When Wireless Localization Meets Artificial Intelligence: Basics, Challenges, Synergies, and Prospects

Abstract

The rapid development of information communication and artificial intelligence (AI) technology is driving innovation in various new application fields such as autonomous driving, augmented reality, and the metaverse. In particular, the advancement of wireless localization technology plays a great role in these cutting-edge technologies. However, traditional wireless localization systems rely on the global navigation satellite system (GNSS), which is ineffective in indoor or underground environments. To overcome this issue, indoor positioning systems (IPS) have gained attention, and various localization techniques utilizing wireless communication were studied. Subsequently, AI technologies are improving the performance of wireless localization and addressing problems that were previously difficult to solve. In this paper, we summarize wireless localization techniques and define the factors that impede their performance. Furthermore, we categorize AI algorithms and present examples of how they can be used to address these hindering factors. Finally, we propose open research directions and prospects for AI-assisted wireless localization.

1. Introduction

Vibrant advancements in modern technology have expanded living areas and necessitated the identification of users’ locations in real time. Traditional methods, such as compasses and maps, have been supplanted by contemporary alternatives, e.g., global positioning systems (GPS). The global navigation satellite system (GNSS), which employs satellites to transmit users’ real-time locations, has enabled location-based systems (LBS) in fields, e.g., aircraft, ships, and private vehicles, resulting in enhanced user convenience. The emergence of state-of-the-art technologies, e.g., autonomous driving and virtual reality, necessitate precise, real-time user location data to guide users to their desired destinations and enhance the quality of their experiences. Thus, it is obvious that localization technology is essential to the use of various LBS services.

During the initial phases of development, location measurement was limited to rough approximations in outdoor environments due to the sole accessibility of GPS or GNSS. However, present-day research focuses on developing ultra-precise localization systems that can extend services to indoor environments and minimize errors from meters to centimeters. In scenarios where satellite signals cannot be accessed, an indoor positioning system (IPS) has become necessary for accurately determining location within indoor or underground spaces [1,2]. Since most of our daily activities occur indoors, IPS can be utilized in various fields such as guiding individuals within buildings, tracking emergency patients’ locations, inventory management in retail stores and warehouses, and more.

As mentioned above, the demand for LBS is increasing nowadays in various ways. Consequently, the advancement of localization technology must keep pace with evolving requirements, a demand well-matched by the rapid progress of mobile communication systems such as cellular and Wi-Fi. However, besides the primary goal of improving wireless communication system specifications, wireless localization encounters its own set of challenges. These include environmental interference, accuracy limitations in densely populated spaces, and potential privacy implications stemming from the collection of location data.

For this reason, active research on artificial intelligence (AI) has also made significant contributions in this regard [3,4,5], which can extract unveiled patterns or hidden features to solve exceedingly complex problems in a simpler manner. More specifically, machine learning is a family of algorithms which are developed through data to enable computers to learn and identify patterns, unlike traditional explicit programming where developers implement the algorithms themselves. Furthermore, deep learning, a special subset of machine learning, uses artificial neural networks (ANNs) that emulate the human brain to extract statistical characteristics of data. These AI techniques are expected to have new roles, such as extracting features that are difficult to achieve through conventional mobile communication system optimization or significantly reducing complexity. Consequently, when combined with localization, these techniques can enhance measurement processes to improve localization accuracy, and resource utilization efficiency, and overcome existing limitations of localization technology.

In this paper, we briefly introduce the wireless communication technologies capable of inducing wireless localization. Next, we outline wireless localization techniques and define the factors that hinder their performance. Additionally, we categorize machine learning algorithms and provide examples of how they can be implemented to mitigate these obstacles. Lastly, we propose potential future research directions and prospects for wireless localization.

2. Wireless Communication Technology

Wireless localization is based on the transmission and reception of analog/digital signals composed of overlapping radio waves. Wireless communication has evolved with varying considerations such as frequency band and network capacity, tailored to specific applications and unique attributes. With the continued development of various wireless communication technologies, they are also being utilized in various applications, as is evident in the localization field [6]. In addition to the GNSS technologies such as GPS and Galileo that are already utilized for outdoor localization, there is an increasing demand for indoor localization due to the inefficacy of GNSS technology in such environments. Related research is being conducted to utilize various wireless communication technologies for effective indoor localization, and some of the commercially available wireless communication technologies include Wi-Fi, ultra-wideband (UWB), Bluetooth, and Zigbee, as shown in

Table 1. Comparison of wireless communication technologies for indoor localization.

	Nominal Range	Maximum Signal Rate	Operating Frequency
Wi-Fi	100–500 m	10 Gbps	$2.4$ GHz; 5 GHz; 6 GHz
UWB	30–100 m	480 Mbps	$3.1$–$10.6$ GHz
Bluetooth	10–100 m	3 Mbps	$2.4$ GHz
Zigbee	10–100 m	250 Kbps	915 MHz; $2.4$ GHz

.

Wi-Fi-based localization uses existing access points (APs) to estimate the location of a device. Initially, the target space was divided into multiple cells, and the received signal strength indicator (RSSI) collected in each cell was used to construct a radio frequency map. Recently, round trip time (RTT) information has been utilized to overcome localization errors. Using existing wireless local area networks (WLANs)offers the advantage of cost-effective initial infrastructure deployment. However, a corresponding drawback emerges as errors can be caused by path loss, and the initial connection time is long due to the security process, authentication protocols, and concurrent wireless frequency congestion.

