Enhancing Indoor Localization with Machine Learning and WiFi Signals from Mobile Devices

View all paper

Abstract

Indoor Positioning Systems (IPS) are crucial technologies that provide accurate location information in enclosed environments such as shopping malls, hospitals, and airports, where GPS signals are either inaccessible or significantly degraded. However, traditional indoor positioning methods like RFID, Ultrasonic, Ultra-Wideband, etc., often face challenges such as high costs, complex infrastructure requirements, and inconsistent accuracy. A new approach involves using Wi-Fi messages from personal devices like smartphones, smartwatches, laptops, etc. When a Wi-Fi-enabled device searches for a network to connect to, it emits containing information such as MAC address and Received Signal Strength Indicator (RSSI). This approach offers several advantages: low deployment costs, no user interaction required, and the ability to estimate the location of the signal-emitting device with reasonable accuracy. Experimental results have shown that the maximum error is only 0.28 meters, accuracy up to 0.25m, and detection rate 93% in a room. The proposed method can also achieve moderate computational complexity and moderate real-time performance compared to other methods.

INTRODUCTION

Indoor Positioning Systems (IPS) are essential for providing location information in environments where GPS is ineffective, such as shopping malls, hospitals, and airports. Unlike outdoor positioning that relies on GPS, indoor positioning utilizes various methods due to the obstruction of GPS signals by indoor structures. The growing need for IPS is evident in its diverse applications. For instance, in shopping malls, IPS can enhance customer navigation to stores; in hospitals, it can track equipment and guide visitors; and in airports, it can assist passenger navigation. Additionally, in industrial settings such as factories, warehouses, and large manufacturing plants, IPS is crucial for asset tracking and crowd management, enhancing safety and operational efficiency.

Common indoor positioning techniques include Angle of Arrival (AoA), which is accurate but requires specialized hardware; Time of Arrival (ToA), which is precise but costly; and Fingerprinting, which uses existing infrastructure but needs extensive calibration [1]. Kalman Filters are also used to enhance accuracy through measurement over time, although they require complex system models [2].

This research introduces a cost-effective approach using Wi-Fi from personal devices, which emit signals containing information like MAC addresses and RSSI when searching for networks. These signals are collected by Wi-Fi access points or receivers, allowing for passive, user-friendly location estimation without needing specialized hardware or user interaction. The study evaluates various machine learning models, including SVM, KNN, Naive Bayes, Linear Regression, and Decision Tree, to identify the most effective method for indoor positioning. It also explores factors affecting accuracy and suggests improvements for future work.

The remainder of the paper is organized as follows. Section II reviews related work and existing methods for indoor positioning. Section III details the proposal methodology. Section IV details experimental including the experimental setup, data collection, and machine learning models used. Section V describes the implementation process and presents the results of our experiments, comparing them with related research. Finally, Section VI concludes with a summary of our findings and discusses potential future directions for research and improvements.

RELATED WORK

Indoor Positioning Systems (IPS) have been extensively researched due to their wide applications and the limitations of GPS indoors. Various technologies and methods have been developed to enhance their accuracy and efficiency.

• Ultrasonic positioning uses sound waves above 20,000 Hz and calculates distances by measuring the time of flight (ToF) between transmitters and receivers [3]. This method offers centimeter-level accuracy but is affected by air density and humidity, limiting its reliability and range.
• RFID technology [4] employs electromagnetic waves for data communication between RFID readers and electronic tags. It is cost-effective and easy to deploy, commonly used in warehouses and libraries.
• Ultra-Wideband (UWB) technology uses anchor nodes and blind nodes for positioning through triangulation or ‘fingerprint’ methods [7].
• Bluetooth positioning, especially using iBeacon technology, relies on RSSI values to estimate positions [5]. It is easy to implement but accuracy depends on the density and power of beacons, and indoor multipath fading can cause high errors. Bluetooth 5.1 introduces new features like direction finding, enhancing BLE technology [8].
• Infrared positioning works well in environments with few obstructions but is affected by heat sources and lighting [9].
• Zigbee, a low-power communication protocol, forms a network between reference nodes and gateways for positioning [6]. It is low-cost and power-efficient but has a small range and poor interference resistance.
• WiFi-based positioning uses a network of access points and RSSI-based fingerprinting for location estimation [6]. It is widely used due to the ubiquity of WiFi-enabled devices, though accuracy depends on calibration density.
• Hybrid methods combine technologies to leverage strengths but can be complex and costly [10].

