Compact, Accurate and Low-cost Hand Tracking System based on LEAP

Motion Controllers and Raspberry Pi

Giuseppe Placidi

1 a

, Alessandro Di Matteo

, Filippo Mignosi

2 b

, Matteo Polsinelli

1 c

and Matteo Spezialetti

2 d

VI-Lab, c/o Dept. MeSVA, University of L’Aquila, Via Vetoio Coppito, 67100 L’Aquila, Italy

Dept. DISIM, University of L’Aquila, Via Vetoio Coppito, 67100 L’Aquila, Italy

Keywords:

Hand Tracking, Virtual Glove, Raspberry Pi, Occlusions, Remote Operating.

Abstract:

The large diffusion of low-cost computer vision (CV) hand tracking sensors used for hand gesture recognition,

has allowed the development of precise and low cost touchless tracking systems. The main problem with CV

solutions is how to cope with occlusions, very frequent when the hand has to grasp a tool, and self-occlusions

occurring when some joint obscures some other. In most cases occlusions are solved by using synchronized

multiple stereo sensors. Virtual Glove (VG) is one of the CV-based systems that uses two orthogonal LEAP

sensors integrated into a single system. The VG system is driven by a Personal Computer in which both a

master operating system (OS) and a virtual machine have to be installed in order to drive the two sensors

(just one sensor at a time can be driven by a single OS instance). This is a strong limitation because VG has

to run on a powerful PC, thus resulting in a not properly low-cost and portable solution. We propose a VG

architecture based on three Raspberry Pi (RP), each consisting of a cheap single board computer with Linux

OS. The proposed architecture assigns an RPi to each LEAP and a third RP to collect data from the other two.

The third RP merges, in real time, data into a single hand model and makes it available, through an API, to

be rendered in a web application or inside a Virtual Reality (VR) interface. The detailed design is proposed,

the architecture is implemented and experimental benchmark measurements, demonstrating the RPi-based

VG real-time behaviour while containing costs and power consumption, are presented and discussed. The

proposed architecture could open the way to develop modular hand tracking systems based on more than two

LEAPs, each associated to one RP, in order to further improve robustness.

1 INTRODUCTION

Computer vision (CV) (Voulodimos et al., 2018) is

actually applied with success for the development of

touchless hand tracking systems due to the low cost,

increasing precision, fastness and high versatility in

recognizing every hand size and silhouette (Oudah

et al., 2020). On the contrary, wearable gloves (WG)

use devices installed directly on the hand and ﬁngers

(Battaglia et al., 2015; Luzhnica et al., 2016; Wang

et al., 2020) which make them haptic and precise.

However, WG expensive, need to be speciﬁcally de-

signed for a hand size and shape and could be very

limiting for the movements. CV-based approaches

https://orcid.org/0000-0002-4790-4029

https://orcid.org/0000-0001-9599-5730

https://orcid.org/0000-0002-4215-2630

https://orcid.org/0000-0001-5786-3999

(Placidi, 2007; Placidi et al., 2013; Erden and Cetin,

2014; Marin et al., 2016; Placidi et al., 2017; Placidi

et al., 2018; Kiselev et al., 2019; Shen et al., 2019;

Yang et al., 2020; Ameur et al., 2020; Placidi et al.,

2021) use the interpretation of video-collecting de-

vices, usually sensors operating also in the visible or

in the infrared (IR) range, placed at a certain distance

from the hand. The key advantage of CV-based sys-

tems is that no physical contact is required and the

movements are free, ﬂuid and natural, being the hand

unforced to wear anything, and it could naturally grip

specialized tools to carry on speciﬁc procedures. For

these reasons, CV hand tracking systems are grow-

ing especially in human-system interaction, to im-

prove the communication between users and comput-

ers, virtual reality environments, devices and robots,

also from remote, to perform medical procedures and

for rehabilitation (Ankit et al., 2011; Avola et al.,

2013; Placidi et al., 2015; Imran and Raman, 2020;

652

Placidi, G., Di Matteo, A., Mignosi, F., Polsinelli, M. and Spezialetti, M.