UWB [7] is a technology that uses very short pulses over a wide frequency band, resulting in reduced interference between pulse signals and good penetration ability, allowing for more precise and sophisticated localization. It has the advantage of being suitable for next-generation localization technology and the ability to transmit large amounts of data with low power. However, there is a drawback that it must solve the interference problems caused by overlapping with the bandwidth of Wi-Fi and Bluetooth, which are currently widely used.

Bluetooth and Zigbee [8] are low-cost communication systems used for indoor localization. Bluetooth uses multiple beacons as anchors to estimate the location, while Zigbee determines the location by using the signal strength of the transmitted signal from a reference node and the received signal at a tag. Bluetooth is economically favorable due to its low cost, the mass production of chipsets, and low power consumption during operation. Zigbee is strategically designed to prioritize low data transmission, resulting in minimized network wait times and efficient power consumption. However, this design choice does come with the trade-off of comparatively slower data transmission speeds. Consequently, devices interconnected via Zigbee benefit from shortened waiting intervals while communicating with each other or with a central hub. This streamlined process facilitates swift connection establishment, effective exchange of small data payloads, and rapid responsiveness to commands or inquiries. Nevertheless, it is worth noting that both Bluetooth and Zigbee share a notable limitation in terms of signal resolution, which is an aspect warranting consideration.

3. Wireless Localization Technology

As mentioned above, wireless communication technologies are diversely being utilized for localization. In fact, communication signals have various physical factors such as power, delay, and phase, which are used for wireless localization. The algorithms for estimating the position can be divided into three categories: range-based techniques, which are based on physical factors of radio waves, range-free techniques, which use known information (such as pre-processed data) rather than physical factors, and hybrid techniques, which utilize the two techniques appropriately to complement each other. These three categories are summarized as shown in Figure 1.

3.1. Range-Based Techniques

Range-based techniques involve estimating a device’s location by considering physical factors (such as signal strength, time, and angle) of transmitted and received signals, along with geometric and linear algebraic methods like trilateration and triangulation between the device and anchor/AP/base stations (referred to as anchor).

First, the distance information between the device and the anchor is obtained using the physical factors of the signal, and then the location of the device is estimated using trilateration or multilateration. In this case, trilateration or multilateration refers to calculating the device’s location as the intersection point of circles (or spheres) centered at each anchor, with each circle’s radius being the distance from the anchor to the device, using at least $n+1$ anchors in n-dimensional space, as shown in Figure 2.

The distance between a device and an anchor is measured by the change in the physical factors of radio waves during transmission and reception. Thus, by using signal strength and propagation time, received signal strength (RSS) and time of arrival (ToA) can be used to estimate the distance between the transceivers.

3.1.1. Received Signal Strength (RSS)

One of the simplest methods for distance measurement is the utilization of received signal strength (RSS). The distance between the device and the anchor through RSS can be obtained through the following equation.

$$P_R=P_0-10nlog\left(\frac{d}{d_0}\right).$$

(1)

d, $d_0$, $P_R$, $P_0$, and n are the distance between the device and the anchor, the reference distance, the actual RSS, the reference RSS, and the path loss exponent, respectively. Here, $P_0$, $d_0$, and n are fixed values, thus the distance d can be simply calculated by substituting $P_R$. Although this method is quite simple regarding hardware implementation, there is an issue of knowing the path loss exponent n in advance, and it is very vulnerable to the signal-to-noise ratio (SNR).

3.1.2. Time of Arrival (ToA)

ToA is the time it takes for a wireless signal to travel from the transmitter to the receiver. It can be used to estimate the distance between the device and the anchor by using the following equation.

$$d=T\times c.$$

(2)

T represents the propagation delay time, d represents the distance between the device and the anchor, and c represents the speed of light. Two methods using the ToA technique, one-way ranging (OWR) and two-way ranging (TWR), are widely used, as shown in Figure 3.

In the case of OWR, ToA can be obtained through the duration of transmission and reception time. While OWR calculates the time on a single path, TWR calculates it using RTT, which means the time from the moment the transmitter transmits the wireless signal to the moment the receiver receives the signal and then retransmits it to reach the transmitter. Through this, ToA can be calculated by subtracting the receiver waiting time (such as network delay time within RTT) from RTT and dividing it by half. Comparing the two methods, OWR is relatively fast and consumes less power in distance measurement; however, TWR does not require precise synchronization for the localization network and has the advantage of lower initial configuration cost as it is robust against network synchronization.

In the case of the ToA technique, if the reference clocks of the anchor and the device are different, it is not possible to obtain the absolute time and the two reference clocks must be identical.

3.1.3. Time Difference of Arrival (TDoA)

TDoA is the difference between the distances between a device and two anchors, divided by the speed of light, and estimates the location of the device by finding the intersection of the parabolic curves with the two anchors as the focus, as shown in Figure 4. This technology is used even when synchronization between the device and the network is not established, by simply obtaining the time difference from multiple anchors. On the other hand, recent research has aimed at translating the time difference of arrivals for two signals transmitted at different frequency bands simultaneously into the distance to the receiving device.

3.1.4. Angle of Arrival (AoA)

AoA refers to the direction of the signal coming from the device toward the anchor. By leveraging and measuring the delay difference of the arrival at individual antenna array elements, AoA-based methods use the anchor’s antenna arrays to determine the angle at which the transmitted signal arrives at the anchor as shown in Figure 5. The received signal angles from two or more anchors are measured and the position of the device is estimated using trigonometry.