Table 1 - Comparison of indoor positioning methods.

Name	Method	Accuracy	Range	Cost	Environmental Influence	Strengths	Weaknesses
Carotenuto, R. et al. (2019) [3]	Ultrasonic	Up to cm level	Small	Low	Affected by air density and humidity	High accuracy	Susceptible to environmental changes
Magnago, V. et al. (2020) [4]	RFID	Up to cm level	Up to tens of meters	Low	Minimal effect	Low cost, easy to deploy	Short range
Ma, W. et al. (2024) [7]	Ultra-Wideband (UWB)	Up to cm level	Medium	High	Minimal effect	High accuracy, minimal interference	High cost
Yu, K. et al. (2019) [5]	Bluetooth	Up to meter level	Medium	Low to Medium	Affected by indoor environment	Easy to deploy, low power consumption	Accuracy depends on beacon density and power
Phutcharoen, K., Chamchoy, M. and Supanakoon, P. (2020) [8]	Infrared	Up to cm level	Small	Medium	Affected by heat sources and lighting	Works well in unobstructed environments	Affected by heat sources and lighting
Yelkovan, Y. et al. (2014) [9]	Zigbee	Up to meter level	Small	Low	Affected by interference	Low power consumption, low cost	Small range, high positioning error
Bianchi, V., Ciampolini, P. and De Munari, I. (2019) [6]	WiFi FIngerprint	Up to meter level	Large	Low to Medium	Affected by calibration density	Widely used, easy to deploy	Accuracy depends on calibration point density
Kao, C.-H. et al. (2017) [10]	Hybrid	Up to meter level	Large	Medium to High	Depends on combined technologies	Combines strengths of multiple technologies	Complex, high deployment cost

PROPOSAL METHOD

The indoor positioning system uses Wi-Fi and machine learning, relying on Wi-Fi fingerprinting to collect signal characteristics at known locations. This technique estimates positions without requiring line-of-sight, associating RSSI values from Wi-Fi access points with specific indoor locations. It is effective in environments with multipath effects and non-line-of-sight issues, aiming to provide an accurate and practical solution for real-world applications.

Wi-Fi capable devices, like smartphones, often automatically send out probe requests in order to detect available Wi-Fi networks in the area. These probe requests include a device identifier known as a MAC address [11]. By taking advantage of probe frame waves, we will measure the device’s broadcast power at different point coordinates, then we propose to use a machine learning model to calculate and determine the device’s location based on the data collected from the previous step. Fig 1 below will clearly describe the system flow chart.

Theoretical principle

In the study, the layout of the example room is depicted in Fig 2. The actual dimensions of the room are nine meters by seven point five meters. It is noted that the Wi-Fi signal experiences a 15% attenuation when passing through walls and around a 25% attenuation when in proximity to strong magnetic fields [12]. The calculation is done using the following formula Equation 1:

\[RSSI = -10 \cdot n \cdot \log\_{10}(d) + A \quad (1)\]

where:

• $A$ represents the signal strength at a distance of one meter from the transmitting source;
• $n$ is the environmental attenuation factor;
• $d$ is the distance between the transmitting source and the receiving end.
• $RSSI$ is the Wi-Fi signal strength value.

So, the estimated distance $d$ between the mobile device points $MD$ and the $m$-th AP ($AP_m$) can be expressed as:

\[d = 10^{\frac{RSSI - A}{10 \cdot n}} \quad (2)\]

To estimate the distance between the testing device and each access point (AP) using RSSI, we employ a triangulation algorithm to create a wireless channel model. This method uses signals from three APs to narrow the possible positioning points to no more than two, minimizing error by determining the intersection of the three signals $(x_k, y_k)$, and a given access point $AP_m$ $(x_m, y_m)$ is expressed using the Euclidean distance formula:

\[d_k = \sqrt{(x_m - x_k)^2 + (y_m - y_k)^2} \quad (3)\]

where: $d_k$ the estimated distance between the mobile devices testing position $(x_k, y_k)$ and the $m$-th access point $(x_m, y_m)$.

