The Challenge of Dynamic Environments in Regard to RSSI-Based Indoor Wi-Fi Positioning—A Systematic Review

Abstract

Among indoor positioning technologies, Wi-Fi fingerprinting using the Received Signal Strength Indicator (RSSI) is the most convenient and cost-effective method for indoor positioning. Instability and interference in wireless signal transmission cause significant variations in the RSSI, especially in a dynamic environment (DE). These factors hamper the accuracy of fingerprint-based indoor positioning system (IPSs), as these systems may struggle to reliably match observed signal patterns with stored fingerprints. Thus, ensuring positioning accuracy is critically important when designing and implementing Wi-Fi IPSs. Currently, there is a lack of surveys that provide a detailed and systematic analysis of the impact of DEs on the accuracy and reliability of Wi-Fi indoor positioning. This systematic literature review (SLR) was conducted to examine three aspects of Wi-Fi indoor positioning based on the RSSI: the impact of a DE on indoor positioning accuracy, the importance of constructing radio maps for indoor localization, and the role of machine learning (ML)/deep learning (DL) models in predicting indoor position with minimal error despite the DE. This review was conducted according to a structured and well-defined methodology to search for and filter relevant studies on Wi-Fi indoor positioning using the RSSI. Through this systematic process, 128 papers (2018–2024) were identified as relevant and then extracted and thoroughly analyzed to effectively answer the specified research questions. Additionally, this review highlights gaps in existing research, suggests directions for future studies, and provides practical recommendations for enhancing Wi-Fi-based indoor positioning in DEs.

1. Introduction

With the widespread use of wireless networks, location-based services have been integrated into our daily lives. Currently, there are two main types of navigation systems: indoor (inside a building) and outdoor (outside a building in the open air). Global Positioning System (GPS) is well known as one of the outdoor navigation methods that can accurately provide the position of an object from satellites. However, this method only works effectively in open areas with a direct line of sight with respect to the satellites [1]. GPS signals may be interfered with or blocked by building structures, such as walls and roofs [2]. Therefore, indoor location cannot be determined when the signal is lost inside a building. GPS is not applicable when indoor positioning is required.

Although GPS is not applicable in these cases, the advancement of technology has made indoor positioning feasible. Such technologies have been proposed to implement indoor positioning, i.e., Radio Frequency Identification (RFID), Wi-Fi, Bluetooth, and ultra-wide band (UWB) [1,3,4]. Along with positioning, these technologies also enable indoor navigation. This has allowed the development of indoor maps, which ease human life in airports, parking garages, alleys, underground locations, or inside multistory buildings. With indoor positioning technology, passengers can receive a customized route with guided directions to the desired destination by using a mobile phone. This reduces unnecessary travel time and improves the passenger experience. Among all technologies for indoor positioning, Wi-Fi is preferred because it does not require additional modifications to the existing infrastructure, has minimal hardware requirements, and is commonly available in public spaces [3,5,6].

In Wi-Fi indoor positioning, the RSSI has become a prominent tool for various reasons. Due to the widespread availability of Wi-Fi in indoor environments and wireless access points (WAPs) that can function as transmitting devices [7], this method has gained increasing popularity and significance in recent years for indoor positioning. The cost of deploying it on a larger scale can be reduced by utilizing existing WAPs [7,8]. In addition, Wi-Fi has a wider coverage area than other methods [9,10]. This expanded reach enables Wi-Fi IPSs based on the RSSI to provide comprehensive coverage across large indoor spaces, including multistory buildings and complex facilities.

The RSSI fingerprinting method usually involves two stages: an offline stage (training stage) and an online stage (positioning stage) [11,12]. In the offline stage, a device such as a smartphone, laptop, or Wi-Fi-enabled device captures RSSI fingerprint data at the target area to build the training data sets (i.e., a radio map). These recorded RSSI values are associated with the corresponding coordinates of each location. In the online stage, the mobile device measures the current RSSI values from surrounding WAPs, and these RSSI measurements serve as input data for the localization system. Then, the current RSSI measurements are compared to the pre-existing fingerprints in the database. Based on the best match, the system estimates the current location of the device, and the result will be sent back to the user. However, the RSSI signal from the wireless access point is vulnerable to various factors such as obstacles, signal fluctuations, noise, environmental changes, non-line-of-sight communication, and multipath interference [13,14,15,16,17]. In short, these factors are part of the DE.

Therefore, considering DEs is essential, especially in real-world implementations. Neglecting DEs in the design and evaluation of indoor positioning systems can lead to significant performance degradation, including reduced localization accuracy, increased latency, and system instability. Since real-world applications require reliable and adaptive positioning solutions, incorporating DEs into discussions ensures that positioning algorithms and models are resilient to environmental changes. With the popularity of RSSI signals, review papers on the RSSI can easily be found online. However, knowledge gaps remain. Detailed explanations are provided in Section 1.1, with the motivation discussed in Section 1.2. In this paper, we attempt to address these gaps by providing a deeper analysis of how a DE could influence Wi-Fi IPSs.

1.1. Existing Survey Articles

Several review papers on Wi-Fi indoor positioning were organized and summarized, as shown in

Table 1. Summary of existing surveys on indoor positioning. “Yes” indicates that the paper addresses or includes information about this aspect. “No” indicates that the paper does not address or include information about this aspect. “Not specified” indicates that the paper does not clearly mention or discuss this aspect.

Paper	Year	Wi-Fi RSSI	DE Challenges	RM Construction Method	ML and DL
[25]	2018	Not specified	Not specified	No	Not specified
[26]	2019	No	No	No	Yes
[3]	2019	Yes	Yes	No	No
[18]	2020	Yes	Not specified	No	Yes
[27]	2020	No	Not specified	No	Yes
[19]	2020	Yes	Not specified	No	Not specified
[20]	2020	Yes	Not specified	No	Yes
[21]	2021	Yes	Not specified	Yes	Yes
[22]	2022	Yes	Not specified	No	Yes
[23]	2023	Yes	Yes	Yes	No
[24]	2023	Yes	Not specified	No	Yes
[4]	2024	Yes	Not specified	Yes	Yes
[13]	2024	Yes	Not specified	Not specified	No
[28]	2024	Yes	Not specified	No	Yes
Our SLR	2024	Yes	Yes	Yes	Yes

. The table indicates that the topic of a Wi-Fi RSSI for indoor positioning is common in most of the reviewed studies. Specifically, papers such as [3,4,13,18,19,20,21,22,23,24] include discussions on the Wi-Fi RSSI in their studies.