3.2. Range-Free Techniques

Range-based technologies can have limitations in terms of accuracy as wireless signals are influenced by various factors such as distance, spatial structure, humans, objects, temperature, humidity, etc. In contrast, range-free technologies estimate the location of a device by using pre-existing information, rather than relying solely on the physical factors of the wireless signal.

3.2.1. Proximity

A representative proximity-based technology is Cell-ID technology. Cell-ID is a technology that uses the unique number assigned to each base station in wireless communication. The localization technology based on Cell-ID determines the position of the device based on the location of the base station it is connected to as shown in Figure 6. It is simple to determine the location through information processing, while the disadvantage is that the entire communication range of the base station is the error range, making it difficult to achieve precise localization.

Another proximity-based technology is device-to-device (D2D) communication. D2D communication means direct communication between devices without passing through the anchor. This communication method can expand the active area by continuously connecting devices through a relay method. This method can also be applied to localization, where the location of the next device can be measured based on the device that has determined its location, without direct communication with the anchor. This reduces the load on the anchor that communicates with multiple devices; however, the error can increase if the connection between devices becomes overlapping.

3.2.2. Dead Reckoning

Dead reckoning (DR) is a localization technique that estimates a device’s current position based on the distance and direction it has traveled over a certain period of time. This technique uses inertial navigation systems to measure the movement of a device and continuously updates its position in real time. It involves both movements relative to fixed axes indicated by red and green arrows, as shown in Figure 7, with reference to the device, as well as rotations represented by the blue arrow.

DR provides stable and continuous position tracking; however, it accumulates errors over time. Moreover, there is also a technology called pedestrian dead reckoning (PDR), which incorporates the estimation of a pedestrian’s direction and stride length into dead reckoning by detecting steps.

3.2.3. Fingerprinting

Fingerprinting technology is a technique that determines a device’s location based on the strength of the received signal and the reference location in a radio map with the most similar signal strength as shown in Figure 8. There is a pre-condition that a radio map must be pre-made. During actual localization, only the process of comparing the received signal and the radio map is performed, thus the location can be estimated relatively quickly. The radio map is a kind of set of on-grid values created in advance by measuring the strength of the signal received from the anchor at several reference locations with known positions. The advantage of this technology, if utilized based on Wi-Fi, is that localization is possible without additional infrastructure or equipment, using only existing commercial devices.

3.3. Comparison

Wireless localization technologies can be classified into range-based technologies and range-free technologies, depending on whether or not they use physical factors of communication signals.

Table 2. Comparison of range-based techniques and range-free techniques.

Comparison Aspects	Range-Based Techniques	Range-Free Techniques
Accuracy	High	Moderate
Power Consumption	High	Low
Deployment	Hard	Easy
Additional Hardware	Required	Not required
Preparation Process	Not required	Required

summarizes the advantages and disadvantages of these two technologies.

The range-based technology estimates the location in real-time by measuring physical factors, does not require prior work, and is accurate but has the characteristic of being resource-intensive and the geometric calculation process is complex. On the other hand, the range-free technology does not measure the physical factor of the signal but requires prior work and has the advantage of being a resource-efficient and simple comparison process.

In more detail, the advantages and disadvantages of range-based technologies can be compared as shown in

Table 3. Comparison between range-based techniques.

Comparison Aspects	RSS	ToA	TDoA	AoA
Accuracy	Moderate	High	High	High
Attenuation	High	Low	Low	Moderate
Cost	Low	Moderate	Moderate	High

. RSS-based technology requires simple hardware implementation, leading to low costs, but it also comes with lower accuracy. Its accuracy is compromised by factors such as multipath and reflection. ToA and TDoA technologies have lower attenuation and cost, but they are influenced by hardware resolution and LoS environments. AOA technology offers high accuracy but comes with a higher cost by employing complex antenna arrays. Additionally, it is susceptible to environmental conditions such as multipath, scattering, and NLoS.

Similarly, range-free technologies can be compared as shown in

Table 4. Comparison between range-free techniques.

Comparison Aspects	Proximity	Dead Reckoning	Fingerprinting
Accuracy	Moderate	Moderate	High
Energy Efficiency	High	High	Low
Cost	Low	Moderate	High

. Proximity technology enables moderate accurate positioning with inexpensive implementation, leading to high energy efficiency. The accuracy and cost of DR technology depend on the measurement device, and estimating movement distance and direction in DR technology may result in some energy consumption. Fingerprinting technology’s accuracy relies heavily on the precision of the radio map. Although creating a precise radio map can incur high costs and may have relatively low energy efficiency, it provides exceptionally high accuracy.

3.4. Hybrid Techniques

While both technologies have distinct properties, there is a growing research effort to integrate the two groups. It aims to leverage their unique strengths and weaknesses to complement each other effectively. When combining various technologies, it is possible to categorize the approach into tightly coupled and loosely coupled based on the interdependencies between the technologies [9]. The tightly coupled approach involves closely integrating various technologies and data sources to achieve a high level of synergy. In positioning technology, this means that each technology simultaneously utilizes the data of others, complementing each other to achieve high accuracy. For example, enhanced Cell-ID is a technology that improves accuracy by adding RTT and AoA information between the anchor and the device to the Cell-ID method. On the other hand, the loosely coupled approach indicates that components operate independently, integrating information only when necessary. In positioning technology, this means that each technology independently measures location data, which is selectively and appropriately corrected depending on the situation. For instance, there are technologies that combine fingerprinting and PDR [10], which resolve the error of judging the reference location unrelated to the device’s position and the cumulative error of the inertial navigation device due to the flow of time, as the device’s position. Depending on the circumstances, a suitable combination between these two approaches can enhance positioning accuracy.

