated by robust algorithms.

Actually, ROS Navigation Stack provides generic

autonomous navigation via path planning and SLAM

system. This allows robot to localize obstacles around

it, plan a path to avoid collision, and navigate in un-

known or partially known environments. Therefore,

online planning is needed to create a collision free

trajectory over the global trajectory toward the goal

w.r.t trajectory optimization. So that, (R

osmann et al.,

2017) policies are used for on-line trajectory planning

which is based on Timed-Elastic-Band (TEB) ap-

proach. That is called teb-local planner

which com-

bines temporal information to reach the goal. This

method based on odometry and laser scan data re-

quires only low computational resources. Conceptual

overview of the system architecture is shown in Fig-

ure 2. based on move base

node in ROS Navigation

stack.

2.2 Human Detection and Localization

To localize humans, we implement depth-based hu-

man detection method using the Point Cloud Library

(PCL) and RGB-D data (Munaro et al., 2016) as il-

lustrated in Figure 3. Initially, to obtain a real time

performance, the 3D points cloud data is reduced by

voxel grid ﬁlter, where each voxel is scaled down by

threshold size close to its centroid coordinates. This

downsize operation reduces voxel size to 0.06 m of

its real size that is sufﬁcient to perform the detection

process.

The ground plane (GP) is estimated and removed

from the obtained voxel grid. The RANSAC-based

least square method is used at each frame to com-

pute the GP equation coefﬁcients and update it with

teb-local-planner, http://wiki.ros.org/teb local planner.

ROS move base, http://wiki.ros.org/move base.

OpenNI

RGB-D camera

current cloud

Voxel grid filter

3D bounding box

Centroid(x,y,z)

Decision

Human clusters

Remove GP

Ground Plane GP

coefﬁcients Initialization

GP Estimation

RANSAC-based

Compute RGB confidence

RGB pre-trained

human classifier

RGB image

Clustering

neighboring 3D points

(Euclidean distances based)

Sub-Clustering

head positions

Figure 3: Block diagram describing input/output data

and the main operations performed by human localization

method.

respect of the previous frame estimation. The remain-

ing 3D points are clustered in 3D boxes based on their

Euclidean distances. Then, these 3D boxes are seg-

mented into sub-clusters to center them on peaks of a

height map which contains threshold distance of the

points from GP. These peaks are representing the hu-

man’s head positions and its 2D RGB image is repre-

senting region of interest (ROIs). Instantly, the RGB

image exhibiting ROIs is used by RGB-based human

detector to compute RGB conﬁdence index for the ob-

tained sub-clusters. From the conﬁdence index, we

can select true positive human detection and elimi-

nate false detection. At this stage, two methods are

investigated:

• A usual machine learning method: HOG-based

human detector with the recommended training

conﬁgurations by (Dalal and Triggs, 2005) based

on support vector machine (SVM).

• A recent deep neural network method based on

reinforcement computation implemented in CPU

or GPU: YOLO-based human detector by (Red-

mon and Farhadi, 2018) in its open source Dark-

net

framework with 53 convolutional layers for

features extraction.

In experiments, a pre-trained model (yolov3-tiny

based on COCO dataset) is used, which is charac-

terised by small model. Then, each human-detector

method is tested individually and the obtained results

are discussed in more details in results section. An ex-

ample for the extended code to output markers for vi-

sualization of ROIs and their RGB conﬁdence score is

given in Figure 4. Finally, the detected humans poses

are published as ROS topic to be used further with

human safety area generation.

Darknet, https://pjreddie.com/darknet/

Human-aware Robot Navigation in Logistics Warehouses

373

(a) Depth image

(b) 3D boxes (c) Sub clusters

(d) RGB human detection

(e) Human localization

Figure 4: Example of output markers for visualization of

ROIs in RViz: First row depth image raw, 3D boxes for

clusters and sub-clusters, second row in RGB detection red

color rectangle for YOLO detector, green color rectangle

for HOG detector. Third row, localized human (yellow 3D

boxes).

2.3 Human Safety Area (HSA) Fusion

The navigation stack needs to know the position of

sensors, wheels, and joints in respect to robot base.

This is done using ROS Transform Frame (TF) pack-

age. The located human position coordinate is trans-

formed from RGB-D frame to 2D LIDAR sensor

frame for further human safety area calculations.