Compact, Accurate and Low-cost Hand Tracking System based on LEAP Motion Controllers and Raspberry Pi.

DOI: 10.5220/0010880900003122

In Proceedings of the 11th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2022), pages 652-659

ISBN: 978-989-758-549-4; ISSN: 2184-4313

Mahdikhanlou and Ebrahimnezhad, 2020; Chen et al.,

2015; Jin et al., 2016; Erden and Cetin, 2014; Liu and

Zhang, 2014; Carrieri et al., 2016; Zhang et al., 2019;

Petracca et al., 2015; Moro et al., 2016).

Among the CV based gloves, those implemented

with the LEAP (LEAP

) sensor are the most diffuse

(Placidi et al., 2017; Placidi et al., 2018; Kiselev et al.,

2019; Shen et al., 2019; Yang et al., 2020; Ameur

et al., 2020) due to the fact that LEAP is accurate,

low-cost, compact (Bachmann et al., 2014) and can

effectively be used for hand tracking.

LEAP sensor uses 3 IR light sources and two de-

tectors to obtain 3D visual information saved and re-

produced almost simultaneously (more than 60fps)

from the server. One of its advantages is that it is

appropriate for different hand sizes (adults and chil-

dren), as well as for different hand shapes (healthy

people and patients with residual inﬁrmities).

However, all CV hand tracking systems, including

LEAP-based, have to face the problem of occlusions,

occurring when some hand joints are invisible to the

sensors, being occlusions either produced by external

objects grasped by the hand or by parts of the hand

itself (self-occlusions).

Due to occlusions, CV-based tracking systems can

fail to correctly reproduce position/trajectory of the

hand or parts of it because the position of the invisible

parts are guessed, thus resulting in inaccurate and un-

stable representations. This could be negligible when

just raw gestures need to be reproduced, but crucial

when ﬁner movements are used in tele-operated ap-

plications, such as tele-surgery or operations in dan-

gerous environments (Choi et al., 2018; Mueller et al.,

2019; Smith et al., 2020).

The only way to reduce the impact of occlusions

is to use multiple sensors surrounding the scene, in a

multiple-view arrangement, all sensors synchronized

each-other and all contributing to recover the hand

model in real-time (occlusions are solved because

points which are invisible to one sensor can probably

be visible to another).

Recently, several works have been published with

the aim of improving hand tracking accuracy by

combining LEAP data with those of other devices

or data from multiple LEAPs (Marin et al., 2016;

Mahdikhanlou and Ebrahimnezhad, 2020; Yang et al.,

2020; Kiselev et al., 2019; Shen et al., 2019; Placidi

et al., 2017; Placidi et al., 2018; Placidi et al., 2021).

In particular, in (Marin et al., 2016) a LEAP is

supported by a Depth camera. The system has very

good accuracy (regarding gesture recognition) but low

frame rate (15fps) making it not suitable for applica-

tions that requires higher frequency (30fps or greater)

https://www.ultraleap.com/

to track natural hand movements. Moreover, due to

the occlusions between ﬁngers (self-occlusions), the

method performs well only when the hand is in ideal

orientations/positions.

In (Mahdikhanlou and Ebrahimnezhad, 2020) a

LEAP is supported by an RGB webcam to improve

the quality of the recognition of symbols in the 3D

American sign language datasets. Aim of the pro-

posed system is to reduce the ambiguities in gesture

recognition: the RGB webcam is used as an auxiliary

system, being it unable to furnish speciﬁc spatial in-

formation. The same gesture recognition problem for

identiﬁcation of American sign language and Handi-

craftGesture is solved accurately with just one LEAP

(Yang et al., 2020).

