Elsevier

Ad Hoc Networks

Volume 114, 1 April 2021, 102445
Ad Hoc Networks

A new deep learning-based distance and position estimation model for range-based indoor localization systems

https://doi.org/10.1016/j.adhoc.2021.102445Get rights and content

Abstract

Many fine-grained indoor localization systems rely on accurate distance estimation between anchors and a target node to determine its exact position. The Received Signal Strength Indicator (RSSI) is commonly used for distance estimation because it is available in most low cost standard wireless devices. Despite the cost efficiency, the distance estimation accuracy in the RSSI-based ranging model needs to be enhanced, especially indoors. The RSSI is sensitive to multiple indoor factors that fluctuate in time and space and lead therefore to its variation. These factors are the origin of the distance estimation error increase in RSSI-based ranging models which in turn raise the position estimation error. Previous works have presented different in-site self-calibration processes to improve the accuracy of distance estimation using RSSI-to-distance samples. It permits to settle the parameters of the RSSI-based ranging model such as the Path Loss Model (PLM) and to mitigate the changing behavior of the RSSI. However, the RSSI measurement depends not only on the distance between the transmitter and the receiver but also on the indoor ambient temperature and humidity variations. Besides, indoor obstacles such as furniture, metallic surfaces or walls have also an impact on the RSSI measurements. We present in this paper a new RSSI-based indoor ranging model using deep learning on collected in-site samples to ensure efficient and autonomic calibration process. This permits to mitigate disturbing factors such as temperature, humidity and noise in order to increase the accuracy of both distance and position estimations. The experimental results have shown that our ranging model has improved not only the precision of distance estimation but also the position estimation in the range-based indoor localization systems.

Introduction

The indoor localization of objects is one of the growing research topics due to plenty services that can be provided in different fields of applications. It can be used for logistic purposes to localize assets inside facilities at real-time [1]. Also, it is useful for security reasons such as monitoring the position of persons or valuable objects inside sensitive and risky closed environments.

Despite its effectiveness outdoors, the Global Positioning System (GPS) is not appropriate for indoor localization because the signals emitted by the GPS-satellites cannot be received correctly inside the buildings [2]. Thus, other technologies such as the Wireless Sensor Networks (WSN) are exploited to design indoor localization systems [3], [4], [5], [6], [7].

In the WSN technology, the objects are localized using Radio Frequency (RF) signals. The WSN-based indoor localization system consists mainly of reference and target nodes. The reference nodes, also called anchors, are mounted at known positions inside the building. The target nodes are attached to each object that we want to localize.

There are two localization approaches in the WSN-based systems: range-free or range-based. In the range-free localization approach, a set of anchors is selected according to the connectivity information with the target node. Thereafter, algorithms such as centroid [8], APIT [9] or DV-hop [10] are used to estimate the target’s location. The range-based approach is more accurate since it relies on point-to-point RF-based measurements between wireless nodes. The Angulation and Lateration are the main range-based techniques [11]. In Angulation technique, at least 2 anchors are required to estimate the position of the target in 2D space. Using built-in multiple antenna arrays, the anchors measure the Angle of Arrival (AoA) of received RF beacons transmitted by the target [12], [13]. In Lateration technique, at least 3 anchors are required to estimate the position of the target in 2D space. The anchors measure their distances to the target node based on the received RF beacons using either a time-based or attenuation-based ranging model.

In the time-based ranging techniques, the anchor estimates the distance by measuring the propagation delay of the beacon sent by the target. The Time of Arrival (ToA) [14] is a time-based measuring model in which the propagation delay is measured from the time of transmitting the beacon to its reception by the anchor. Other ToA-based techniques are presented to overcome the clock synchronization issue between the sender and the receiver. For example, in the two-way ToA [15], also called round-trip ToA, the receiver (e.g. anchor) sends back the received beacon, and therefore the sender (e.g. target) measures the propagation delay which is the half round-trip time of the beacon while taking into account the computational delay in the receiver. Other works [16], [17], [18] have also studied the clock synchronization issue in the ToA-based localization systems and they have proposed solutions to achieve better positioning accuracy.