In warehouse environment, the mobile manipula-

tor needs to come close to the pallets to pick up ob-

jects on the pallets. So the parameter for the security

distance used to avoid collision between robots and

pallets is set low. In the meantime, humans must not

be scared by robots and the proxemics theory gives a

higher value for the optimal distance between robots

and humans (1.2 m). Once the distinction between

human and non-human obstacle is done, the fusion

process 2D laser plane must take this constraint into

account. The HSA is virtually inserted as an obstacle

in the warehouse environment and merge with the sig-

nal coming from the 2D lidars. This new signal and

associated occupancy grid is then used by the robot

local planner to navigate safely and human integrity

in the warehouse.

The tested 2D-LIDAR is of type Hokuyo UST-

Algorithm 1: Human Safety Fusion.

1 HumanDetector (rgb.img , depth.cloud);

2 2D Laser Scan; //scan.msg;

3 while HumanDetections(x , y , z) do

4 Calculate HSA ← (x , y , z);

5 //Generate 2D laser data message;

6 (safety.msg) ← HSA ;

7 end

8 //Async Spinner;

9 obstacle.msg ← (scan.msg , safety.msg);

20LX at a height of 0.12 m with 20 m detection range,

and 270

◦

wide detection angle [−135

◦

: 135

◦

]. The

start laser ray angle φ

◦

= −135

◦

forward along the X

axis in laser frame, alongside high angular resolution

∆ = 0.25

◦

(angle between 2 beams) with measure-

ment discrete steps (η

,i = [1 : 1081]) and 25 msec

high speed response. To sum up, the laser ray angle

∈ [φ

◦

+ η

· ∆]. As shown in Figure 5a, suppose the

HSA is represented as a Virtual Safety Circle (VSC)

centroid with detected human coordinates C =





and radius r. The goal is to ﬁnd all the range µ

and

intensity for each angle θ

of index i (i ∈ 1 : 1081)

which is done using the following calculations:

First, we determine the angle θ

, the θ

mini-

mum, and θ

maximum angles of the

PN arc corre-

sponding to intersection points P,N displayed on Fig-

ure 5a.

= arctan2(x

) , η

= E

− φ

∆

(1)

= min(θ

+ ψ,θ

− ψ) , η

= E

− φ

◦

∆

(2)

= max(θ

+ ψ,θ

− ψ) , η

= E

− φ

◦

∆

(3)

where ψ = arccos(r/d). Let ω

= θ

− θ

and µ

, θ

respectively the range and angle associated with Q

(see Figure 5b). Applying Al Kashi theorem in trian-

gle LQ

= µ

+ d

− 2d · µ

· cos(ω

) (4)

The resolution of this quadratic equation leads to two

solutions and the minimum is chosen to be µ

. Thus,

= µ

cosθ

sinθ

(5)

ICINCO 2019 - 16th International Conference on Informatics in Control, Automation and Robotics

374

(a)

(b)

Figure 5: Calculations of the arc HSA in 2D-LIDAR coor-

dinates system.

The Async Spinner

is utilized to merge the gen-

erated circular obstacles to 2D LIDAR sensor as de-

scribed in human safety fusion algorithm 1. Whereas,

a delayed message is controlled in custom time stamp

queue and empty message is eliminated to keep track

to the detected obstacles.

3 EXPERIMENTS AND RESULTS

The performance and capabilities of the proposed sys-

tem are tested using a PC with Intel Core i7-6700HQ

CPU 2.60 GHz, 32 GB of RAM and Nvidia Quadro

GPU M3000M with 4 GB of memory and 1024

CUDA cores under Ubuntu 16.04 (Xenial) with ROS

Kinetic. Several experiments are designed to rein-

force the demands invoked in the introduction. First,

human detection is investigated and evaluated, then

the evaluation of the safety navigation approach is fo-

cused on avoiding obstacles. Testing the navigation

system is carried out in logistic warehouse simula-

tion environment using Summit XL mobile robot with

Orbbec Astra RGB-D camera mounted on top of the

robot at 1.63 m and with Hokuyo laser range ﬁnder at

0.12 m from ground plane as given in Figure 6.