Kiselev et al. (Kiselev et al., 2019) use three

LEAPs for gesture recognition. The Authors show

that by increasing the number of sensors, the accu-

racy also increases due the fact that the number of

occlusions decreases. Moreover, the use of multiple

sensors of the same type greatly improves the perfor-

mance of the data integration strategy due to the eas-

iness in comparing similar models. However, since

just one LEAP at time can be driven by a single op-

erating system instance, the used client/server archi-

tecture described in the paper suggests that at least

three different computers have been used (expensive

and critical in terms of synchronization). In addition,

as two of the three LEAPs are coplanar, they mostly

contribute to increase the active region but have low

inﬂuence in reducing occlusions. Finally, the perfor-

mance of the system, in terms of frame rate, has not

been discussed.

Shen et al. (Shen et al., 2019), solve the problem

of occlusions in gesture recognition by proposing the

use of three LEAPs placed with their long axes on the

medium points of the sides of an equilateral triangle.

Though the paper deeply discusses on the system as-

sembly, calibration, data-fusion and results in terms

of position/orientation accuracy, no mention is dedi-

cated to the resulting efﬁciency of the system in terms

of fps.

Virtual Glove (VG) (Placidi et al., 2017; Placidi

et al., 2018; Placidi et al., 2021) is a system based

on the synchronized use of two orthogonal LEAPs

(Fig. 1) for reducing the probability of occlusions.

In particular, in (Placidi et al., 2017), VG was ﬁrst

presented in a raw assembly; in (Placidi et al., 2018)

it was reﬁned and calibrated and, ﬁnally, in (Placidi

et al., 2021) a new strategy for real time data integra-

tion from both sensors was presented and discussed.

Better results regarding occlusions reduction

could have been obtained with three LEAPs on a equi-

lateral conﬁguration, as in (Shen et al., 2019), but

Compact, Accurate and Low-cost Hand Tracking System based on LEAP Motion Controllers and Raspberry Pi

653

with serious problems with the real-time maintenance

(at least 30fps) on a low-cost computer. Though, in

principle, the paradigm of VG (Placidi et al., 2017)

is applicable to any number of sensors placed in any

angular conﬁguration, the choice of two orthogonal

sensors represents a good compromise between opti-

mization/positioning of the region of interest (ROI),

precision and efﬁciency. In fact, position/dimensions

of the ROI and precision with respect to the angle are

leaved to project-related considerations: what is crit-

ical in VG is the possibility of maintaining the real

time (at least 30 fps are necessary for hand motion

ﬂuidity).

Since now, also in its ﬁnal ensemble (Placidi et al.,

2021), VG was limited to the use of just two LEAP

sensors because of maintaining efﬁciency in driving

the LEAP sensors with a single powerful personal

computer.

Aim of the present paper is to design and test a

different hardware architecture to drive LEAP sensors

in VG in order to allow, in the future, the extension of

its concept to more than 2 sensors while maintaining

both cost and power consumption low.

The rest of the manuscript is structured as follows:

Section 2 reviews VG assembly. Section 3 details the

proposed architecture. Section 4 presents experimen-

tal benchmark data and discussion. Finally, Section 5

concludes the manuscript and delineates future devel-

opments.

2 ORIGINAL VG

CONFIGURATION

As discussed above, and referenced in (Placidi et al.,

2018), VG is composed by two orthogonal LEAP

sensors each lodged on a side of a square-angle alu-

minium support of side 25 cm (see Figure 1). The

centre of each LEAP was positioned at 18.5cm from

the internal part of the corner of the support: this

was established to optimize the signal in a cylindri-

cal region of interest (ROI) of radius 10cm and height

21cm while also reducing the effects of IR interfer-

ence between sensors. The two sensors were cali-

brated to allow data integration between them and to

construct a single hand model with data coming from

both sensors in real time (about 60 fps). Different fu-

sion strategies were presented: the original, simple,

used data from just one sensor at a time, those hav-

ing the most favourable view of the hand, in mutual

exclusion (Placidi et al., 2018); the second, smarter

than the ﬁrst (but slower), used data from both sensors

in each frame and merged the joints from one sensor

with those from the other according to the calculation

of temporal behaviour (temporal smoothness) of each

joint (Placidi et al., 2021).