Besides the clock synchronization issue, the time-based ranging models require wireless nodes with advanced built-in hardware available in the Ultra-wideband (UWB) based modules. As presented in [19], the UWB-based modules provide both ranging and positioning accuracy in millimeter scale thanks to the built-in ADC which has high sampling frequency and enables to read the UWB signals in large bandwidth. However, these devices are still expensive and characterized with a high power consumption. They are therefore not adequate for low-cost indoor localization systems.

In the attenuation-based ranging models, the distance is obtained by measuring the power attenuation of the received RF beacon as it propagates through space. The anchors read the RSSI of the received RF signal and then estimate the distance using RSSI-based ranging model such as the Path Loss Model (PLM). RSSI-based ranging models do not require expensive hardware, since the RSSI metric can be provided by the majority of the wireless modules in markets. This technique can be a good solution to develop low-cost range-based indoor localization systems. However, the RSSI-based ranging model is still less accurate than the one using UWB modules. For this reason, we aim in this paper to improve the RSSI-based ranging accuracy.

The high sensitivity of the RSSI measurements is one of the main factors that leads to decreasing the accuracy of the RSSI-based ranging model. The RF wave can face obstacles in its path, such as walls or furniture. These obstacles disturb the physical characteristics of the RF signal, in particularly its power and consequently the RSSI. In addition, the RSSI shows sensitivity to certain climatic variations such as temperature and humidity [20], [21], [22]. In spite of their slight variation indoors, they have an important impact on the RSSI [23].

Online calibration inside the building enhances the accuracy of the RSSI-based ranging model [24]. It permits to accurately tune the values of its parameters that are related to the indoor environment, such as the indoor attenuation factor. Since the handmade calibration is neither cost-effective nor time-saving, many works in the literature have presented different self-calibration methods. These methods permit to tune the parameters of the ranging model by applying regression techniques on in-site RSSI-to-distance samples. These samples are collected from anchor-to-anchor beacons exchanges since the anchors are in preset positions. However, the studies confirm that the RSSI value can change even at a fixed distance between the sender and the receiver because of the external factors mentioned above. Thus, the accuracy of the ranging model is affected by the RSSI disturbance. For this reason, the self-calibration process needs to be launched periodically to readjust the values of the model’s parameters.

In this paper, we present a new RSSI-based ranging model that takes into account the values of certain factors like temperature, humidity and noise to mitigate their impact on distance estimation. Our ranging model is built using supervised deep learning technique via Artificial Neural Network (ANN) technology. In order to train our model, we have collected in-site samples from exchanged beacons between anchors and Calibration Nodes (CNs). In the previous works, the training or calibration process is carried out using anchor-to-anchor samples. In indoor localization scenario, the anchors are typically fixed on the ceiling [4], whereas the targets are generally placed on the ground. For this reason, we have used the CNs, which are located at known positions in the same space plan of the targets. Thus, the collected samples are more realistic and describe better the attenuation model between anchors and target nodes.

We have proposed a centralized, distributed and simplified training processes to our new ANN-based ranging model. In the centralized design, we use the network’s samples to train the ANN-based ranging model. Subsequently, the obtained model is used for all anchors to estimate the distance. However, the wireless nodes can differently measure the RSSI due to manufacturing miscalibrations. For this reason, we have designed a distributed training process in which each anchor has its own ANN-based ranging model. In extended indoor localization platforms, the number of anchors becomes significant, and in this case, the distributed training process will require more time and powerful computational resources to handle the significant computational load. For this reason, we have designed a simplified training process that provides a better compromise between the lightness of the centralized architecture and the accuracy of the distributed one. To summarize, we present below the main contributions of the paper:

  • Development of a new ANN-based ranging model to learn the implicit relationship between RSSI and distance,

  • Consideration of certain factors like temperature, humidity and noise in the input layer of the ANN-based ranging model to improve the accuracy,

  • Collection of in-site samples using CNs and anchors to ensure efficient training process,

  • Realization of statistical study on the collected samples to theoretically validate the approach and to prove the statistical significance of each input variable on the distance estimation,

  • Design of three training processes: centralized, distributed and simplified with different accuracy and overhead requirements,

  • Test the distance estimation accuracy of the new ANN-based ranging model with different settings, compare it to the existing ranging models and highlight its impact on enhancing the position estimation accuracy in the indoor localization platforms.