However, earlier studies, such as [25,26,27], either do not specify or do not include discussions on the Wi-Fi RSSI, indicating a potential gap in their methodologies. Regarding the challenges associated with DEs, only [3,23] mention this aspect, while most studies do not discuss DE challenges or do not provide clear information. This is a significant gap in the literature, as DEs can impact indoor positioning accuracy. The lack of attention to this aspect in most studies presents an opportunity for future research to better explain and analyze this topic. Besides that, many papers do not include discussions of radio map construction methods, with only three papers discussing this aspect [4,21,23]. In contrast, the application of ML and DL techniques is covered in most studies, with only five papers either not mentioning them or briefly mentioning them without providing detailed insights into their implementation [3,13,19,23,25]. To the best of our knowledge, there are no systematic literature reviews (SLRs) that focus on DEs and how ML/DL impact DEs in indoor positioning. Additionally, the role of radio map construction in improving positioning accuracy and time efficiency in DEs is rarely mentioned. A list of review or survey papers published between 2018 and 2024 and the discussed features are summarized in Table 1, which shows that existing reviews explore these areas separately but rarely address DEs specifically.

1.2. Motivation and Contributions

Based on the existing knowledge gap (refer to Section 1), the primary motivation of this work is to shed light on this intricate domain and provide readers with essential references about Wi-Fi indoor positioning based on the RSSI regarding DEs. The contributions of this SLR are summarized as follows:

• It shows the reasons behind the impact of DEs on Wi-Fi indoor positioning accuracy.
• It analyzes how the construction of a radio map can improve positioning accuracy and training time efficiency in DEs.
• It highlights the role of ML and DL in adapting indoor positioning to different environmental conditions.
• It identifies areas for further exploration and proposes innovative ideas to advance this field.

The main content and contributions of this SLR are organized as follows. In Section 2, the research methodology of this study is explained. Section 3 discusses the fundamental concept of the RSSI and evaluates the advantages and challenges of RSSI-based techniques through insights drawn from various research studies. Section 4 concludes and discusses the research outcomes obtained from relevant studies that address the defined research questions. Section 5 identifies the limitations mentioned in existing works and clarifies areas for further exploration. Finally, Section 6 summarizes the entire research study.

2. Research Methodology

In this study, an SLR was conducted in accordance with the PRISMA 2020 guidelines. We applied a systematic method to identify, evaluate, and select research publications in an organized manner. The steps, which include the definition of the eligibility criteria, are illustrated in Figure 1. The review protocol was not registered.

2.1. Research Questions

The following research questions were formulated to align with the primary objective:

• RQ1: How does a DE affect the accuracy of indoor positioning?
• RQ2: How can constructing a radio map improve positioning accuracy and training time efficiency in DEs?
• RQ3: How can ML/DL models predict indoor position with minimal error despite the challenges posed by a DE?

RQ1 pertains to the influence of a DE on the accuracy of indoor positioning. Indoor positioning plays a vital role in various areas, ranging from industrial automation, asset tracking, and logistics to healthcare, retail, and smart homes [4,29]. However, the effectiveness of these systems can be significantly influenced by the dynamic nature of indoor environments. Therefore, this investigation identifies dynamic factors and their impact on the precision and reliability of indoor positioning technologies.

With RQ2, we aim to specify the impact of constructing a radio map on enhancing the accuracy of localization systems and reducing training time. In the context of indoor positioning, radio maps serve as a fundamental basis for localization algorithms, providing critical information for accurate positioning of devices. This study seeks to examine the methodologies and techniques employed in creating radio maps and analyze how the advantages of these construction methods contribute to improving accuracy in localization and training time efficiency, even in DEs.

RQ3 was specifically formulated to explore the capabilities of ML or DL models in predicting indoor positions despite the challenges posed by a DE. By investigating this question, we seek to assess the effectiveness of ML/DL techniques in adapting to continuously changing environments, such as areas featuring variations in signal strength, dynamic object movement, and evolving interference patterns within indoor spaces. Overall, this investigation is important for identifying the potential of ML/DL approaches in addressing the challenges pertaining to IPSs and predicting indoor position with minimal error in practical scenarios.

2.2. Data Search Strategy

The data search strategy plays a crucial role in conducting an SLR. The steps involve identifying the primary sources for the research and defining the relevant keywords. To ensure comprehensive coverage and access to high-quality scholarly articles, we selected the following digital libraries: Google Scholar, IEEE Xplore, ScienceDirect, ACM Digital Library, MDPI Journals, and Springer.

To enhance the precision and relevance of the search results, the search terms (keywords) were determined based on the following steps:

• Identifying primary terms aligned with the respective research questions;
• Exploring alternative spellings and synonyms for the designated main terms;
• Confirming the validity of the search terms through reference to pertinent studies;
• Using Boolean operators (OR/AND) to systematically combine these terms for a comprehensive search strategy.

Table 2. Keywords used in the search.

“WiFi indoor positioning” OR “WiFi fingerprinting” OR “Wi-Fi indoor positioning” OR “Wi-Fi fingerprinting” OR “Machine learning” AND “WiFi indoor positioning” OR “Machine learning” AND “Wi-Fi indoor positioning” OR “RSSI” AND “WiFi indoor positioning” OR “RSSI” AND “Wi-Fi indoor positioning”

presents the search terms (keywords) used in this research. The specified search terms are applied within the titles and abstracts in the identified digital libraries.

2.3. Paper Selection Criteria

Initially, an investigation was conducted on 607 papers to assess their relevance to the research question. The next step involved the removal of duplicate articles collected from different digital libraries. Next, these articles were examined and selected based on the following criteria:

• Consideration was given to the sources of publication, such as journals and conferences, excluding theses, white papers, and dissertations.
• Articles were reviewed to determine whether they constituted original research, excluding reviews or summaries.
• Inclusion was restricted to articles published within the timeframe from 2018 to 2024.
• The articles needed to be written in English.
• The articles had to apply to the specified research questions

2.4. Paper Selection Process

In this step, papers were selected based on the search string, title, abstract, and keywords. After the screening process, 196 papers that contained relevant information in their titles and abstracts were identified. Next, the abstracts of these papers were read to filter out irrelevant articles, as although some may contain the keywords, their contents are not related to indoor positioning. Finally, 128 selected papers proceeded to the extraction process to evaluate their relevance and contribution to the research objectives of this study.

2.5. Data Extraction

In this step, a critical analysis of the selected papers was conducted by examining and evaluating the complete content of each paper. The primary outcomes extracted from each study were matched with the research questions. When studies presented multiple scenarios, the most relevant data corresponding to these outcomes were collected. Then, the process involved organizing relevant information into a table format to address the specified research questions. The following information was extracted from each paper: the title of the paper, publication year, RQ1, RQ2, RQ3, limitations, and notes. Potential sources of bias were identified and noted, including differences in sample size, restricted testing environments, and limited performance metrics. Any missing or unclear information was marked as “-” and verified based on the original study. The primary effect measure used in this paper was mean positioning error (in meters). Additional measures, including RMSE, MAE, median error, and percentile error, were also included, enabling cross-study comparisons. No automation tools were employed in this process.