In either cases, data from the sensors had to be

collected by different machines. In fact, a limitation

of the LEAP sensor is that an instance of the Oper-

ating System (OS) can drive no more than one sen-

sor. The simultaneous management of multiple LEAP

sensors was solved by using a virtual machine. The

virtual machine (Slave) was installed on the physical

machine (Master): one sensor was assigned to Master

and the other to Slave, thus allowing to each machine

to instantiate its own, single, driver. Data provided by

both machines were rerouted towards a server (hosted

on the Master) that provided to send data of both de-

vices to the Master machine.

The server driven data from both LEAP sensors

and provided data fusion: data fusion strategy is also

implemented on the Master OS.

The previous stratagem allowed to drive 2 LEAP

sensors in a single machine: however, the host com-

puter was required to be powerful. Another choice

would be to use different PCs for different sensors,

but this would make VG architecture even more com-

plex, expensive and cumbersome than the one used,

at least until now.

Figure 1: Orthogonal LEAP sensors assembly in VG.

3 RPi-BASED VG

CONFIGURATION

Recent advances in the ﬁeld of microprocessors has

allowed the development of cheap, compact and

powerful computers to be used in several applica-

tions. Raspberry Pi (RP) 4 model B (https://www.

raspberrypi.org/) is a single board computer equipped

with a 64 bits quad-core ARM8 Broadcom BCM2711

ICPRAM 2022 - 11th International Conference on Pattern Recognition Applications and Methods

654

Mongo

Database

Server

API

Application

RPi Box

Figure 2: VG architecture based on RPi: each LEAP sen-

sor is associated to one RPi (1 and 2) and data from hand

models of both sensors are passed RP3. On each RPi an

instance of Linux OS runs. Besides OS, RP1 and RP2 just

collect data from the respective sensor; RP3 executes the

software to merge the two hand models into a single one (to

resolve the occlusions) and the graphic interface for the user

interaction.

1.5 Ghz processor, 4GB LPDDR4 Ram and an exten-

sible MicroSD. The operating system is derived from

Linux and it is called Raspberry Pi OS. The cost of

each RP is about 60 USD and the dimensions of the

board are 60mm x 90mm. There also exist more re-

cent and powerful versions but the previous one is that

we have integrated into our project. The new hard-

ware architecture of VG is based on three RPi (i=1:3),

as reported in Figure 2. Instead of using a single pow-

erful PC, we have used three RP: two assigned to drive

the LEAP sensors and the other, in cascade, to collect

data from the others, to perform data fusion in a single

hand model and to run the virtual reality environment.

The RPs are interconnected through a small router us-

ing Ethernet cables. Data from the RP1 and RP2 are

collected into the RP3 by using a web socket commu-

nication protocol.

Since the LEAP driver is available just for X86

processors, we have used the emulator Box86 to em-

ulate X86 instructions into an ARM processor. In fact,

it allows to execute Linux x86 applications on Linux

not X86, as the ARM case. In order to run Box86, the

OS has to be 32-bits. Further, Box86 uses native ver-

sions of some OS libraries, thus ensuring good perfor-

mance. Box86 integrates the dynamic re-compilation

(Dynarec) for the ARM platforms: in this way the ex-

ecution time is from 5 to 10 time faster than using an

interpreter. Box86 source code is released by GitHub

(https://github.com/ptitSeb/box86).

The other two processes necessary to be run on the

RPi are Leapd and Visualizer.

Leapd is a daemon process responsible for the cre-

ation of the numerical hand model to be made avail-

able for the other processes. A connection to Leapd is

possible either from a proprietary SDK or from a web-

socket client. In fact, the daemon hosts a websocket

server. Leapd allows the hand model formation and

its transfer outside the board (we use websocket to

transfer it to RP3).

Visualizer is an application, furnished by the

SDK, which allows to visualize the hand model gen-

erated from the LEAP sensor. The ﬁnal ensemble of

the RPi-box is reported in Figure 3.