The receivers in the system are equipped with antennas that have a gain of 5 dBi. These antennas enhance the ability of the receivers to capture Wi-Fi effectively by providing a higher received signal strength. The received power can be estimated using the Friis transmission equation:

\[P*r = P_t + G_t + G_r - 20 \log*{10}(d) - 20 \log*{10}(f) - 20 \log*{10}\left(\frac{4\pi}{c}\right) \quad (4)\]

where:

• $P_r$ is the received power in dBm;
• $P_t$ is the transmitted power in dBm;
• $G_t$ is the gain of the transmitting antenna in dBi;
• $G_r$ is the gain of the receiving antenna in dBi;
• $d$ is the distance between the transmitter and receiver in meters;
• $f$ is the frequency of the transmitted signal in Hz;
• $c$ is the speed of light in m/s.

Preprocessing the Captured Data:

Once the Wi-Fi were captured, the raw data needed to be preprocessed to ensure it was suitable for training the machine learning models. The preprocessing steps included:

• Noise Removal - Filtering: Noisy and irrelevant data is filtered out to maintain the quality of the data set. This step will delete any frames that are incomplete on all 3 receivers or are corrupted.

• Outlier Detection and Removing: Outliers are due to transient disturbances or abnormalities of the receiver’s hardware. We used the Hampel filter to identify and remove RSSI outliers. It replaces the outlier-sensitive estimates of the mean and standard deviation with the more robust median and median absolute deviation (MAD). The latter is defined as:

  \[R*{ji} = 1.4826 \times \text{median} | r*{ji}(t*m) - \text{median}(r*{ji}) | \quad (5)\]

  The factor 1.4826 was selected to ensure that the expected value of $R_{ji}$ matches the standard deviation for normally distributed data. $R_{ji}$ represents the robust estimate of the standard deviation for the data set indexed by $i$ and $j$. $r_{ji}(t_m)$ is the specific RSSI data value at time $t_m$, and $\text{median}(r_{ji})$ is the median of a RSSI data set. The MAD-scaled version of the data is:

  \[MAD*{ji}(t_m) = \frac{r*{ji}(t*m) - \text{median}(r*{ji})}{R\_{ji}} \quad (6)\]
• Normalization RSSI Values: The RSSI values were normalized to standardize data for comparison, accounting for varying signal strengths caused by environmental factors like distance and obstructions.
• Data Structuring: The processed data was organized into a structured format. Each data entry included the coordinates of the grid point, the RSSI values from the three receivers, the MAC address of the mobile device, and the timestamp of the capture.

Evaluate model performance:

To determine the best model for indoor positioning, we evaluated several popular machine learning algorithms known for their effectiveness in classification tasks. The chosen models include:

• Support Vector Machine (SVM) [14]: SVM is a powerful classification algorithm that works well with high-dimensional data aimed to find the optimal hyperplane.
• K-Nearest Neighbors (KNN) [15]: KNN is a simple, yet effective algorithm that classifies a data point based on the majority class of its nearest neighbors.
• Decision Tree (DT) [16]: Decision Tree algorithms partition the data into subsets based on feature values, and each leaf node represents an outcome.
• Linear Regression (LR) [17]: is commonly associated with predicting linear relationships between variables. This study proposes using LR to detect trends and patterns in data, thereby making classification decisions.
• Naive Bayes (NB) [18]: is a probabilistic classification method, known for its ability to efficiently handle large and complex datasets quickly.

Real time positioning:

The Real-Time Positioning stage of the indoor positioning system uses a trained machine learning model to estimate live positions based on real-time data. Key steps include:

1. Deploy positioning system: Deploy the top-performing model from the Data Processing stage into the operational environments.
2. Capture Real-time Wi-Fi : These contain crucial information such as the device’s MAC address and RSSI values, which are essential for position estimation.
3. Estimate position using best model: Feed real-time data, including RSSI values and other relevant features into the machine learning model to estimate the device’s position in the indoor environment.
4. Provide position estimate to user: Provide real-time position estimates to the user for applications like navigation, asset tracking, or crowd management.

EXPERIMENTAL AND RESULT

Experimental Setup

To implement the indoor positioning system, we installed three Wi-Fi receivers in a 9m by 7.5m rectangular room with the following characteristics:

• Wall composition: Standard drywall with metal studs
• Flooring: Vinyl tiles over concrete
• Ceiling: Suspended ceiling tiles at a height of 2.8m
• Furniture: The room contained typical office furniture including desks, chairs
• Electronics: Several desktop computers, monitors, and a network printer were present in the room
• Lighting: Fluorescent overhead lighting fixtures

Receivers were strategically placed at fixed, precise coordinates to enhance data accuracy. The room was divided into a grid with specific intervals, allowing systematic data collection at coordinates from (1,1) to (8,7) (Table 2). These receivers captured Wi-Fi from a Samsung Galaxy Note 20 at each grid point during data sessions. We placed the device on a non-metallic tripod at a height of 1.2m to simulate typical hand-held use. The setup allowed detection from multiple angles for a comprehensive dataset. Devices emitted continuous Wi-Fi probes with a fixed MAC address for simplified later filtering.