The paper is organized as follows: The next section presents the different works in the field. Section 3 presents the statistical study on the training sample to validate our approach theoretically. Section 4 presents the design of the ANN-based ranging model with its different training methods. In Section 5, we present the obtained results related to the ranging accuracy of the ANN-based model and its impact on the position estimation. Also, we compare its performance with the ranging models which are only based on RSSI measurements. Section 6 concludes this paper.

Section snippets

Related works

The Path Loss Model (PLM), also called the Log Distance Path Loss Model, is commonly used in WSN-based indoor localization systems as an RSSI-based ranging model thanks to its simplicity (1) [4]. However, the PLM has many limitations: First, the preliminary settings of the parameters do not take into account the anisotropic characteristic of the wireless channel. For this reason, the PLM needs manual calibration inside the working environment to adjust its parameters. However, the handmade

Statistical study of our approach

Contrary to the current ranging models, the distance output depends on the input variables RSSI, LQI, T and RH in our proposed approach. For this reason, we have carried out statistical tests to study the statistical significance of each input on the distance and to evaluate the new approach. The statistical study is performed on offline collected training samples.

System design

We have developed our new ranging model using the Artificial Neural Network (ANN). The ANN is among the deep learning technologies that approximates any arbitrary function to any degree of accuracy using the multilayer architecture [32]. Our new ANN-based ranging model is built using a supervised training algorithm on the already collected samples. We have proposed different training process architectures to build our ANN-based ranging model which are the centralized, the distributed and the

Results and analysis

We have carried out the cross-validation technique, as stated in [36], to validate our ANN-based ranging model. In this technique, we have split the samples into training and test sets. The training set is used to train and build the model. Subsequently, we have used the obtained model to get predictions based on the test set and we have measured the ranging error. Also, we have tested different models settings such as the number of hidden layers and the number of their neurons. We have tested

Conclusion

We have presented in this paper a new ranging model referred to as ANN-based. It accurately estimates the distance between wireless nodes and enhances the position estimation accuracy of the indoor localization platforms. Contrary to the existing RSSI-based ranging models, the ANN-based one relies not only on the RSSI readings to estimate the distance but also on the factors that are responsible for its disturbance like temperature, humidity and noise. The ANN-based ranging model takes as

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Amir Guidara graduated in Computer Science engineering from the National Engineering School of Sfax (ENIS), Tunisia, in 2012. He received his M.S. degree in Robotics in 2014, from ENIS. Since December 2014, he has joined the Research Unit of Development and Control of Distributed Applications (ReDCAD) at ENIS as a Ph.D. candidate. He is also part of the research team at the laboratory of EIT in HTWK Leipzig, Germany since 2016. In August 2020, he received his Ph.D. degree in Computer Systems

References (37)

  • A. Guidara, F. Derbel, A real-time indoor localization platform based on wireless sensor networks, in: 2015 IEEE 12th...
  • GuidaraA. et al.

    Energy-efficient model for indoor localization process based on wireless sensor networks

  • GuidaraA. et al.

    Lookup service for fog-based indoor localization platforms using chord protocol

  • B. Deng, G. Huang, L. Zhang, H. Liu, Improved centroid localization algorithm in WSNs, in: 2008 3rd International...
  • HeT. et al.

    Range-free localization schemes for large scale sensor networks

  • KumarS. et al.