Router

Power

Supply

Fan 3

2 1

Fan 2

Fan 1

Figure 3: Upper side view of the RPi-box: it hosts the three

RPi (1, 2 and 3), a multiple USB power supply on the right

(to reduce the number of wires exiting from the box) and a

modem to switch data between RPi (left side, on the ﬂoor).

The cooling system is composed by three fans (black com-

ponents in the bottom). The transparent box is speciﬁcally

produced to host a maximum of 4 RPi.

Each RPi is equipped wit a 64 GB Sandisk SD

card, transfer rate of 120 MB/sec.

The installation procedure starts by compiling

Box86 on the RPi s(in the proprietary OS of RPi, the

installation requires the use of the apt package man-

ager, being the command ”cmake” absent). In order to

save the execution ﬁle of Box86 in the OS, the com-

mand ”make install” has to be executed followed by a

restart of the system.

Once Box86 is installed, it is possible to execute

the deamon of the LEAP sensor ﬁrst by download-

ing the ﬁle ”LeapDeveloperkit” for Linux from the

proprietary site ( https://developer.leapmotion.com/

sdk-leap-motion-controller). The executable ﬁle has

to be in the path x86/usr/sbin: its execution is possible

by running the command ”sudo box86 leapd”.

A similar procedure needs to be executed to in-

stall the process Visualizer. However, it requires the

Qt library: for this reason, Qt has to be installed on

the OS before running the Visualizer. After that, the

command ”sudo box86 Visualizer” is sufﬁcient to run

the Visualizer (as shown in Figure 5).

Finally, the client/server software developed in

(Placidi et al., 2018) with the fusion strategy pre-

sented in (Placidi et al., 2021) has been tested on the

new architecture. As discussed above, the web-based

Compact, Accurate and Low-cost Hand Tracking System based on LEAP Motion Controllers and Raspberry Pi

655

Figure 4: RPi-based VG in its ﬁnal embodiment, composed by: RPi-box (lateral view, the cooling system is now visible),

sensor support, wide-screen, keyboard and mouse.

Figure 5: A screenshot collected from the RP1. The system

resources are shown alongside the Leap Visualizer which

renders a stick model of the hand. In order to represent the

hand movement, colored lines at the tips of the ﬁngers were

added. This negatively affects the Frame Rate.

rendering software is only necessary on RP3. In fact,

the data models from RP1 and RP2 have just to be

passed to RP3 for model fusion and visualization in

the virtual environment.

4 RESULTS

The ﬁrst test was conducted on RP1 and RP2. The ex-

periment consisted on tracking a free moving hand for

a total of 10 minutes. Since RP1 and RP2 are twins

with respect to the hardware and software they use,

the results are very similar and, for this reason, re-

ported only for RP1, in Figure 6. Data in Figure 6

show that RP1 and RP2 are capable to run the whole

Leap Motion Controller software, including the Leap

Visualizer, at a frame rate which is about 33 fps. In

this conﬁguration, the CPU is not completely used (in

the average, just 63% of the CPU is used). For our

purposes, a ﬁrst for of optimization can be obtained

by excluding the Leap Visualizer from RP1 and RP2

(the model is visualized as a result of RP3). In this

case, the resulting frame rate of both RP1 and RP2

is increased to about 37 fps and the CPU usage is re-

duced to 57%, in the average. The reduced CPU usage

suggests that further optimization is possible to boost

the frame rate.

The ﬁnal test was conducted on RP3 in the ﬁnal

assembly for a free hand moving experiment of the

same type, and the same duration, of that used for RP1

and RP2. Figure 7 shows that RP3 is capable to ren-

ICPRAM 2022 - 11th International Conference on Pattern Recognition Applications and Methods

656

0 100 200 300 400 500 600

Acquisition Time(s)

33.5

34.5

35.5

36.5

37.5

fps

fps and avg CPU for RP1 and RP2

AVG CPU 57%

AVG CPU 63%

fps with Visualizer

avg fps with Visualizer

fps with Client

avg fps with Client

Figure 6: RP1 performance, in terms of frame rate and average CPU usage. Both parameters were sampled every 5 seconds

for a total of 600 seconds.