Async Spinner, http://wiki.ros.org/roscpp/Overview/

Callbacks and Spinning

3.1 Human Detection Evaluation

Detection-based metrics (Goutte and Gaussier, 2005)

are utilized for evaluating human detectors. These

metrics include Accuracy, Precision, Recall, F1-

Measure and FP-Rate which are depending on the

number of correctly detected (True Positive: TP),

falsely detected (False Positive: FP), missed detected

objects (False Negative: FN) and true rejected object

(True Negative: TN). The following performance in-

dices are considered :

• Precision of predictions = T P/(T P + FP)

• Recall = T P/(T P + FN)

• Accuracy = ((T P+T N)/(T P +T N +FP +FN))

• F1-Measure = 2 ×

precision×recall

precision+recall

• FP-Rate = FP/(FP + T N)

For human detection in 2D laser range data, we

compare between different trained classiﬁers on laser

features described in (Linder et al., 2016) using three

supervised learning techniques: Adaboost, SVM and

the random forest (with 15 trees and maximum depth

of 10). Comparatively to the recorded results for

human detection which addressed a highly perfor-

mance in (Linder et al., 2016) due airport environment

dataset given in Table 1. Different from the reported

result of the same pre-trained human classiﬁers in lo-

gistics environment. Unfortunately, human detection

by 2D laser range data is ruled out, where it scores

high false alarms and missed detection in respect to

the appearance similarity between human leg features

and logistics warehouse equipment such as selective

racks and pallets at ground plane. 2D-LIDAR fea-

tures sample for false alarm and missed detection re-

sults are shown in Figure 7. In details, false alarms by

euro pallet are formatted with a 2D laser features most

similarly to the human leg features when coworker

is in standing state; where both of them are approxi-

mately sharing some features (Circularity, Radius and

number of reﬂected points). Additionally, when the

Hokuyo

Orbbec Astra RGB-D

2D Lidar

Figure 6: Summit XL robot and its URDF model with TF

in RViz.

Human-aware Robot Navigation in Logistics Warehouses

375

jump distance value for 2D laser data segmentation

is chosen with high value to ignore the false alarms

by the selective racks, a highly score of missed detec-

tion is found for the coworker in walking state. Under

those circumstances, severity rate of human damage

is increased. The performance evaluation of 2D laser-

based human detector in logistics warehouses simula-

tion is shown in Table 2. Each single experiment is

run at least 3 times and detection metrics have been

averaged to ensure stable results for each classiﬁer.

For evaluating the performance of human detec-

tor in RGB-D data, the classical supervision machine

learning technique are compared with deep learning

in two computation scenarios: the ﬁrst scenario uses

CPU computation only, the second scenario uses CPU

and GPU computation. Due the classical supervision

machine learning technique, a pre-trained HOG de-

scriptor is used for human with svm classiﬁer. On the

other hand, a pre-trained model (yolov3-tiny) is used

based on COCO dataset. Nevertheless, both HOG-

based and YOLO-based detectors achieve high detec-

tion rates as reported in Table 3.

The YOLO-based detector scores very slow de-

tection time 1.961 (sec) per frame under CPU calcu-

lations only. Conversely, it is run in real time only

by GPU computation with 53 Hz, very fast detection

time 0.008 (sec) per frame and scores less FP-Rate

0.018. While, the HOG-based human detector is run

in real time at 18 Hz on CPU with average detection

time 0.039 (sec) per frame and score a few FP-rate

0.154, also the HOG detector based on CUDA im-

plementation is run at 34 Hz and fast detection time

0.021 (sec) per frame. Under these circumstances,

non-expensive GPU is useful to obtain high speed hu-

man perception, but, when mobile robot platform is

not equipped with GPU, the HOG-based human per-

ception is adequate in real time.

3.2 Navigation Simulation

Logistics warehouse environment simulation is de-

signed under Gazebo simulator where standard par-

allel racks are set up at 3.0 m distances of each other.

For mapping and simulation experiments, we have

created a map by ROS Hector

SLAM (Kohlbrecher

et al., 2011) with two costmaps conﬁgurations: global

costmap and local costmap. In the local costmap, a

8×8 rolling window covers 4 m around mobile robot

and 0.05 m inﬂation radius for obstacles in inﬂation

layer. Besides, local trajectory planning parameters

are tuned with optimal parameters allowing the robot

to maneuver in narrow aisles and let human workers

Hector SLAM, http://wiki.ros.org/hector slam

warehouse equipment

human leg

walking

standing

(multi-poses)

racks

pallet

true detection

missed detection

false positive

Jump distance parameter

Severity rate of human damage

no human detection

human detection

true negative

Figure 7: 2D LIDAR features samples for logistics ware-

house equipment and human leg with strip rectangle for

each object, ﬁrst column is for logistics warehouse equip-

ment (selective racks, euro pallet), while second column is

for human leg (walking and standing states).