• Grid Points: The room was divided into a grid with coordinates ranging from (1,1), (1.5, 1.5) to (8,7) in the x-y plane. Each grid point represented a known location where data would be collected (Fig. 2).

Fig. 2 -The 3 reference points (AP1, AP2 and AP3) and 195 mobile device points (MD1–MD195).

Data Collection

In the proposed indoor positioning system, accurate data collection is essential for reliable machine learning models. This involves capturing, preprocessing, and organizing the data to ensure precise position estimation.

Table 2 - Parameters values used in this study

Parameters	Values
Dimension (Unit: m)	9m x 7.5m
Samplings (N) (Unit: samples)	5048
Sampling points (N) (Unit: points)	195
Location (x, y) (Unit: m)	$AP_1(x_1, y_1), AP_2(x_2, y_2), \dots, AP_m(x_m, y_m)$
Mobile Device points (x, y) (Unit: m)	$MD_1(x_1, y_1), MD_2(x_2, y_2), \dots, MD_k(x_k, y_k)$

Wi-Fi , broadcast by devices searching for networks, contain MAC addresses and RSSI values indicating signal strength. Receivers log these values, timestamps, and MAC addresses. Data is collected at each grid coordinate every 15 minutes, with approximately 40-70 samples per point. In our experiments, we measured that when a Wi-Fi enabled device is actively searching for a network, it will send a probe request every 1-10 seconds. However, this may vary depending on the device’s power management settings and whether the device is active or idle.

Data collection was conducted during peak (10:00-12:00, 14:00-16:00) and off-peak (08:00-10:00, 12:00-14:00, 16:00-18:00) hours over one week. Room temperature (24-26°C) was monitored with a digital hygrometer. While no significant effects on RSSI values were observed, potential long-term impacts are worth future consideration. The number of people in the room averaged 2-3, with occasional peaks of up to 8, causing signal attenuation of 3-5 dB.

Table 3 Collected RSSI values

No.	P1	P2	P3	x	y
1	-73	-34	-68	1	1
2	-67	-44	-71	1.5	1
3	-70	-57	-69	2	1
4	-68	-59	-67	2.5	1
5	-69	-63	-65	2	1.5
…	…	…	…	…	…
5048	-70	-75	-57	8	7

Table 3 presents the Received Signal Strength Indicator (RSSI) values collected from three Wi-Fi receivers at different grid points in the room. The columns include RSSI values from the three receivers ($P_1, P_2, P_3$), grid coordinates (x, y). This data went through cleaning steps to train machine learning models to accurately estimate locations based on Wi-Fi signal strength.

Fig. 3 - RSSI Distributions Before and After Preprocessing

Fig 3 demonstrates that after performing noise removal and outlier detection, the RSSI distributions for $AP_1, AP_2, AP_3$ show significant improvement. $AP_1$ distribution becomes more concentrated around -60 dBm, indicating less noise. $AP_2$ distribution, which was initially wide and noisy, becomes much narrower after processing, suggesting effective noise reduction. $AP_3$ shows slight improvement, with fewer outliers and a clearer signal range. After noise removal total sampling remaining is 4396 samples.

Evaluated Machine Learning Models

This section details the machine learning models used to estimate mobile device positions based on collected Wi-Fi probe frame data, with the goal of identifying the most effective algorithms for accurate indoor positioning. The evaluated models include Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Decision Tree (DT), Linear Regression (LR), and Naive Bayes (NB). An overview of these models is provided, along with the supervised learning approach and methods for feature extraction and selection.

Supervised learning method

Supervised learning involves training models using labeled data. In this case, the input to the models consists of reference RSSI values at each point in time ($P_1, P_2, P_3$), and supervised learning is used based on the coordinates ($x, y$) of the access points (APs). The output of the models is the predicted coordinates ($x, y$), rounded to two decimal places, representing the expected position of a mobile device in a two-dimensional space.