0 100 200 300 400 500 600

Acquisition Time(s)

33.5

34.5

35.5

36.5

37.5

fps

fps and avg CPU for RP3

AVG CPU 60%

fps

avg fps

Figure 7: RP3 performance, in terms of frame rate and average CPU usage. Both parameters were sampled every 5 second

for a total of 600 seconds.

der about 34 fps in the web-based rendering software

with an average CPU usage of 60%. The ﬁnal frame

rate is lower than that in RP1 and RP2 because the

merging operation among the models coming from

RP1 and RP2 is time consuming, though RP3 was free

from the driver of the LEAP device. Also in this case,

the CPU is not completely used and further optimiza-

tion is possible through, for example, the use of smart

Compact, Accurate and Low-cost Hand Tracking System based on LEAP Motion Controllers and Raspberry Pi

657

shape-based segmentation strategies (Franchi et al.,

2009). however, for our purposes, the goal of main-

taining a frame rate of at least 30 fps is completely

ﬁt.

It is important to note that a fundamental role is

assumed by the used SD: in fact, when we tried to

change the SD with one of the same capacity but of

80 MB/s in transfer rate, the performance fell to about

17 fps. This fact authorized us to imagine that the

use an SD with a transfer rate greater than 120 MB/s

could contribute to improve the frame rate above 35

fps, though this has not been attempted and it is out

the scope of our project (our goal is to obtain a ﬁ-

nal frame rate which is above 30fps in the ﬁnal hand

model reproduced in a virtual environment).

However, we have veriﬁed that the hardest role in

our project is assumed by RP3: in fact, its assigned

tasks are data fusion, model reconstruction and re-

production in the graphic interface. The ﬁnal frame

also using the graphic interface was 34 fps when using

the smart fusion strategy presented in (Placidi et al.,

2021). Though still acceptable, that number is very

close to the lower limit, though a further gain in fps

can be obtained by stressing the CPU usage.

The ﬁnal assembly, showing RPi-Box, LEAP sup-

port, wide-screen, keyboard and mouse is reported in

Figure 4.

In this ﬁnal version, keyboard and mouse were

used: a touch-screen could allow the elimination of

both these devices.

5 CONCLUSION

We have demonstrated that a VG can be implemented

by using cheap PCs instead of a costly, bulky and

power consuming PC. In fact, we have implemented

a low-cost and compact embodiment of the VG by

using three light RP 4 model B. The results is a VG

version with a frame rate of 34 fps which is accept-

able for most of VG purposes, though it could be too

low for high (temporal) resolution procedures (such

as medical interventions). Improvements could be ob-

tained along different directions:

1. by using last, most powerful, versions of RP;

2. by using faster SDs;

3. by optimizing the software in order to better use

the CPU power;

4. by using the 64 bit version of the emulator when

it will be available;

5. by using more than three RP.

The last case could improve performance by di-

viding the tasks of RP3 among two RPi in series: one

to collect data from RP1 and RP2 and to fuse the mod-

els and the other to render the ﬁnal model in the VR

environment.

Thanks to the scalability of the system, a further

extension of the proposed architecture would be to

drive more than two LEAP sensors, maybe by imple-

menting the designs proposed in (Kiselev et al., 2019;

Shen et al., 2019), for further reducing occlusions. In

that case, however, the RP acting as a hub would re-

ceive and process information from several RPi and

an efﬁcient and smart data fusion strategy would be

necessary to maintain real-time.

REFERENCES

Ameur, S., Ben Khalifa, A., and Bouhlel, M. S. (2020). A

novel hybrid bidirectional unidirectional lstm network

for dynamic hand gesture recognition with leap mo-

tion. Entertainment Computing, 35:1–10.

Ankit, C., Jagdish, R. L., Karen, D., and Sonia, R. (2011).