4. The Factors for Degrading Performance of Localization

As mentioned above, various techniques are used for localization. The performance of localization is hindered by various factors in the application of these techniques [11,12,13,14]. However, in this paper, we will introduce them in three main categories.

4.1. NLoS Events

Non-line-of-sight (NLoS) events refer to situations where the direct path between the anchor and the device is obstructed by an object, such as a building or a car. In such cases, the wireless signal is reflected or diffracted around the obstacle and reaches the receiver via a different path, not the direct path. This can cause problems in localization performance.

As a main problem, NLoS can lead to multipath. Multipath refers to the propagation phenomenon that results in radio signals reaching the receiver by two or more paths, which occurs when wireless signals are reflected from surrounding surfaces and arrive at the receiver through many ways. This causes phase and amplitude distortion, interference, and signal fading, leading to inaccuracies in localization performance.

Another problem is the difficulty in accurately predicting and modeling the behavior of the reflected signals. The reflected signals are affected by the shape and location of the obstacles, as well as the frequency and polarization of the signal. Therefore, it is difficult to predict the exact path of the reflected signals, which can lead to localization errors.

4.2. Spatiotemporal Asynchronization

Spatiotemporal asynchronization refers to the discrepancy in the clock pulse and relative position. The discrepancy in the clock pulse can occur when the system clock pulses of the anchor and devices are not the same. This causes a difference in time standards of the wireless signal, which impairs the localization performance.

Next, the discrepancy in relative position can occur when the signal is transmitted from a moving transmitter or when the receiver is moving relative to the transmitter. In such cases, the received signal may arrive at different times and from different directions, causing errors in the estimation of the transmitter location. This can significantly affect the localization performance, as it can result in inaccurate range and angle measurements, which are used to determine the location of the transmitter.

4.3. Signal Quality Deterioration

The quality of the wireless signal can be degraded by various factors, including noise and the medium through which the signal travels. Noise can be produced by various sources, such as other wireless devices, and electrical equipment. The presence of noise can make it difficult for the receiver to distinguish between the desired signal and the unwanted noise, which can result in bit error, packet loss, and reduced signal strength in the estimation of the location.

Moreover, atmospheric conditions can also have a significant impact on the performance of localization systems that rely on radio signals. This is because atmospheric conditions can cause the radio signal to be attenuated or refracted, leading to errors in signal timing or strength. For example, changes in temperature, humidity, and pressure can make the atmosphere more or less dense, which in turn affects the speed of radio waves. This can cause signal delays or phase shifts, leading to errors in range measurements.

Additionally, atmospheric phenomena such as ionospheric scintillation, which is caused by charged particles in the ionosphere, can cause rapid fluctuations in signal strength and phase. This can lead to errors in position estimates or even complete signal loss.

5. A Brief Introduction to AI

The recent explosion of numerous mobile applications, particularly those powered by AI, has sparked intense debates over how wireless localization will develop in the future. As the problems of wireless localization become more complex, AI technology is receiving increasing attention as a means to find better solutions to the various problems and issues that arise, which cannot be addressed by conventional methods alone. Recently, there have been studies on utilizing AI technology to address the issues of degraded localization performance caused by the issues mentioned above. In this regard, AI refers to the artificial implementation of some or all of the intellectual abilities that humans possess. If utilized, it can solve the same problems more efficiently and accurately than humans. Although AI is divided into various fields, in the field of localization, it is mainly used in conjunction with machine learning and deep learning to complement the limitations of existing conventional localization technologies [15].

Machine learning and deep learning have led to breakthroughs in various fields, such as natural language processing, computer vision, and speech recognition. These advancements have enabled machines to understand and interpret complex data, make informed decisions, and perform tasks that traditionally require human intelligence. As a result, the integration of machine learning and deep learning technologies has not only revolutionized industries but also paved the way for innovative applications. In general, the ultimate goal of machine learning and deep learning is to discriminate between different types of data, and in the process of discriminating, machines define the features of each data point. The defined features are multiplied by weights to discriminate the data, and the process of training the model to perform this discrimination accurately is machine learning [16,17]. In other words, machine learning is the field of developing algorithms that allow machines to learn, and in machine learning, humans define the features directly, which are called handcrafted features.

In addition, as one of the technical models of machine learning, deep learning involves creating a model by stacking multiple layers of ANN that mimic the human brain and optimizing it [18,19,20]. In other words, deep learning involves extracting the characteristics of data through ANN during the learning process. Sometimes, new paradigms that go beyond human experiential categories can be extracted as they are extracted by mechanical algorithms.

In other words, since accurate discrimination is possible with the selection of appropriate features, many studies using AI have been conducted to find good features. Deep learning can be applied to problems that are difficult to identify with handcrafted features, such as NLoS discrimination in localization techniques. However, there are limitations to using only deep learning, such as insufficient data or the inability to exclude inappropriate data such as bias and distortion, and various studies are underway to address these issues.

We classified AI algorithms into two categories as shown in Figure 9 based on whether the algorithm used ANN or not. We will briefly introduce these algorithms in the following two subsections.

5.1. Machine Learning Algorithms without ANN

The algorithms, such as K-nearest neighbor (K-NN) [21], support vector machine (SVM) [22], and random forest (RF) [23], have been widely used since the early days of machine learning as they have relatively simple structures as classification models.