The steps to apply supervised learning in our research are as follows:

1. Data Splitting: The collected dataset (4396 samples after preprocessing) was divided into three sets:
   • Training Sets: 75% of the data (3297 sets) used to train the models.
   • Validation Sets: 15% of the data (659 sets) used to tune hyperparameters and evaluate intermediate model performance.
   • Testing Sets: 10% of the data (440 sets) used for unbiased evaluation of the final model.
2. Model Training: Each model was trained on the training set using the extracted features from RSSI signals. Models were trained using RSSI signal features with the following parameters: 100 epochs, batch size of 10, Adam optimizer with default learning rate, and hyperparameter tuning to enhance performance.
3. Model Validation: After training, the models were validated on the testing set to assess their accuracy and robustness. The validation process included:
   • K-Fold Cross-Validation [19]: To prevent overfitting, we used k-fold cross-validation (typically k=5 or k=10), which divides the training set into k subsets. Each time, k-1 subsets train the model and the remaining subset validates it. This process cycles k times, with each subset serving as validation data once.
   • Performance Evaluation: Metrics such as Root Mean Square Error (RMSE), Precision, Recall, and F1-Score were used to evaluate the overall performance of the models on the validation set. RMSE served as a metric to measure the model’s predictive accuracy, calculating the square root of the average of squared differences between predicted and actual values [20]. Lower RMSE values indicate better model accuracy.

The results are summarized in Table 4.

EVALUATION

Table 4 compares the performance of various models for indoor positioning. KNN is the best performer with 91.65% accuracy within 0.25m and the lowest RMSE (0.24). Naive Bayes x also performs well with 87.36% accuracy and an F1-score of 91.55%, while Naive Bayes y has lower scores. SVM performs well, especially on the y-axis with 79.11% accuracy. Linear Regression x shows 76.16% accuracy, but Linear Regression y underperforms with only 49.16% accuracy. Decision Tree performs moderately. Overall, KNN is the top model, with SVM and Naive Bayes x also being solid options.

KNN’s ability to handle non-linear data without requiring assumptions about data distribution makes KNN particularly effective. KNN leverages local information from nearby points, which is especially useful for location estimation based on variations in Wi-Fi signal strength (RSSI). Moreover, KNN has low complexity during the training phase, allowing it to efficiently process large datasets.

Table 4 Accuracy, F1-score, RMSE, Recall of classifications.

Classifier	Accuracy within 0.25m (%)	F1-Score (%)	RMSE	Recall (%)
SVM x	76.34%	84.67%	0.58	85.67%
SVM y	79.11%	88.34%	0.63	88.34%
Naive Bayes x	87.36%	91.55%	0.69	87.36%
Naive Bayes y	77.00%	87.01%	0.83	77.00%
LR x	76.16%	86.47%	0.46	76.16%
LR y	49.16%	65.91%	0.63	49.16%
KNN	91.65%	91.93%	0.24	91.65%
DT	77.85%	87.54%	0.45	77.85%

Table 5 provides a comprehensive comparison of different Wi-Fi based indoor positioning methods, highlighting their principles, advantages, and disadvantages.

Table 5 Comparison of Wi-Fi Based Indoor Positioning Methods

Paper	Method	Principle	Accuracy	Cost	Complexity	Advantages	Disadvantages
Yimwadsana, B. et al. (2019) [1]	Angle of Arrival (AoA)	Measures the angle at which signals arrive at the receiver	Medium (up to m level)	Medium	High	High accuracy, effective in multipath environments	Requires specialized hardware
Kul, G. et al. (2014) [21]	Time of Arrival (ToA)	Measures the travel time of signals to estimate distance	Medium (up to m level)	High	High	High accuracy, well-suited for line-of-sight conditions	Affected by synchronization errors, high cost
Nabati, M. and Ghorashi, S.A. (2023) [22]	Fingerprinting	Compares real-time signal strength to a pre-recorded database	Medium to High (up to cm level)	Low to Medium	Medium to High	No need for line-of-sight, leverages existing infrastructure	Labor-intensive calibration, sensitive to environmental changes
Sircoulomb, V. and Chafouk, H. (2022) [2]	Kalman Filter	Uses a series of measurements observed over time to estimate unknown variables	Medium	Medium	High	Smooths noise from measurements, improves estimate accuracy over time	Requires accurate model of system dynamics
Li, S. et al. (2018) [23]	Time Difference of Arrival (TDoA)	Measures the difference in arrival times of signals at different receivers	Medium (up to m level)	High	High	High accuracy, suitable for large areas	Requires precise synchronization
Proposed approach	Machine learning	Utilizes Wi-Fi and machine learning to estimate device location	High (up to 0.25m)	Low to medium	Medium	Low deployment costs, no user interaction required, moderate real-time performance	Requires sufficient training data, sensitive to environmental changes

Fig 4 describes the multiclass ROC curve. KNN is still a pretty good model with the second highest AUC for all classes (0.91, 0.90, 0.86). Decision Tree performs the worst with lower AUC scores. SVM and KNN are recommended for similar classification tasks.