Intelligent approaches to interact with machines using

hand gesture recognition in natural way: A survey. In-

ternational Journal of Computer Science & Engineer-

ing Survey, pages 122–133.

Avola, D., Spezialetti, M., and Placidi, G. (2013). Design

of an efﬁcient framework for fast prototyping of cus-

tomized humancomputer interfaces and virtual envi-

ronments for rehabilitation. Computer Methods and

Programs in Biomedicine, 110(3):490–502.

Bachmann, D., Weichert, F., and Rinkenauer, G. (2014).

Evaluation of the leap motion controller as a new

contact-free pointing device. Sensors, 15(1):214–233.

Battaglia, E., Bianchi, M., Altobelli, A., Grioli, G., Cata-

lano, M. G., Serio, A., Santello, M., and Bicchi, A.

(2015). Thimblesense: a ﬁngertip-wearable tactile

sensor for grasp analysis. IEEE transactions on hap-

tics, 9(1):121–133.

Carrieri, M., Petracca, A., Lancia, S., Moro, S. B., Brigadoi,

S., Spezialetti, M., Ferrari, M., Placidi, G., and

Quaresima, V. (2016). Prefrontal cortex activation

upon a demanding virtual hand-controlled task: A

new frontier for neuroergonomics. Frontiers in Hu-

man Neuroscience, 10(53):1–13.

Chen, S., Ma, H., Yang, C., and Fu, M. (2015). Hand ges-

ture based robot control system using leap motion. In

International Conference on Intelligent Robotics and

Applications, pages 581–591. Springer.

Choi, P. J., Oskouian, R. J., and Tubbs, R. S. (2018).

Telesurgery: Past, present, and future. Cureus.

Erden, F. and Cetin, A. E. (2014). Hand gesture based

remote control system using infrared sensors and a

camera. IEEE Transactions on Consumer Electron-

ics, 60(4):675–680.

Franchi, D., Gallo, P., Marsili, L., and Placidi, G. (2009). A

shape-based segmentation algorithm for x-ray digital

ICPRAM 2022 - 11th International Conference on Pattern Recognition Applications and Methods

658

subtraction angiography images. Computer Methods

and Programs in Biomedicine, 94(3):267–278.

Imran, J. and Raman, B. (2020). Deep motion templates and

extreme learning machine for sign language recogni-

tion. The Visual Computer, 36(6):1233–1246.

Jin, H., Chen, Q., Chen, Z., Hu, Y., and Zhang, J.

(2016). Multi-leapmotion sensor based demonstration

for robotic reﬁne tabletop object manipulation task.

CAAI Transactions on Intelligence Technology, 1.

Kiselev, V., Khlamov, M., and Chuvilin, K. (2019). Hand

gesture recognition with multiple leap motion devices.

In 2019 24th Conference of Open Innovations Associ-

ation (FRUCT), pages 163–169. IEEE.

Liu, Y. and Zhang, Y. (2014). Toward welding robot with

human knowledge: A remotely-controlled approach.

IEEE Transactions on Automation Science and Engi-

neering, 12(2):769–774.

Luzhnica, G., Simon, J., Lex, E., and Pammer, V. (2016).

A sliding window approach to natural hand gesture

recognition using a custom data glove. In 2016 IEEE

Symposium on 3D User Interfaces (3DUI), pages 81–

90. IEEE.

Mahdikhanlou, K. and Ebrahimnezhad, H. (2020). Mul-

timodal 3d american sign language recognition for

static alphabet and numbers using hand joints and

shape coding. Multimedia Tools and Applications,

79(31):22235–22259.

Marin, G., Dominio, F., and Zanuttigh, P. (2016). Hand

gesture recognition with jointly calibrated leap motion

and depth sensor. Multimedia Tools and Applications,

75(22):14991–15015.

Moro, S. B., Carrieri, M., Avola, D., Brigadoi, S., Lan-

cia, S., Petracca, A., Spezialetti, M., Ferrari, M.,

Placidi, G., and Quaresima, V. (2016). A novel

semi-immersive virtual reality visuo-motor task ac-

tivates ventrolateral prefrontal cortex: a functional

near-infrared spectroscopy study. Journal of Neural

Engineering, 13(3):1–14.