2.6. Data Synthesis

Finally, the data collected from the selected papers had to be synthesized systematically to address the research questions effectively. Studies reporting similar types of information, such as dynamic environments or performance indicators, were grouped. Reported performance metrics were used as presented in the original studies. In cases where different but related measures of accuracy were reported (e.g., RMSE, MAE, and mean error), they were treated as comparable indicators of positioning error to enable meaningful comparison. Studies with insufficient or missing data were summarized narratively rather than excluded. Additionally, the results of the studies were tabulated to support comparison across key variables, including dataset, method, cost, complexity, scalability, and experimental environment. Visual representations such as tables and comparative summaries were used to highlight similarities and differences across studies. A narrative synthesis approach was applied to provide a descriptive and organized summary of findings. No subgroup analyses, meta-regression, or sensitivity analyses were conducted, as the primary objective was to compare and synthesize findings descriptively rather than statistically.

3. RSSI

In this section, the fundamental concept of the RSSI will be described in terms of its significance and role within IPSs. Moreover, this section will draw insights from various research papers to evaluate the advantages and challenges of employing RSSI-based techniques for indoor positioning.

RSSI is a measure of the power level of a radio infrastructure’s Received Signal Strength (RSS) that can estimate the distance between transmitters and receivers [30]. Figure 2 illustrates the RSSI positioning method. In this method, signal strength is obtained from the target device in decibel-milliwatts (dBm) [31]. Since the RSSI is measured in dBm, RSSI values that are negative and closer to 0 dBm represent stronger signal quality, indicating closer proximity to the corresponding AP. For signal strengths not detected by the device, the RSSI is commonly set to +100 dBm or 100 dBm by default [32,33]. In a wireless environment, the RSSI is the relative RSS; signal quality depends on the value of the RSSI [34].

Table 3. RSSI signal strength.

Signal Strength	Grade
>−50 dBm	Excellent
−50 dBm to −60 dBm	Good
−60 dBm to −70 dBm	Fair
<−70 dBm	Weak

presents the levels of different signal strengths.

In Wi-Fi-based IPSs, the RSSI is widely used and considered a practical method for IPSs compared to other approaches. However, in fact, the RSSI often struggles to meet the accuracy requirements due to the inherent limitations of RSSI measurements. To better understand the landscape of RSSI-related research, a comprehensive review of RSSI discussions from various papers is presented in

Table 4. Comparative analysis of RSSI studies.

Study Reference	Year	Advantages	Limitations
[35]	2018	Availability for indoor positioning Low implementation costs	Device-specific fingerprinting Challenges in fast-changing environments
[36]	2018	Widely available	Readings of the RSSI often fluctuate Large storage space is required to store reference RSSI data
[37]	2018	Easy to obtain from wireless devices Low cost	Sensitivity to environmental changes Limited range Device heterogeneity
[14]	2018	Does not require any additional hardware or infrastructure Cost-effective Wide coverage	Sensitivity to environmental changes Variability in RSSI values
[38]	2019	No need for additional hardware	Not specified
[17]	2020	Wide availability Low cost	Signal interference Sensitivity to environmental factors Limited range Dependency on Wi-Fi infrastructure
[39]	2020	Widely available	Sensitivity to environmental changes Requires labor-intensive fingerprint database construction
[16]	2021	Low cost and easy implementation Widespread availability	Sensitivity to environmental factors Limited range Lack of environmental adaptability
[40]	2022	Requires less spending on physical equipment. Easy implementation and control. Good positioning effect	Signal amplitudes will interfere with each other due to the rapid increase in the number of available access points
[41]	2022	Does not require any additional hardware or infrastructure Wide availability Supports multiple users simultaneously	Signal interference Signal propagation variability Limited accuracy in a dynamic environment One needs to spend a significant amount of time creating a reference database for calibration
[15]	2022	Simplicity Low hardware requirements	Random fluctuations caused by fading and multi-path phenomena
[42]	2024	Low cost Multipath and Non-Line-Of-Sight	Complexity variability of the indoor environment

.

The analysis indicates that researchers prefer using the RSSI for Wi-Fi indoor positioning primarily because of its widespread availability, the lack of need for additional hardware or infrastructure, and its cost efficiency [43]. Despite being widely used and offering many advantages, the RSSI still faces significant challenges related to DEs and signal stability. First, the accuracy of the RSSI can be affected, especially in situations where there are obstructions to the line of sight, resulting in decreased precision caused by signal attenuation and multi-path fading [43,44]. Additionally, device heterogeneity can also cause inconsistencies in RSSI measurements due to variations in hardware sensitivity across different devices. Due to this situation, substantial efforts may be required to collect and maintain accurate signal strength data, making the construction of the fingerprint database more labor-intensive and time-consuming. Achieving high position accuracy using RSSI values remains a challenge because of these dynamic factors.

4. Results

4.1. RQ1: How Does a DE Affect the Accuracy of Indoor Positioning?

Today, Wi-Fi has emerged as the most widely used technology for indoor positioning because of the extensive deployment of APs, particularly in urban areas. However, Wi-Fi IPSs still face many challenges, and one of the primary issues is accuracy. In indoor positioning using Wi-Fi, signal stability can directly affect the positioning estimation and performance of the trained model. This problem is usually caused by DEs. Therefore, RQ1 was formulated to identify and characterize a DE and examine its effects on the accuracy of Wi-Fi IPSs. Specifically, the focus is on four key aspects: (1) the complexity of the environment and environmental changes over time; (2) the movement and presence of objects; (3) device heterogeneity; and (4) the number and distance of access points. Table 5 summarizes these dynamic factors by listing the number of papers associated with each category.

In fact, complex environments with different layouts and structures, including walls, furniture, and doors, are some of the dynamic factors that cause the inconsistencies in Wi-Fi signal strength [14,16,45,46,47,48,49,50,51,52,53,54,55,56,57]. Besides that, the indoor environment may undergo temporal changes. These dynamics also involve alterations in furniture placement, introducing new obstacles or modifications in the spatial layout. These environments or changes introduce a complexity that can lead to signal attenuation, reflection, or multipath effects, causing fluctuations in the RSSI [7,11,16,36,37,45,46,47,48,50,52,53,58,59,60,61,62,63,64,65,66,67,68,69]. Therefore, effectively navigating through temporal changes and reducing the impact of signal interference pose a critical challenge.

Apart from that, one of the key dynamic factors arises from the presence of moving objects within indoor spaces. As people or objects move, they can intermittently obstruct or reflect these signals, causing variations in RSSI values [14,48,52,53,54,55,70,71,72,73,74,75]. The unpredictable nature of these movements makes it challenging to maintain consistent and accurate RSSI measurements [41]. Consequently, the dynamic nature of moving objects within indoor spaces introduces a level of uncertainty that directly influences the overall accuracy of IPSs.

Furthermore, device heterogeneity is also regarded as a significant dynamic factor affecting Wi-Fi IPSs. This heterogeneity includes differences in hardware components (e.g., sensors, antennas, and signal-processing chips) or software configurations (e.g., operating systems and firmware versions). These variations can lead to inconsistencies in signal measurements, such as differences in RSSI values, despite the devices being in the same position [76,77,78,79,80,81,82,83,84]. This variability poses challenges for maintaining accurate and reliable positioning results, as the system must adapt to the unique attributes of each device type to ensure consistent performance across a wide range of devices.