5.1.1. K-Nearest Neighbor (K-NN)

The K-NN algorithm is a classification model that compares the features of samples to be compared on a graph with n features and uses the parameter K, defined by the user, to classify by voting K nearest samples based on the distance between data points, as shown in Figure 10. When K is too small, overfitting may occur, causing the model to be closer to the training samples’ distribution than to the actual distribution, resulting in inaccurate prediction of new data. Conversely, if K is too large, underfitting may occur, as it becomes more dependent on the number of samples rather than the characteristics of the data, making it difficult to identify patterns based on features. Therefore, selecting an appropriate value for K is important.

5.1.2. Support Vector Machine (SVM)

The SVM algorithm is a model that defines a reasonable decision boundary for classification. It defines this boundary based on the existing data and determines which side of the boundary new, unclassified data belong to, as shown in Figure 11.

To do this, the algorithm follows a similar approach to K-NN by graphing the samples on n characteristic axes and then defining an appropriate decision boundary. The appropriate decision boundary is to maximize the margin, which is the perpendicular distance between the line and the data. Outliers are data points that deviate significantly from the range of the same characteristic samples. The algorithm seeks to distinguish data points according to their characteristics while sometimes ignoring outliers that fall outside of a certain range. When the margin is too small, overfitting may occur, and when it is too large, underfitting may occur. Therefore, proper handling of outliers is important.

5.1.3. Random Forest (RF)

A decision tree is a model that divides data into true/false based on specific criteria selected by questions. It does not require data scaling, such as feature normalization or standardization. However, if modeled without any parameters, it can lead to overfitting.

However, the RF algorithm is a model that is designed to compensate for the overfitting of decision trees, forms multiple decision trees, and simultaneously passes newly selected data points from the dataset through each tree, as shown in Figure 12. Afterwards, a vote is taken to select the result with the most votes as the final result. RF has the advantage of maintaining the advantages of decision trees while compensating for their shortcomings, but it does not work well with high-dimensional and sparse data.

5.2. Machine Learning Algorithms with ANN

An ANN is a computing system that mimics the appearance of neurons in the human brain and is an algorithm that forms the basis of deep learning, as mentioned earlier. An ANN consists of an input layer representing input signals, a hidden layer consisting of nodes that correspond to neurons in the brain, and an output layer representing the final result value. The nodes in the hidden layer receive multiple input signals, multiply the weight representing the importance of each signal for each signal, and output the sum. At this time, the weight is not calculated by any formula, but is initially given an arbitrary value and gradually updated by finding the value with the least error between the output value and the correct answer when inputting the training dataset input value.

A deep neural network (DNN) is an advanced form of ANN with more than two hidden layers, as shown in Figure 13. The computer works by repeatedly creating classification labels based on the input and output layers and separating the data to construct the optimal hidden layers. DNNs are capable of modeling complex data using fewer units compared to when there is only one hidden layer. They are currently widely used in various forms through repeated and pre-training, and backpropagation techniques. Some examples of the most actively used algorithms include convolutional neural network (CNN) [24], recurrent neural network (RNN) [25], autoencoder (AE) [26], graph neural network (GNN) [27], and generative learning models [28,29].

5.2.1. Convolutional Neural Network (CNN)

CNN refers to a type of neural network that extracts features from data by identifying patterns in the extracted features, unlike the conventional method of extracting knowledge from data as shown in Figure 14. The hidden layers of CNN are composed of convolutional layers and pooling layers. Convolution involves creating a matrix called a kernel and sequentially convolving the kernel with a matrix of the same size as the data, overlapping from the beginning to the end. This process extracts the features of the data, and after passing through an activation function, a compressed result with reduced parameters is obtained, creating a convolutional layer based on a feature map. Subsequently, pooling is performed on the created feature map. This process reduces noise and creates a pooling layer with a smaller size compared to before.

5.2.2. Recurrent Neural Network (RNN)

RNN is a specialized neural network for sequential data, which uses its internal recurrent structure to reflect past learning in current learning. Unlike other neural networks where the direction of learning progress is unidirectional, this neural network is composed of a structure where the result from the activation function of a node in the hidden layer is sent to the output layer direction and then again as the input of the next calculation in the hidden layer node. It has the characteristic of sharing parameters at multiple stages, resulting in a small number of parameters to train.

The long short-term memory (LSTM) model, a type of RNN, is designed to overcome the drawback of traditional RNN that cannot remember information that is distant from the output. It is a neural network structure that enables both long-term and short-term memory. The LSTM model also has a recurrent structure like RNN, but it is composed of four layers that exchange information with each other and repeat the process.

5.2.3. Autoencoder (AE)

AE is a neural network composed of an encoder and decoder, as shown in Figure 15, and has a structure that encodes and decodes inputs to produce copies of the outputs. In this case, the neural network is utilized in various ways by applying constraints to it, such as reducing the number of nodes in the hidden layer to compress data or training the neural network to restore the original input from noisy input by adding noise.

5.2.4. Graph Neural Network (GNN)

A GNN is a machine learning model that learns information within a graph by utilizing the feature vectors of nodes and their neighboring nodes, as shown in Figure 16. In this process, the feature vectors of each node and the weights of the edges connecting the nodes are calculated, and a new feature vector for each node is computed through interactions with its neighboring nodes. This process is repeated in multiple layers to update the feature vectors of nodes, and each node gradually incorporates the characteristics of its neighboring nodes to include comprehensive information.