Mueller, F., Deuerlein, C., and Koch, M. (2019). Intuitive

welding robot programming via motion capture and

augmented reality. IFAC-PapersOnLine, 52(10):294–

299.

Oudah, M., Al-Naji, A., and Chahl, J. (2020). Hand gesture

recognition based on computer vision: A review of

techniques. Journal of Imaging, 6(8):73.

Petracca, A., Carrieri, M., Avola, D., Basso Moro, S.,

Brigadoi, S., Lancia, S., Spezialetti, M., Ferrari, M.,

Quaresima, V., and Placidi, G. (2015). A virtual

ball task driven by forearm movements for neuro-

rehabilitation. In 2015 International Conference on

Virtual Rehabilitation (ICVR), pages 162–163.

Placidi, G. (2007). A smart virtual glove for the hand tel-

erehabilitation. Computers in Biology and Medicine,

37(8):1100 – 1107.

Placidi, G., Avola, D., Cinque, L., Polsinelli, M., Theodor-

idou, E., and Tavares, J. M. R. S. (2021). Data inte-

gration by two-sensors in a LEAP-based virtual glove

for human-system interaction. Multimedia Tools and

Applications, 80(12):18263–18277.

Placidi, G., Avola, D., Iacoviello, D., and Cinque, L.

(2013). Overall design and implementation of the

virtual glove. Computers in Biology and Medicine,

43(11):1927–1940.

Placidi, G., Cinque, L., Petracca, A., Polsinelli, M., and

Spezialetti, M. (2017). A virtual glove system for

the hand rehabilitation based on two orthogonal leap

motion controllers. In Proceedings of the 6th Inter-

national Conference on Pattern Recognition Applica-

tions and Methods - Volume 1: ICPRAM,, pages 184–

192. INSTICC, SciTePress.

Placidi, G., Cinque, L., Polsinelli, M., and Spezialetti, M.

(2018). Measurements by a leap-based virtual glove

for the hand rehabilitation. Sensors, 18(3):1–13.

Placidi, G., Petracca, A., Spezialetti, M., and Iacoviello, D.

(2015). A modular framework for EEG web based

binary brain computer interfaces to recover communi-

cation abilities in impaired people. Journal of Medical

Systems, 40(1).

Shen, H., Yang, X., Hu, H., Mou, Q., and Lou, Y. (2019).

Hand trajectory extraction of human assembly based

on multi-leap motions. In 2019 IEEE/ASME Inter-

national Conference on Advanced Intelligent Mecha-

tronics (AIM), pages 193–198.

Smith, R., Cucco, E., and Fairbairn, C. (2020). Robotic

development for the nuclear environment: Challenges

and strategy. Robotics, 9(4):94.

Voulodimos, A., Doulamis, N., Doulamis, A., and Protopa-

padakis, E. (2018). Deep learning for computer vi-

sion: A brief review. Computational Intelligence and

Neuroscience, 2018:1–13.

Wang, Z., Wang, D., Zhang, Y., Liu, J., Wen, L., Xu, W.,

and Zhang, Y. (2020). A three-ﬁngered force feed-

back glove using ﬁber-reinforced soft bending actu-

ators. IEEE Transactions on Industrial Electronics,

67(9):7681–7690.

Yang, L., Chen, J., and Zhu, W. (2020). Dynamic hand

gesture recognition based on a leap motion controller

and two-layer bidirectional recurrent neural network.

Sensors, 20:2106–2123.

Zhang, W., Cheng, H., Zhao, L., Hao, L., Tao, M., and

Xiang, C. (2019). A gesture-based teleoperation sys-

tem for compliant robot motion. Applied Sciences,

9(24):1–18.

Compact, Accurate and Low-cost Hand Tracking System based on LEAP Motion Controllers and Raspberry Pi

659