    An advanced DV-hop localization algorithm for wireless sensor networks

    Wirel. Pers. Commun.

    (2013)
  • ZhangY. et al.

    A range-based localization algorithm for wireless sensor networks

    J. Commun. Netw.

    (2005)
  • LiX. et al.

    Super-resolution TOA estimation with diversity for indoor geolocation

    IEEE Trans. Wireless Commun.

    (2004)
  • Cited by (0)

    Amir Guidara graduated in Computer Science engineering from the National Engineering School of Sfax (ENIS), Tunisia, in 2012. He received his M.S. degree in Robotics in 2014, from ENIS. Since December 2014, he has joined the Research Unit of Development and Control of Distributed Applications (ReDCAD) at ENIS as a Ph.D. candidate. He is also part of the research team at the laboratory of EIT in HTWK Leipzig, Germany since 2016. In August 2020, he received his Ph.D. degree in Computer Systems Engineering from ENIS. His current research focuses on indoor localization systems based on Wireless Sensor Networks and Internet of Things protocols and architectures. Besides, he is focusing on the different technologies in the fields of artificial intelligence and data science.

    Ghofrane Fersi graduated in Computer Science engineering from the National School of Engineers of Sfax (ENIS), Tunisia, in 2008. She received her M.S. degree (DEA) in Systems of Information and dedicated New Technologies (NTSID) in 2009, from ENIS. Since December 2010, she has joined the Research Laboratory of Development and Control of Distributed Applications (ReDCAD) at the National School of Engineering of Sfax (ENIS) as a Ph.D. candidate. She received the PhD. degree in computer science from ENIS, University of Sfax in 2014. Since 2013, she has been an Assistant Professor in the Department of computer science in the university of Sousse, Tunisia. She becomes Associate Professor since 2016. Her current research focuses on Distributed Hash Table-based protocols in Wireless Sensor Networks, Internet of Things protocols and architecture.

    Maher Ben Jemaa obtained his diploma of Engineer in Computer Science from the National School of Computer Sciences ENSI (Tunisia) in 1989 and his Ph.D. from the Department of Electrical Engineering at the National Institute of Applied Sciences (INSA) Rennes (France) in 1993. He joined the National School of Engineers of Sfax (ENIS) as Assistant Professor of Computer Science in the Department of Computer Science and Applied Mathematics in 1995. He became an Associate Professor in 1997 and a keynote professor in March 2011. He is full professor since Mai 2016. His current research areas include Fault Tolerance of distributed systems, Quality of Service in Ad hoc Networks and routing issue in Wireless Sensor Networks.

    Faouzi Derbel is professor for Smart Diagnostics and Online Monitoring at Leipzig University of Applied Sciences. He received the M.Sc. degree in electrical engineering from the Technical University of Munich, Germany, in 1995 and the Ph.D. degree from the University of the Bundeswehr, Munich, in 2001. From 2000 to 2012 he was in different positions in the industrial area e.g. strategic product manager and systems engineer, responsible for wireless and future technologies within Siemens Building Technologies in Munich (Germany) as well as Head of Research and Development within Siemens Building technologies in Mühlhausen (Germany) and QUNDIS Advanced Measuring Solutions (Germany). In 2006 he won together with his team the “Product Innovation Award” within the company Siemens AG on worldwide level and in 2009 the “Product Success Story Award” of Texas Instruments. Prof. Derbel is member of ETSI ERM TG28 dealing with short range wireless devices within the European Telecommunication and Standardization Institute, as well as the technical group Smart Cities within the “Deutsche Kommission Elektrotechnik, Elektronik und Informationstechnik im DIN und VDE”. He is editor in Chief of the ASSD (Advances in Systems, Signals and Devices, Issues on Communication and Signal Processing and Power Systems and Smart Energies, De-Gruyter Verlag, Germany) and holds many awards and patents. His research activities focus on power aware design of wireless sensor networks and sensor systems with smart signal processing as well as state estimation of machines and smart grids.

    View full text