pass through. Then, the proposed method is tested us-

ing 2D navigation goal tool in RViz with avoiding ob-

stacles using laser human safety fusion messages (see

Figure 8). Figure 9 shows how the mobile robot avoid

standing and walking coworkers using the proposed

HSA around their localization in logistics warehouses

simulation

4 CONCLUSIONS

In this paper, we study a difﬁcult problem namely hu-

man aware navigation in warehouses where research

results for human detection do not give the expected

good results. To overcome this problem, we proposed

a new strategy based on a 2-steps method: the ﬁrst

step is to use the depth signal for clustering and identi-

fying 3D boxes likely to be humans; while, the second

step is to compute a human presence conﬁdence index

from the RGB signal. For the second step, several

methods are compared and found solutions able to be

run in real time on CPU with the HOG based method

or on GPU with deep learning methods. The choice

of the second step method depends on the available

resources, but at least in the cheapest conﬁguration,

one method can solve the problem. The last contribu-

tion consists of fusing the data coming from the hu-

man detection algorithm and the 2D laser signal to

create a 2D map enabling the robot to navigate not

only safely but also human friendly, i.e., taking into

account the proxemics theory optimal distances be-

tween robots and humans. The experiments in the

simulation environment proved that the method can

improve greatly the robot behaviors when crossing

Video, https://vimeo.com/325696517

ICINCO 2019 - 16th International Conference on Informatics in Control, Automation and Robotics

376

Table 1: Quantitative results for human detection in the 2D laser range data (Linder et al., 2016).

Method Accuracy Precision Recall F1-Measure FP-Rate

SVM 0.914 0.741 0.615 0.672 0.031

Adaboost 0.969 0.895 0.846 0.870 0.014

Random Forest 0.962 0.906 0.766 0.830 0.011

Table 2: Quantitative results for human detection in the 2D laser range data in warehouse simulated environment.

Method Accuracy Precision Recall F1-Measure FP-Rate

SVM 0.310 0.221 0.249 0.235 0.636

Adaboost 0.440 0.285 0.497 0.363 0.588

Random Forest 0.291 0.214 0.331 0.260 0.733

Table 3: Quantitative results for human detection by RGB-D data in warehouse simulated environment.

Method GPU detection time (sec) Hz Accuracy Precision Recall F1-Measure FP-Rate

HOG Off 0.039 18.82 0.718 0.664 0.619 0.640 0.154

HOG On 0.021 34.57 0.741 0.761 0.672 0.713 0.136

YOLO Off 1.961 5.29 0.966 0.906 0.813 0.856 0.021

YOLO On 0.008 53.80 0.974 0.912 0.830 0.869 0.018

(a) 3D box visualization (b) 2D laser fusion (c) Local costmap visualization

Figure 8: Visualisations of 2D laser fusion and local costmap outputs using RViz for logistics warehouse simulation.

human coworker in logistics warehouses. In future,

the proposed framework will be tested on real robot

platform follow through compress human localization

using deep neural network models for reducing mem-

ory and computational requirements and allowing al-

gorithm implementation on the CPU only.

ACKNOWLEDGEMENTS

This research is supported and funded by ERDF

XTERM project action 5: Factory of the future. Grant

HN-002106.

REFERENCES

Arras, K. O., Mozos, O. M., and Burgard, W. (2007). Us-

ing boosted features for the detection of people in

2d range data. In IEEE International Conference on

Robotics and Automation, pages 3402–3407. IEEE.

Choi, S., Kim, E., Lee, K., and Oh, S. (2017). Real-time

nonparametric reactive navigation of mobile robots

in dynamic environments. Robotics and Autonomous

Systems, 91:11–24.

Dalal, N. and Triggs, B. (2005). Histograms of oriented

gradients for human detection. In IEEE International

Conference on Computer Vision & Pattern Recogni-

tion, volume 1, pages 886–893. IEEE.

Goutte, C. and Gaussier, E. (2005). A probabilistic interpre-

tation of precision, recall and f-score, with implication

for evaluation. In European Conference on Informa-

tion Retrieval, pages 345–359. Springer.