In addition, the effectiveness of indoor positioning often correlates with the distance between or number of access points [47,55,61,62,64,66,71,85,86,87]. For example, a higher density of access points typically will improve accuracy [40]. The signals from nearby access points can overlap or interfere with each other, leading to difficulties in accurately determining the position of a target [40]. Although conventional wisdom suggests that an increased number of access points should enhance precision [49,51,53,88], the distance between access points is also important. As the number of access points increases, the signals emitted by each access point can intersect and interfere with one another. This interference may affect the reliability of Wi-Fi fingerprinting methods and degrade prediction performance. Here, achieving the optimal balance between access point density and spacing is important for maximizing the accuracy of indoor positioning and reducing the potential dynamic factors. The taxonomy of the dynamic factors is summarized in Figure 3.

As mentioned above, these dynamic factors can gradually degrade the accuracy and reliability of RSSI-based IPSs. These factors change the signal propagation paths and received signal strength, leading to mismatches between stored fingerprints and real measurements. Moreover, such inconsistencies reduce the validity of the existing radio map, making periodic recalibration necessary to maintain positioning accuracy. To provide a clearer understanding of the factors that require radio map updates, in Table 6, we summarize the key environmental and system variations influencing the need for recalibration.

Table 5. Dynamic factors in RSSI-based IPSs.

No of Paper	Dynamic Factors
[7,11,14,16,36,37,45,46,47,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63,64,65,66,67,68,69]	Complex environment with different layouts and structures Environmental changes
[14,41,48,52,53,54,55,70,71,72,73,74,75]	Movement and presence of objects
[77,78,79,80,81,82,83,84,85]	Device heterogeneity
[43,48,50,52,54,56,62,63,65,67,72,86,87,88,89]	Number of or distance between access points

Table 5. Dynamic factors in RSSI-based IPSs.

No of Paper	Dynamic Factors
[7,11,14,16,36,37,45,46,47,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63,64,65,66,67,68,69]	Complex environment with different layouts and structures Environmental changes
[14,41,48,52,53,54,55,70,71,72,73,74,75]	Movement and presence of objects
[77,78,79,80,81,82,83,84,85]	Device heterogeneity
[43,48,50,52,54,56,62,63,65,67,72,86,87,88,89]	Number of or distance between access points

Table 6. Summary of key dynamic factors affecting radio map stability and the necessity of updates.

Factor	Example of Changes	Impact of Signal	Need for Radio Map Update
Environment layout	Empty or full; open or closed doors	Alters signal propagation paths and attenuation.	Moderate High
Structural modification	Addition/removal of walls or partitions	Causes major RSSI pattern changes.	Very High
Human movement	Walking individuals or dense crowds	Introduces temporal signal fluctuation.	Moderate High
Object movement	Furniture or equipment relocation	Affects reflection and scattering.	High
Device heterogeneity	Using different Wi-Fi chipsets or antenna orientations	Produces inconsistent RSSI readings.	Moderate
Access point relocation/configuration	Moving APs and adjusting transmission power and channel or antenna direction	Shifts signal coverage and strength.	Very High

Table 6. Summary of key dynamic factors affecting radio map stability and the necessity of updates.

Factor	Example of Changes	Impact of Signal	Need for Radio Map Update
Environment layout	Empty or full; open or closed doors	Alters signal propagation paths and attenuation.	Moderate High
Structural modification	Addition/removal of walls or partitions	Causes major RSSI pattern changes.	Very High
Human movement	Walking individuals or dense crowds	Introduces temporal signal fluctuation.	Moderate High
Object movement	Furniture or equipment relocation	Affects reflection and scattering.	High
Device heterogeneity	Using different Wi-Fi chipsets or antenna orientations	Produces inconsistent RSSI readings.	Moderate
Access point relocation/configuration	Moving APs and adjusting transmission power and channel or antenna direction	Shifts signal coverage and strength.	Very High

4.2. RQ2: How Can Constructing a Radio Map Improve Positioning Accuracy and Training Time Efficiency in DEs?

Creating a radio map is a crucial stage in deploying the fingerprinting indoor positioning technique [89,90]. A well-constructed radio map, including detailed signal strength measurements at various locations, can help improve overall system performance during indoor positioning. Besides that, the use of advanced ML algorithms (e.g., K-Nearest Neighbors (KNNs), Neural Networks, or other methods) trained on a quality radio map can increase a model’s reliability to ensure more precise position estimation. Hence, it is evident that the accuracy of Wi-Fi indoor positioning is closely related to the quality of the constructed radio map. However, the creation or maintenance of a radio map requires considerable time and effort through manual calibration [90].

Indoor personnel movement or indoor environment may change over time, which will affect positioning accuracy. Therefore, it is necessary to update the radio map regularly based on environmental changes. Most of the time, a radio map may require several days or weeks to build or maintain, so its construction is time-consuming and labor-intensive. These substantial time and labor requirements have become a major barrier to the development of fingerprint indoor positioning.

Accordingly, finding ways to quickly and cost-effectively construct radio maps has become an important topic in this field in recent years. Researchers have introduced various methods based on interpolation, crowdsourcing, semi-supervised learning, simulators, unsupervised learning, and inertial sensors. The primary purpose of RQ2 was to discuss the methodologies and techniques different researchers have applied in creating radio maps and analyze how these construction methods contribute to improving accuracy in localization and training-time efficiency in DEs. A detailed discussion of each technique is presented in the following sections.

4.2.1. Crowdsourcing

Using crowdsourcing for indoor map construction has attracted growing interest in recent years. The basic concept of crowdsourcing involves collecting contributions from a large group of users and other participants, including both professional surveyors and common users, to gather data within indoor environments [91]. Currently, smartphones are widespread and contain many sensors, such as IMUs, cameras, microphones, Bluetooth, and Wi-Fi scanners [92]. When a user walks indoors, these sensors in the phone can be used passively or actively to record map-related information in order to extract the indoor map structure from the sensor data [91,93].

In passive mode, sensors record information that will be gathered in the background without direct user input, interaction, or awareness [94]. In contrast, active mode typically requires user intervention or input to collect data [91]. This may involve starting sensor measurements through an application or selecting certain features or functionalities that depend on sensor data.

Regardless of whether passive or active modes are employed, the main purposes of crowdsourcing are reducing labor costs, time, and storage during the construction of the radio map [77,95,96,97,98]. However, there remain some potential challenges associated with using crowdsourcing, including issues related to device heterogeneity [77,97], variations in sample dimensionality and spatial distribution [97,99], and inaccurate sample annotation and location information [97,100].

The typical positioning accuracy of crowdsourced radio map construction ranges from approximately 0.6 m to 3 m [95,97,98]. However, the actual accuracy depends heavily on the density of the crowdsourced data, the algorithm used to clean or interpolate the radio map, and the stability of the wireless infrastructure.

Even though crowdsourcing methods have accelerated the collection of RSS samples throughout a building with minimal effort and time, these samples may not effectively represent their distributions, affecting positioning performance [91]. Therefore, while crowdsourcing offers relative simplicity in collecting data, verifying the reliability of the data remains a challenging issue [6].