5.2.5. Generative Learning Model

Generative learning models generate similar data following the probability distribution of the given training dataset by learning from it. In other words, the most important thing in the generative model is to learn the probability distribution of the training data. There are two main approaches to generate data based on the learned probability distribution: the explicit density approach, which directly models the probability distribution of the training data, and the implicit density approach, which can generate samples without explicitly modeling the probability distribution of the training data.

In the explicit density method, there are two approaches: the tractable density method, which directly calculates the probability distribution of training data based on a defined model, and the approximate density method, which approximates and estimates the probability distribution simply due to the complexity of the model. One of the representative algorithms of the approximate density method, as shown in Figure 17, is the variational autoencoder (VAE), which is very similar to an AE but encodes latent variables of the input as a normal distribution represented by mean and variance instead of a single value. The VAE is a probabilistic model, allowing flexible calculations based on diversity, but its performance is not as good as that of fixed values.

Meanwhile, one of the representative algorithms among the indirect distribution methods is the generative adversarial network (GAN). This algorithm consists of a generator and a discriminator, which compete with each other to improve their respective performances. The generator’s role is to create data that are similar to real data so that the discriminator cannot distinguish between them, while the discriminator’s role is to appropriately distinguish between the data generated by the generator and real data. Through the repeated process of the discriminator distinguishing between the generator’s data and real data, the generator learns to create data that are more similar to real data. Eventually, the generator generates data that are most similar to real data.

6. AI-Assisted Wireless Localization Technology

We previously defined the limitations of the existing localization technologies. Recently, various AI-assisted wireless localization technologies have been studied to overcome the limitations of existing wireless localization technologies. In this section, we introduce how AI algorithms deal with the problems of localization performance deterioration factors.

6.1. Identifying NLoS Events

Discriminating between line-of-sight (LoS) and NLoS paths is a crucial aspect of wireless localization. While measurements from LoS paths can be highly reliable, distance estimates from NLoS paths are inherently less reliable. Accordingly, in the past, attempts have been made to use radio signal parameters such as phase, skewness, and kurtosis to discriminate between the two environments; however, it was difficult to obtain high accuracy. Recently, discriminating LoS and NLoS through AI-based classification techniques has become an active research area. There are several methods, but for comparison, the following is an introduction to using different AI algorithms.

Firstly, an RNN method is proposed that utilizes channel state information (CSI) [30]. Although it is difficult to find the existence of LoS through CSI measurements, it can be inferred through patterns of variation in CSI over time. A new RNN model has been proposed that uses a sequence of CSI measurements to discriminate between LoS and NLoS channels, enabling the implementation of a localization system in challenging NLoS environments.

Secondly, a method using SVM is proposed [31]. In this method, LoS and NLoS are distinguished using an SVM-based signal classification. Then, the CNN is applied to the discriminated NLoS signal to recognize patterns and correct errors. Finally, a hybrid weighting algorithm is applied to the signal data obtained in this way, and the coordinates of the target point are obtained.

Thirdly, a CNN method is proposed that utilizes channel impulse response (CIR) [32]. In this method, features are extracted using a fully convolutional network that uses CIR data in UWB systems as an input dataset. Afterwards, the information necessary for classification is strengthened through the self-attention algorithm, and through this, the classification accuracy of LoS and NLoS is improved. This method may require the addition of other information to supplement the limited information of UWB CIR data.

Fourthly, there is a method using GNN [33], which is different from the previous methods that distinguish between NLoS and LoS. In this method, a GNN-based data-driven approach is used to achieve robust large-scale network localization and address mixed LoS/NLoS environments.

Finally, the features of CIR and channel parameters are combined using an ANN-based approach, named a channel impulse response-based neural network (CIRNN) [34]. In this method, LoS and NLoS can be classified more accurately by inputting the CIR and a small number of channel parameters into the CIRNN.

6.2. Resolving Spatiotemporal Asynchronization

The problem of spatiotemporal asynchronization can occur in various situations. When the internal clocks of the anchor and device are different, the device that receives the signal cannot know the transmission start time of the signal. Hence, it causes an unknown time difference between the ToA measurement and leads to localization errors. To solve this problem, there is a method that uses a Lagrange programming neural network (LPNN) [35], which is a type of DNN. In this method, the LPNN uses a set of ToA measurements as input training data. At this time, the input includes the uncertainty of the transmission start time due to clock asynchronization. The LPNN can estimate the position of the terminal, the starting transmission time, and clock asynchronization simultaneously from the input, and based on this, the refined position of the device can be extracted.

Meanwhile, when there is a spatiotemporal asynchronization caused by the relative position changes of the devices, appropriate features are extracted and probabilistic data are obtained based on them to correct the location. One approach is to use a CNN [36] to learn the spatial–temporal-frequency characteristics of the multiple-input and multiple-output (MIMO) channel from CSI data. This method collects CSI data in MIMO-orthogonal frequency division multiplexing (OFDM) Wi-Fi systems and constructs a 3D amplitude and phase matrix. It then uses this as input for the CNN and learns amplitude and phase features as the data pass through 3D convolutional and pooling layers. Based on newly collected MIMO data, the output layer generates probabilistic classification results for location estimation.

6.3. Preventing Signal Quality Deterioration

Signal quality deterioration issues caused by noise contained in wireless signals, atmospheric conditions, and ionospheric scintillation are resolved by excluding external factors other than the original wireless signal. Unlike the existing measurement model that estimates location using only the RSSI value of the received wireless signal, there is a method that uses all the RSSI, temperature, humidity, and link quality indicator (LQI) values as inputs for the ANN [37]. Additionally, this method can be simplified by adding the ID of the anchor as an input value.