Kohlbrecher, S., Von Stryk, O., Meyer, J., and Klingauf, U.

(2011). A ﬂexible and scalable slam system with full

3d motion estimation. In IEEE International Sympo-

sium on Safety, Security, and Rescue Robotics, pages

155–160. IEEE.

Krajn

ık, T., Cristoforis, P., Kusumam, K., Neubert, P., and

Duckett, T. (2017). Image features for visual teach and

repeat navigation in changing environments. Robotics

and Autonomous Systems, 88:127–141.

Linder, T., Breuers, S., Leibe, B., and Arras, K. O. (2016).

On multi-modal people tracking from mobile plat-

forms in very crowded and dynamic environments. In

Human-aware Robot Navigation in Logistics Warehouses

377

(a) Green arrow for target (f/t=0s) (b) Localize human safety (f/t=5s) (c) Wait human to pass (f/t=10s)

(d) Change local plan (f/t=20s) (e) Keep safety crossing (f/t=25s) (f) Reach to target (f/t=45s)

Figure 9: Snap-shots of visualisations for robot navigation in simulation experiment using RViz (local planner: blue line,

global planner: green line, human safety: green circle (1.2 m radius), local costmap: light gray).

IEEE International Conference on Robotics and Au-

tomation, pages 5512–5519. IEEE.

Linder, T., Griesser, D., Vaskevicius, N., and Arras, K. O.

(2018). Towards accurate 3D person detection and

localization from RGB-D in cluttered environments.

In IEEE/RSJ International Conference on Intelligent

Robots and Systems (IROS), Workshop on Robotics

for Logistics in Warehouses and Environments Shared

with Humans.

Lindner, F. and Eschenbach, C. (2017). An affordance-

based conceptual framework for spatial behavior of

social robots. In Sociality and Normativity for Robots,

pages 137–158. Springer.

Malone, N., Chiang, H.-T., Lesser, K., Oishi, M., and Tapia,

L. (2017). Hybrid dynamic moving obstacle avoid-

ance using a stochastic reachable set based potential

ﬁeld. IEEE Transactions on Robotics, 33(5):1124–

1138.

Munaro, M., Lewis, C., Chambers, D., Hvass, P., and

Menegatti, E. (2016). RGB-D human detection

and tracking for industrial environments. In Intel-

ligent Autonomous Systems (IAS-13), pages 1655–

1668. Springer.

Nardi, L. and Stachniss, C. (2017). User preferred behaviors

for robot navigation exploiting previous experiences.

Robotics and Autonomous Systems, 97:204–216.

Quigley, M., Conley, K., Gerkey, B., Faust, J., Foote, T.,

Leibs, J., Wheeler, R., and Ng, A. Y. (2009). ROS: An

open-source robot operating system. In ICRA Work-

shop on open source software.

Redmon, J. and Farhadi, A. (2018). Yolov3: An incremental

improvement. arXiv preprint arXiv:1804.02767.

osmann, C., Hoffmann, F., and Bertram, T. (2017). Inte-

grated online trajectory planning and optimization in

distinctive topologies. Robotics and Autonomous Sys-

tems, 88:142–153.

Song, K.-T., Jiang, S.-Y., and Wu, S.-Y. (2017). Safe guid-

ance for a walking-assistant robot using gait estima-

tion and obstacle avoidance. IEEE/ASME Transac-

tions on Mechatronics, 22(5):2070–2078.

Wahrmann, D., Hildebrandt, A.-C., Schuetz, C., Wittmann,

R., and Rixen, D. (2017). An autonomous and ﬂexible

robotic framework for logistics applications. Journal

of Intelligent & Robotic Systems, 93:419–431.

Yang, H., Fan, X., Shi, P., and Hua, C. (2016). Nonlin-

ear control for tracking and obstacle avoidance of a

wheeled mobile robot with nonholonomic constraint.

IEEE Transactions on Control Systems Technology,

24(2):741–746.

Zimmermann, C., Welschehold, T., Dornhege, C., Burgard,

W., and Brox, T. (2018). 3D human pose estimation

in RGBD images for robotic task learning. In IEEE

International Conference on Robotics and Automation

(ICRA), pages 1986–1992. IEEE.

ICINCO 2019 - 16th International Conference on Informatics in Control, Automation and Robotics

378