4.2.2. Interpolation

Interpolation is a mathematical tool that can estimate a value at a specified point based on the spatial relationship among nearby points. The authors of [89,101,102] suggest that interpolation methods can improve the accuracy of indoor positioning or reduce the time required for data collection while also maintaining accuracy. Although some systems utilize interpolation methods, different approaches may involve the application of interpolation in constructing the radio map. Examples include the Biharmonic spline interpolation method [89], Inverse Distance Weighting [101], KNN interpolation [101], Linear interpolation [102], and Gaussian process regression (GPR) [102].

These interpolation methods can achieve accurate positioning with a sufficiently small sampling interval and uniform sampling density. When the sampling interval is large and there are numerous APs, performance deteriorates. Moreover, this approach may neglect factors such as wall attenuation and device heterogeneity, which pose challenges in real-world environments [91,103]. Interpolation methods have difficulty handling dynamic factors.

4.2.3. Semi-Supervised and Unsupervised Learning Methods

Semi-supervised and unsupervised learning methods are also applied in radio map construction. In the semi-supervised learning method, a prediction model can be trained using a limited labeled dataset with a large set of unlabeled data [6,104]. This method utilizes unlabeled data effectively during the construction of the radio map to reduce the need a great deal of labeled data and annotation effort [6,105,106]. However, it is undeniable that researchers still need to spend some time collecting accurate data for initializing the learning process [91].

The unsupervised learning method rarely depends on labeled data or any manual calibration to train a prediction model [107]. By using this method, a radio map can still be constructed for specific indoor environments that contain only a few reference locations [90,108]. Obtaining a significant number of unlabeled fingerprints has become easier through crowdsourcing, especially now. Hence, the unsupervised learning method emerges as a promising solution [91]. In fact, there are several limitations to unsupervised learning methods, including their inability to handle dynamic indoor environments and challenges in accurately modeling path loss. Here, unsupervised learning methods usually need to be combined with other techniques to improve overall reliability [6].

4.2.4. Inertial Sensors and SLAM

Currently, smartphones typically incorporate several inertial sensors, including accelerometers, magnetometers, and gyroscopes, which can also be called Inertial Measurement Units (IMUs) [109,110,111]. These sensors provide crucial data on linear acceleration, angular velocity, and magnetic field strength. In addition, this data is fundamental for techniques like Pedestrian Dead Reckoning (PDR), which uses an accelerometer to count steps and combines data from a magnetic field and gyroscope to estimate orientation angles [91,112]. Thus, this information offers reliable solutions for using indoor positioning to enhance the accuracy of a radio map during construction and makes real-time location tracking possible.

Similarly, SLAM is a technology used to create a radio map of an unknown environment while simultaneously tracking a device’s location in real time [109]. This approach utilizes data from various sensors, including cameras, lidars, and IMUs, to perceive the environment and build a radio map. The positions obtained through IMUs or SLAM often suffer from significant inaccuracies due to the accumulation of errors from noisy sensors and feature mismatching [5,91,107,111]. Additionally, new challenges continue to arise, including sensor drift, orientation estimation, feature extraction, device heterogeneity, and power consumption [91,112,113]. These limitations may affect reliability during radio map construction and lead to an overall performance decline of IPSs in DEs.

Figure 4 and

Table 7. Summary of methods for radio map construction.

Method	ID	Positioning Method/Testbed Area	Dataset/Radio Map Source	Public Availability	Advantages	Disadvantages	Accuracy
Crowdsourcing	[98]	Weighted K-nearest neighbor (WKNN) 4th floor of Building 695 WLAN APs	Self-collected dataset	No	Avoids the need to repeatedly conduct site surveys using expert surveyors, which is power-consuming, time-consuming, and labor-intensive Ensures accuracy while updating the indoor radio map using crowdsourcing	Dependency on crowdsourcing participation	0.6325 m (Mean/Best)
	[97]	Nearest neighbor (NN) A typical office environment The total area is 592 m2	Self-collected dataset	No	Labor cost and time can be reduced	Crowdsourced samples may be inaccurate due to the dynamic environment Device heterogeneity Crowdsourced samples may not be uniformly distributed	Average localization error of around 1.5 m
	[95]	WKNN One floor of a typical office building 3600 m2 194 APs	Self-collected dataset	No	Solves the high-sampling-cost problem Guarantees localization accuracy	Difficult to ensure the data is collected consistently and accurately Crowdsourced data collection is prone to including duplicates	The 60th percentile (60%ile) localization error for RACC is 3 m.
Interpolation	[91]	WKNN 211 m × 2.4 m 56 TP-Link 2.4 GHz APs	Self-collected dataset	No	Reduces the workload and time pertaining to radio map construction Ensures the same positioning accuracy as a complete manual radio map	Device heterogeneity The proposed method may not reflect the distance between the AP and device due to dynamic environments	4.1 m (ME)
	[101]	KNN and inverse distance weight (IDW) A floor area of 150 m × 150 m 3 Wi-Fi transmitters	Simulated dataset	No	Improves indoor positioning accuracy Reduces the offline workload while updating the radio map	Interpolation errors	Interpolation error reduces by 54% when a correlation exists in the collected Wi-Fi measurements
	[102]	KNN An office 16.71 × 10.76 m	Self-collected dataset	Yes	Computational requirements can be significantly lowered without affecting the positioning error	Interpolation does not completely remove large positioning errors near the edges of the office	2.29 (MAE)
Semi-supervised	[104]	ANN 4-floor university building	Wi-Fi crowdsourced fingerprinting dataset	Yes	Solve the problem of overfitting due to insufficient labeled data Reduces costs	-	Average floor detection accuracy is up to 93.94%
	[6]	Variation auto-encoder (VAE) Surface measuring 108,703 m2 Three buildings with 4 or 5 floors in a university	UJIIndoorLoc dataset	Yes	Reduces the need for labeled data	A dynamic environment can impact signal strength and positioning accuracy	4.65 m (RMSE)
Unsupervised	[107]	Weighted least squares (WLS) Ninth floor of an office building 1100 m2 (41 m by 27 m) A total of 35 fixed access points	Self-collected dataset	No	Optimizes model-based radio maps Radio maps can automatically cope with changes in the environment	Influence of dynamic environment	The median accuracy with the WHIPP path loss model improved from 2.90 m to 2.07 m
	[108]	Support vector regression (SVR) Surface measuring 108,703 m2 Three buildings with 4 or 5 floors in a university	UJIIndoorLoc dataset	Yes	Does not require location annotation and is less sensitive to user privacy	-	1.9 m (RMSE)
Inertial Sensors	[112]	PDR, Wi-Fi Constraints and Landmark Constraints Mall	Self-collected dataset	No	There is no need for extra hardware; this approach only relies on a smartphone Simplifies the site survey process	Device heterogeneity	Mean error of 1.10 m
	[111]	Backpropagation (BP) neural network Third and fourth floors of an office building 128.4 m × 55.5 m	Self-collected dataset	No	Low cost and easy implementation Adequate accuracy	Wi-Fi signal fluctuations Sensor deviations	1.12 m (RMSE)
Simultaneous Localization and Mapping (SLAM)	[109]	Tango-based PDR and step-counter-based PDR One building on campus 130 m × 70 m	Self-collected dataset	No	Cost-effective alternative for estimating the trajectory of multiple users in unknown environments. Reduces labor-intensive phase in collecting measurements in the existing infrastructure Enhances accuracy by incorporating PDR	Incorrect loop closures due to data association errors	0.6 m with Tango-based PDR and 4.76 m of a step counter-based PDR (RMSE)
	[5]	Extended Viterbi Tracking Signal Fluctuation Matrix (EVSFM) A medium-scale office building 89 m2 118 APs	Self-collected dataset	No	Increases the reliability of training-data collection Improves the accuracy of Wi-Fi radio map construction	The construction radio map process has to be repeated in a new building or space. Open-space limitations	Positioning accuracy close to 1 m can be obtained