On the other hand, there is a way to utilize external factors instead of eliminating them. An example of this is using the ionosphere layer based on LSTM [38] even in long-distance situations of over-the-horizon (OTH) to reduce the localization error. In fact, this method has a disadvantage in that it cannot be used when the ionosphere layer is unstable, but the significance lies in the fact that external factors that can cause errors are used in reverse.

6.4. Enhancing Performance Itself via Miscellaneous Methods

In addition to the methods that solve limitations caused by external factors, there are also active research efforts to improve the performance of localization technologies themselves. Many attempts are possible, but mainly, methods using RSSI with various AI algorithms are being studied. If there is a sufficient RSSI dataset that can be used as training data, it may be possible to estimate the location of the device with simple classification tasks.

First, there is a method using K-NN [39]. After obtaining RSSI fingerprinting data within the area to be located, the data are used as input to apply K-NN and complete the rough localization calculation based on the RSSI data received by the module to be located, the RSSI fingerprint database, and environmental noise parameters. This approach can improve accuracy by applying the Kalman filter.

Next, the RF algorithm can be an appropriate solution in the case of classification [40]. This method labels the RSSI for multiple anchors communicating with the device and repeats the process of randomly selecting some of them as independent trials to create a training dataset. They are used as input for the RF algorithm. It can be confirmed that the approach using ensemble techniques shows better performance compared to other algorithms using a single classification process.

While there are situations where the training dataset is sufficient as above, there are also cases where the training dataset is insufficient for using localization technology. In such cases, there are ways to supplement it. Among them, there is a method of estimating the location by generating virtual datasets based on GAN [41]. This method collects RSSI fingerprinting data from the existing training dataset and generates fake data based on real labeled collected data in order to boost localization performance. Accordingly, it can be confirmed that localization accuracy can be increased by generating fake data to supplement insufficient training datasets. In addition, there is a VAE [42] method of estimating location by utilizing video data instead of RSSI data. In other words, this approach uses image data obtained through a camera rather than a wireless signal. In this method, the VAE network is trained by database images. After this, the VAE network generates global features of the database images and queries images for subsequent image retrieval tasks. At the same time, local features of database images are applied to structure-from-motion (SFM) to reconstruct a 3D model. For the localization process, given a query image, this method generates the global features through the trained VAE network and subsequently performs a global retrieval to determine similar images in the database.

Specifically in range-based methods, the distance measurement can have the entire above-mentioned limitations depending on the specifications of the system. Therefore, if the measured distance is appropriately processed through correction, it can improve the performance of localization. The method of the noise-learning-based denoising autoencoder, nlDAE for short [43], is an algorithm that focuses on noise learning to restore the original for input with added noise, and is one type of denoising autoencoder (DAE). Unlike existing methods that focus on finding the original for the input, it focuses on noise learning rather than the distribution of true distance, making it relatively simple and more efficient with low uncertainty in many scenarios. In localization technology, it is used for error correction through noise removal in the measured distance that appears as a mixed value of error factors, including the actual distance and obstacles.

7. Discussion

As mentioned above, incorporating AI into localization technology using machine learning and deep learning algorithms can enable accurate measurements such as addressing external factors that degrade localization performance and enhancing the internal performance of the localization technology. However, as it is still at an early stage in the development of related fields; there are still challenges that need to be addressed. Additionally, this section discusses how cutting-edge technologies that are in the spotlight in 6G or further communication systems can affect AI-based wireless localization.

7.1. Open Challenges and Potential Solutions

7.1.1. Data Privacy Problem—Federated Learning

One of the most significant challenges is the privacy issue caused by accessibility. Deep learning requires a large amount of data to achieve adequate performance. However, in the process of securing a large amount of data, problems may arise that include sensitive personal information or information that should not be exposed to unidentified persons. Indoor localization systems collect real-time location information of users connected to the Internet through various means such as smartphones, cars, and IoT devices. If this personal location information is used without consent, it can cause many problems. Furthermore, if location information, such as a radio map of a country’s military facility and infrastructure, is misused, it can lead to national security issues. However, if accessibility to information is reduced to solve these problems, there is a problem that the performance of AI may decrease due to the insufficient amount of data necessary for learning. Therefore, a solution for sensitive information and privacy is needed while securing a large amount of data. In order to address privacy issues caused by accessibility, federated learning has been recently studied. Unlike the conventional method of gathering data in a central server for training, federated learning means leaving the data on each individual device and training on each device to obtain parameters, which are then aggregated and weighted to update the model as one [44,45]. In this case, private data are not transmitted to the server, but only model parameters are sent, which can help alleviate privacy issues.

7.1.2. Data Heterogeneity Problem—Distillation Methods

Considering the data stored on each device are personalized and have non-uniform and non-independent distributions, the learning model has difficulty achieving stable performance. This property is called heterogeneity, and the most common issue is distribution imbalance among nodes. Numerous research efforts are being conducted to address this problem. Another heterogeneity issue arises when different types of devices acquire data. Location information is collected by a wide variety of devices, and the nature of the data differs due to various differences in usage environments, communication methods, and so on. This also affects the application of the learned parameters, and a universal model must be adapted to the characteristics of each device. Dataset distillation [46] can be a solution to this problem, and it is expected to be a technique that can resolve the heterogeneity of wireless localization data during dataset collection.