summarizes various methods for radio map construction. In the table, it can be observed that several methods have been proposed to minimize the effort required for creating radio maps and enhance accuracy during indoor positioning. These include techniques based on crowdsourcing, interpolation, semi-supervised or unsupervised learning, inertial sensors, and SLAM. It is worth mentioning that SLAM-based, inertial sensor-based, semi-supervised learning, unsupervised learning, and interpolation methods are typically combined with the crowdsourcing method, with no clear distinction among them.

4.3. RQ3: How Can ML/DL Models Predict Indoor Position with Minimal Error Despite the Challenges of a DE?

As previously mentioned, Wi-Fi IPSs have attracted a significant amount of attention because of their simplicity and low hardware requirements [35]. However, ensuring accuracy, adaptability, and scalability still represents a severe challenge for widespread deployment, especially in DEs.

With the rapid development of Artificial Intelligence (AI) in recent years, various ML and DL techniques are increasingly being explored by researchers to address these challenges, with reasonable success (e.g., supervised learning, unsupervised learning, or deep learning methods) [2,17,40,41,46,58,60,114,115,116]. In fact, achieving high accuracy in a complex environment requires models that can not only process large amounts of data but also adjust to environmental changes in real time.

ML and DL techniques have the potential to provide more reliable solutions by learning patterns from data and compensating for uncertainties. Hence, RQ3 aims to identify a method that delivers accurate indoor position estimates while minimizing errors and overcoming the limitations of dynamic conditions.

4.3.1. Traditional ML Models

Traditional ML models for Wi-Fi indoor positioning are typically categorized into two groups: supervised learning and unsupervised learning. Supervised learning methods depend on labeled data to train models to predict the user’s position based on signal patterns. Techniques such as KNN [17,117,118], Support Vector Machines (SVMs) [15], Decision Trees (DTs), and Random Forests (RFs) [15,36] are widely applied. One notable advantage of supervised learning methods is their ability to achieve high accuracy, as the corresponding model learns directly from labeled real-world data. These models perform well in stable environments where signal readings closely correlate with specific locations. However, it is well recognized that indoor environments are dynamic, with changes such as shifting furniture, new obstacles, and fluctuations in human movement, all of which can affect signal strength and reduce the accuracy of Wi-Fi IPSs [119]. For this reason, models often require re-training or fine-tuning to maintain accuracy in DEs [73]. This is because accuracy depends heavily on the density and distribution of access points and the quality of the labeled dataset.

In contrast, unsupervised learning does not require labeled data with known positions; the corresponding model independently identifies patterns and features within the dataset. This is beneficial in scenarios where obtaining labeled data is difficult or costly [120]. Moreover, Wi-Fi fingerprinting has a major limitation: the performance of these techniques can decline when the corresponding dataset contains hundreds or thousands of stored fingerprints. This leads to slower response times during the online phase since each incoming fingerprint must be compared with every fingerprint in the radio map [121]. Consequently, this method is inefficient for real-time indoor positioning. To improve response time in the online phase, some researchers have proposed using unsupervised methods, including clustering algorithms. Techniques such as k-Means [122,123], Density-based Spatial Clustering of Applications with Noise (DBSCAN) [124], or C-Means [125] are commonly used to group similar fingerprints in a radio map. These clusters can then be associated with specific areas or locations, thus providing an estimate of a user’s current position. However, unsupervised methods usually offer lower accuracy compared to supervised methods, as these approaches do not learn from explicit examples of locations [126]. Additionally, mapping discovered clusters to actual physical locations can be challenging and may require manual intervention or supplementary data.

In summary, supervised learning is valuable due to its ability to provide accurate location predictions. However, its capacity becomes limited when handling unpredictable environmental changes. Although unsupervised learning is more adaptable to DEs, the lack of labeled training data often results in less precise location estimates compared to supervised methods. Therefore, integrating semi-supervised or hybrid learning models is essential for enhancing the reliability of IPSs, enabling them to better address challenges posed by varying conditions, especially in DEs.

4.3.2. Deep Learning (DL) Models

Motivated by the development of AI, recent years have shown the potential of DL methods such as the Multilayer Perceptron (MLP) [127], Convolutional Neural Network (CNN) [2,11,128], Recurrent Neural Network (RNN) [129,130], and Deep Q-Network (DQN) [131], which have outperformed traditional ML approaches in terms of Wi-Fi indoor positioning. Unlike traditional ML methods, DL has the ability to learn complex features from large datasets with no extensive manual feature engineering. Moreover, DL models can adapt to varying conditions and incorporate temporal dependencies. This makes DL particularly suitable for addressing the complexities of Wi-Fi signals in dynamic indoor environments.

However, training DL models requires substantial computational resources and large amounts of data, which can be difficult and expensive to obtain. Additionally, DL-based IPSs can be prone to overfitting, especially if the training data is not representative of the real-world deployment environment. To address these issues, some researchers have proposed to use transfer-learning techniques since this method allows the learning of new information in a different environment by applying knowledge acquired previously [8,132,133,134,135,136]. For example, transfer learning can improve system scalability without requiring extensive site surveys and without compromising accuracy, particularly in situations where labeled data is inadequate [137]. This approach can significantly reduce the amount of labeled data and computational resources needed for training, ensuring the accuracy of indoor positioning, especially in a DE.

Fluctuations in RSSI values are the primary issue that will impact positioning accuracy. DL has demonstrated significant potential in enhancing localization in DEs, especially in situations where extracting and modeling non-linear correlated features prove challenging [138]. Traditional methods may struggle to effectively capture and model these intricate relationships. However, DL models are adept at learning hierarchical representations of data, enabling them to automatically extract relevant features from raw input data [18,139,140,141,142,143]. This capability makes DL techniques suitable for addressing the challenges posed by complex environments, where traditional methods may fall short. Therefore, DL techniques have shown great promise in improving localization accuracy in such scenarios [108].

Table 8. Overview of indoor localization schemes.

ID	Method	Public Datasets	Classification Accuracy	Positioning Error (Best Result)	Cost	Complexity	Scalability	Experiment Environment
[2]	CNN	√	98.37%	-	Medium	Medium	Medium	Multiple floors
[11]	CNN	√	-	Avg 7.60 m	Medium	High	High	Multiple floors
[144]	DBNs	×	-	Avg 1.9 m	Low	Medium	Medium	Single floor (540 m2)
[15]	RF, XGBoost, KNN, SVM	×	-	Avg 1.53 m	Low	Low	Medium	Multiple rooms (88 m2)
[17]	KNN	√	>90%	-	Medium	Medium	Medium	Multiple floors and buildings
[38]	IPC + FS-KNN	×	-	2.46 m (RMSE)	Low	Medium	Medium	Single room (300 m2/192 m2)
[41]	BP neural network	×	-	3.52 m (MSE)	Low	Medium	Low	-
[36]	1-KNN, GNB, SVC, RF	√	68.50%	Avg 5.65 m	Medium	Low	Low	Multiple floors and buildings
[37]	Gauss filtering + Bayes probability	×	-	Avg 1.3 m	Low	Medium	Low	Single floor (72 m2)
[125]	KNN + FPC	√	-	Avg 1.425 m	Medium	Low	High	-
[115]	IAGA-BP	√	-	Avg 1.07 m	Medium	High	Medium	-
[122]	K-means	√	-	Avg 0.871 m	Medium	Low	High	Multiple floors and buildings
[145]	SAS	√	-	Avg 2.04 m	Medium	Medium	High	-
[50]	1-D CNN	×	Floor (70.50%) and Region (81.23%)	Avg 3.47 m	High	Medium	Medium	Multiple floors (740 m2)
[61]	AReAE + FADet	×	-	Avg 3.4 m	Medium	High	High	Single building and floor (8400 m2)
[14]	Two consecutive multi-layer perceptrons	√	-	Floor (9.34833 × 10−1 and Variance (4.42699 × 10−5)	Medium	Low	Medium	Multiple buildings and floors
[146]	CNN	√	95.41%	-	Low	Medium	Medium	Multiple buildings and floors
[11]	CNN	√	96.03%	Avg 11.78 m	Medium	High	High	Multiple buildings and floors
[117]	NN, KNN, WKNN, Bayesian fingerprint matching	×	-	Avg 0.8323 m	Low	Low	Low	Single floor (900 m2)
[147]	1-D CNN	×	-	1.24 m (MSE)	Low	Medium	Low	Single floor (740 m2)
[144]	Deep belief networks	×	-	Avg 1.38 m	Low	Low	Low	Single floor (540 m2)
[148]	GNN	√	97.2%	-	Medium	Medium	Medium	Multiple buildings and floors
[149]	CAE + CNN	√	Building (99.96%) and Floor (93.87%)	Avg 6.95 m	Medium	High	High	Multiple buildings and floors
[118]	LR, DTR, BR, LSVR, XGBR, KNN, ANN	√	-	34.197 (MSE) and 5.848 (RMSE)	High	High	High	-
[129]	LSTM + RNN	√	5-layer LSTM (99.7%)	Avg 2.5–2.7 m	Medium	High	Medium	-
[130]	Vanilla RNN, LSTM, GRU, BiRNN, BiLSTM, BiGRU	√	-	Avg 0.75 m	High	High	Medium	Single floor (336 m2)
[131]	DQN	√	-	0.55 m	Medium	High	Low	Single building and floor (25,000 m2)
[123]	SVM + Kmeans + ReliefF + SAE	√	96%	-	Medium	Low	Medium	Multiple buildings and floors

provides a comparative analysis of different schemes used for Wi-Fi IPSs. It compares their performance, accuracy, and practical considerations of various approaches applied in this domain. Based on the results, one can see that tools like CNNs and LSTMs show high accuracy and excel in large and complex environments but have increased complexity. Simpler methods, such as KNN and traditional ML approaches, are cost-effective and easy to implement but lack scalability. Comparing these methods directly is challenging, as different datasets (either public or self-collected) may be used for experiments. Therefore, this also highlights the importance of public datasets, which enable standardized evaluation and support fair comparisons across methods.

4.4. Risk of Bias, Reporting Bias, and Certainty of Evidence

A formal risk of bias assessment was not conducted for each included study because existing tools are not well-suited to the methodological diversity of Wi-Fi-based indoor positioning research. Instead, potential sources of bias were addressed narratively. Common concerns included incomplete reporting of experimental conditions, selective presentation of favorable outcomes, and limited validation under realistic dynamic settings. In addition, a formal statistical assessment of reporting biases, such as funnel plot asymmetry, was not feasible due to the heterogeneity of studies.

5. Discussion and Suggestions

By examining the existing literature and drawing upon established research findings, in this section, we aim to address the limitations mentioned in existing works, clarifying areas for further exploration and proposing recommendations for future research.

5.1. Optimization of ML Algorithms

As discussed in RQ1, complex environments with different layouts and structures, moving objects within indoor spaces, device heterogeneity, and the distance between or number of access points are dynamic factors that impact the inconsistencies of Wi-Fi signal strength. These dynamic factors induce variability and unpredictability into indoor environments, potentially leading to signal attenuation, reflection, multipath effects, or fluctuations in the RSSI. Although various approaches have integrated ML to address environmental change, there appears to be limited emphasis on developing ML models specifically designed for this issue. Existing solutions that utilize general algorithms face challenges related to high computational complexity and significant demands on device performance. Therefore, developing specialized ML algorithms that are both cost-effective and optimized for handling environmental changes represents a promising direction for future research. For instance, approaches such as Interval Random Parameter Lognormal Shadowing–Adaptive Bayesian Comprehensive Learning (IRPLS-ABCL) and Inverse-Distance-Weight-Assisted Particle-Swarm-Optimized Indoor Localization (IDWPSOInLoc) have shown enhanced robustness and accuracy in dynamic indoor environments by effectively mitigating signal fluctuations and adapting to varying environmental conditions [150,151].

5.2. Hybrid Method for DEs

With regard to strategies for addressing DE factors, integrating technologies such as inertial sensors and fingerprinting also presents a promising approach to enhancing indoor positioning accuracy. Currently, smartphones are equipped with built-in inertial sensors, including accelerometers, gyroscopes, and magnetometers. These sensors provide data on movement and orientation, while fingerprinting uses pre-recorded signal characteristics to estimate positions. Combining these two methods can reduce their respective limitations. For example, fingerprinting can correct the drift errors typically associated with inertial sensors, while inertial sensors can provide continuous location tracking in areas where fingerprinting alone may be less reliable. This hybrid method may improve overall accuracy, making it a valuable solution for dynamic indoor environments [152,153].