7.1.3. Data Insufficiency Problem—Low-Rank Matrix Completion

Through the multi-dimensional scaling (MDS) method, it is possible to construct a Euclidean distance matrix (EDM) to obtain relative positions and then use eigenvalue decomposition to determine the absolute node positions. In this case, a sufficient number of known anchor positions are required to accurately determine the absolute positions of nodes. However, when the wireless communication range is restricted for various reasons, the acquisition of distance information will be limited, and the estimating node positions must rely on a small amount of data. To address this, a matrix completion method has been proposed to reconstruct the incomplete EDM by recovering some entries.

Since there are infinitely many completion options for generally unknown entries, it is impossible to recover the original matrix from information about a subset of entries. Nevertheless, if the matrix has a low-rank property, the rank of the EDM [47] in a k-dimensional Euclidean space is at most $k+2$. Therefore, it can be easily modeled as a low-rank matrix, allowing for the recovery of the partly empty matrix. By employing this approach, it is expected that node positions in scenarios with inadequate distance information can be accurately determined. Thanks to developments in AI methods, particularly NN-based denoising methods such as those in [43,48,49,50] promise to leverage the statistical inference as a novel and potentially more robust alternative to deal with noisy measurements for the EDM-based localization. However, the existing NN-based denoising techniques require computations proportional to the square of the number of entire sensors as the inputting and outputting vectors for NN frameworks. This results from the pairwise nature of generating and measuring the distance between two nodes, entailing a combinatorial approach. Hence, as the number of nodes increases, the complexity grows quadratically. The quadratic increase in the number of computations requires the handling of a substantial amount of training data and the employment of large-size NN models, which may give rise to interesting research topics.

7.2. Collaborations with 6G Key-Enabler Technologies

7.2.1. Integrated Sensing and Communication (ISAC)

With the rapid evolution of advanced communication technologies, including 6G, integrated sensing and communication (ISAC) with wireless localization systems has become a focal point for cutting-edge research. The high data rates, low latency, and massive device connectivity provided by advanced communication networks present new opportunities for enhancing the capabilities of integrated sensing and communication systems. As a result, ISAC systems play a crucial role in leveraging advanced communication technologies for enhanced localization capabilities. The integration of sensing and communication components allows for real-time data exchange and collaborative decision-making, particularly in dynamic and complex environments. If such systems are applied to comprehensive localization techniques, such as hybrid technologies, it is anticipated that faster and more precise localization will be achievable. In addition to these advancements, it is expected that further progress can be achieved through the integration of AI [51]. For example, a predictive model can be implemented by incorporating machine learning algorithms within the ISAC framework to predict and compensate for dynamic changes in a user’s location. Moreover, contextual information obtained through ISAC can enhance the understanding of a user’s environment and use it to enable more precise localization. These are particularly relevant in scenarios where traditional positioning systems may face challenges due to obstructions or signal interference.

7.2.2. Reconfigurable Intelligent Surface (RIS)

Reconfigurable intelligent surfaces (RIS) stand as an innovative technology in the realm of wireless communication, aiming to manipulate the characteristics of radio waves through the use of smart surfaces. These surfaces, comprising various antennas and reflective elements, enable functionalities such as signal concentration, directional control, and optimization of wireless links [52]. Primarily composed of antennas and reflective surfaces, RIS (v3) is software-controlled to adequately reflect or transmit radio waves. This capability regulates signal strength, minimizes multipath interference, and enhances overall communication performance. Especially in wireless localization technology, RIS can notably enhance accuracy in NLoS environments [53]. It improves signal paths in various settings, enabling precise location measurements. Enhancing accuracy in NLoS environments will result in possibilities for high accuracy of complex indoor and urban location-based services. Moreover, the flexibility and reconfigurability of RIS can optimize positioning performance across various application areas. Its application is expected to bring forth new possibilities in modern wireless communication and location-based services through improved performance and accuracy.

8. Conclusions

In this paper, we categorized wireless localization technologies into three distinct groups: range-based, range-free, and hybrid methods. Range-based technologies measure physical quantities of radio waves to estimate real-time position. In contrast, range-free technologies estimate position by utilizing existing information rather than analyzing the received radio waves, and hybrid technologies complement each other’s drawbacks to reduce localization errors.

Next, we provided three main external hindering factors that degrade the performance of these localization technologies. Firstly, NLoS occurs when communication between the anchor and device takes place through a different path due to reflection and diffraction, rather than a direct LoS path. Secondly, spatiotemporal asynchronization refers to performance degradation caused by changes in the coordination of space-time between the anchor and device due to differences in the internal system clocks of the anchor and device, or due to the relative rotations of the device. Finally, signal quality deterioration refers to degradation caused by various factors including additive noise, and so on.

To address these hindering factors, we classified machine learning algorithms that can be combined with wireless localization technologies into two distinct categories: machine learning algorithms without ANN (such as K-NN, SVM, and RF) and with ANN (such as CNN, RNN, LSTM, AE, GNN, GAN, and VAE). Lastly, we presented examples of cases where the three main external hindering factors mentioned earlier are addressed using these algorithms. Although AI-assisted wireless localization technologies have overcome various limitations, there are still challenges to be addressed, such as issues of user privacy, data heterogeneity and data insufficiency. It is expected that these issues will be appropriately addressed in future research related to more advanced techniques, such as federated learning, distillation-based models, and algorithm using the low-rank property of the matrix. Additionally, 6G or further spotlight technologies, such as ISAC and RIS, are anticipated to demonstrate outstanding positioning performance, especially when combined with AI.