5.3. Emerging Role of Wi-Fi 6/6E/7

Currently, most existing IPSs rely on Wi-Fi (802.11 standards), utilizing the signals emitted by the Wi-Fi access points to estimate the position of the target [16,47,48,154,155,156]. Recent advancements in Wi-Fi technology, especially Wi-Fi 6 (IEEE 802.11ax), Wi-Fi 6E, and the emerging Wi-Fi 7 (IEEE 802.11be), have significantly improved the potential for high-accuracy indoor positioning. These newer standards utilize wideband channels (up to 160 MHz or higher) and advanced physical-layer features, enabling more precise fine-time measurements (FTMs) and time-of-flight (ToF) estimations [157,158]. By harnessing these capabilities, Wi-Fi systems can achieve sub-meter-level accuracy, substantially improving upon traditional RSSI-based methods, which are often affected by multipath fading and environmental noise. In the future, the adoption of these advanced Wi-Fi standards could enable more reliable and scalable indoor localization frameworks, especially when combined with machine learning or sensor fusion approaches. Therefore, exploring the potential of wideband Wi-Fi standards represents a promising research direction for enhancing both the accuracy and resilience of indoor positioning systems.

5.4. Adaptability of Radio Maps

In indoor localization, a well-constructed radio map is essential for enhancing the efficiency and accuracy of Wi-Fi indoor positioning. However, constructing a radio map is a complex and demanding task. Therefore, several methods have been proposed to simplify this process, as mentioned in RQ2. The results indicate that each technique for radio map construction in Wi-Fi indoor positioning has its own advantages and limitations. Creating a radio map that can adapt to a DE remains a significant challenge. Consequently, there is an increasing need for adaptive radio map construction algorithms that can identify environmental changes and update the map automatically. Approaches such as crowdsourcing for data collection and the use of semi-supervised learning methods show potential in addressing this challenge of adaptive radio map construction [159]. Researchers should carefully consider the trade-offs and choose techniques based on the specific requirements of the application environment to effectively utilize the benefits of these methods for improving indoor localization performance.

5.5. The Capabilities of DL and Transfer Learning

As outlined in research question RQ3, DL-based models require large amounts of annotated data to train. However, this kind of extensive data collection requires a time-consuming and labor-intensive site survey, which may not be suitable for practical scenarios [135,160,161]. The number of labeled samples is often restricted, which may lead to overfitting of the trained model [143,161].

Moreover, the models based on DL are commonly trained on data from a single building and cannot be applied to other buildings since these models learn the relationship between RSSI values and a location within a particular environment. Therefore, there is a need to explore and develop an adaptive Wi-Fi positioning algorithm using DL that can generalize to different environments with only limited data. For example, Few-Shot Learning (FSL) techniques are an ML paradigm designed to address cases where limited annotated training data is available for model development [162,163,164]. In FSL, a model can learn about new or unseen data from a limited set of labeled samples. The model can then continuously refine and improve its performance by adapting to new data. By periodically incorporating new labeled data into the training process, the model can adjust to changes in the environment, device features, or user behavior, leading to more accurate and reliable positioning estimates. Therefore, future research on Wi-Fi IPSs based on RSSI values should explore the potential of this approach as a dependable method. The changing environment requires innovative IPSs, encouraging competition and progress in the field.

Most of the existing RSSI-based IPSs are designed to operate within single-level structures such as individual floors of buildings or specific areas like laboratories or rooms [35,38,45,46,165,166]. However, the fluctuation of wireless signals resulting from environmental uncertainties has become a significant challenge on a single floor. It can be expected that this challenge may become more substantial when expanding positioning coverage to multiple floors or three-dimensional spaces [73]. Hence, future research should focus on identifying floors within multi-level structures, a task that would require addressing the challenges posed by DEs, such as signal attenuation, reflection, or multipath effects. To achieve multi-level positioning, more advanced ML or DL techniques should be further investigated to enhance the corresponding models (e.g., by improving accuracy and accelerating or stabilizing training). By employing the capabilities of ML or DL to process complex data and infer vertical position, accurate multi-level positioning can be attained.

5.6. The Possibility of Channel-State Information (CSI)

Today, the RSSI has become mainstream in Wi-Fi indoor positioning due to its advantage of imposing no additional equipment requirements. However, extracting the abundant multipath information present in the subcarriers of orthogonal frequency division multiplexing (OFDM) poses a significant challenge, primarily because of the limitations of RSSI values in providing only a rough representation of the wireless channel [167,168]. This can cause difficulties in adapting to dynamic environmental conditions and lead to inaccuracies in indoor positioning. Compared with the RSSI, CSI utilizes the fine-grained features for accurate positioning, and integrating visual elements with Wi-Fi localization expands the potential for acquiring more comprehensive and detailed location information [4,167,169,170,171]. In future research, it will be essential to investigate the feasibility of using CSI as a reliable approach.

6. Conclusions

This paper is an SLR of Wi-Fi indoor positioning based on the RSSI. First, the foundational concepts of RSSI, such as signal strength levels, advantages, and limitations, were presented. Before writing this study, a comprehensive research methodology was developed. This methodology encompassed defining research questions, developing a data search strategy, defining paper selection criteria, selecting papers, extracting data, and synthesizing data. Finally, 128 review papers published between 2018 and 2024 were selected to address the specified research questions.

Based on the results regarding RQ1, it can be concluded that the performance of Wi-Fi-based indoor positioning systems is strongly related to a variety of dynamic factors. These dynamic factors include complex layouts and structures, environmental changes, the movement and presence of objects, device heterogeneity, and variations in the number of and distance between access points. These dynamic factors introduce variability and unpredictability into an indoor environment, potentially leading to signal attenuation, reflection, multipath effects, or fluctuations in the RSSI. Despite many efforts to address these dynamic factors in Wi-Fi indoor positioning systems through improved algorithms and optimized deployment strategies, the dynamic nature of these factors makes it challenging to fully adapt existing solutions, thus emphasizing the need for further research and innovation in order to enhance the reliability and robustness of such systems in real-world applications.

For RQ2, the results indicate that each method of radio map construction in Wi-Fi indoor positioning possesses its own set of advantages and limitations. Researchers should consider the trade-offs and select techniques based on the specific requirements of the application environment to effectively utilize the benefits of these methods for improving indoor localization performance. For example, integrating crowdsourced data with automated techniques could increase scalability and reduce manual effort, while the application of advanced machine learning models could enhance the accuracy and adaptability of radio maps in dynamic environments. For further research, it is essential to examine methods to ensure the long-term stability and real-time updating capability of radio maps, especially in environments subject to frequent changes.

For RQ3, the results have indicated that both ML and DL hold significant potential for predicting indoor positions with minimal error. Notably, DL demonstrates exceptional promise, particularly in addressing the challenges posed by DEs. By utilizing the power of data-driven algorithms, DL models can adjust to changing conditions, enabling the extraction and modeling of complex features from raw input data. Hence, a promising direction for further research is to explore the development of advanced DL architectures designed for indoor positioning in dynamic environments. This includes examining techniques such as transfer learning or Few-Shot learning to adapt models across different environments.

Lastly, a discussion section has been included to explore the findings, limitations, and potential directions for future research based on the review papers. Although Wi-Fi indoor positioning based on RSSI values still faces various challenges, it is firmly believed that positioning will play an increasingly important role in wireless networks in the future.