Fig. 4 Multiclass ROC Curve with Various Models

Overall, using machine learning for indoor positioning is highly feasible for real-world applications like shopping malls and airports. Machine learning models provide high accuracy and robustness to environmental changes, adapting well to dynamic indoor settings. They can scale efficiently with large datasets, making them suitable for extensive indoor areas. Using existing Wi-Fi infrastructure makes these models cost-effective. A current challenge is the large data requirement, which could be addressed by using programmed robots for data collection. Modern tools make implementation simple, allowing real-time position estimates and customization for specific environments.

CONCLUSION

In this study, we explored the use of Wi-Fi for indoor positioning through various machine learning models in environments where GPS is ineffective, such as shopping malls and airports. The findings demonstrate the feasibility and benefits of applying machine learning for indoor positioning, with a maximum error of 0.28 meters and accuracy of 0.25 meters, achieving a 93% detection rate. Models like Naive Bayes and KNN performed exceptionally well, showcasing their potential to enhance the precision and efficiency of location-based services in enclosed spaces.

REFERENCES

[1] B. Yimwadsana, V. Serey, and S. Sanghlao, “Performance Analysis of an AoA-based Wi-Fi Indoor Positioning System,” in 2019 19th International Symposium on Communications and Information Technologies (ISCIT), Sep. 2019, pp. 36–41. doi: 10.1109/ISCIT.2019.8905238.

[2] V. Sircoulomb and H. Chafouk, “A Constrained Kalman Filter for Wi-Fi-Based Indoor Localization with Flexible Space Organization,” Sensors, vol. 22, no. 2, Art. no. 2, Jan. 2022, doi: 10.3390/s22020428.

[3] R. Carotenuto, M. Merenda, D. Iero, and F. G. Della Corte, “An Indoor Ultrasonic System for Autonomous 3-D Positioning,” IEEE Transactions on Instrumentation and Measurement, vol. 68, no. 7, pp. 2507–2518, Jul. 2019, doi: 10.1109/TIM.2018.2866358.

[4] V. Magnago et al., “Ranging-Free UHF-RFID Robot Positioning Through Phase Measurements of Passive Tags,” IEEE Transactions on Instrumentation and Measurement, vol. 69, no. 5, pp. 2408–2418, May 2020, doi: 10.1109/TIM.2019.2960900.

[5] K. Yu, K. Wen, Y. Li, S. Zhang, and K. Zhang, “A Novel NLOS Mitigation Algorithm for UWB Localization in Harsh Indoor Environments,” IEEE Transactions on Vehicular Technology, vol. 68, no. 1, pp. 686–699, Jan. 2019, doi: 10.1109/TVT.2018.2883810.

[6] Bianchi, V., Ciampolini, P. and De Munari, I. (2019) “RSSI-Based Indoor Localization and Identification for ZigBee Wireless Sensor Networks in Smart Homes,” IEEE Access, vol. 7, pp. 112505-112517, doi: 10.1109/ACCESS.2019.2934930. (Note: Updated citation based on [6] likely referring to Bianchi et al.) Original citation [6] pointed to an IEEE Xplore page without full details.

[7] W. Ma, X. Fang, L. Liang, and J. Du, “Research on indoor positioning system algorithm based on UWB technology,” Measurement: Sensors, vol. 33, p. 101121, Jun. 2024, doi: 10.1016/j.measen.2024.101121.

[8] K. Phutcharoen, M. Chamchoy, and P. Supanakoon, “Accuracy Study of Indoor Positioning with Bluetooth Low Energy Beacons,” in 2020 Joint International Conference on Digital Arts, Media and Technology with ECTI Northern Section Conference on Electrical, Electronics, Computer and Telecommunications Engineering (ECTI DAMT & NCON), Mar. 2020, pp. 24–27. doi: 10.1109/ECTIDAMTNCON48261.2020.9090691.

[9] Yelkovan, Y. et al. (2014) “Infrared beacon based sub-meter indoor localization,” in 2014 International Conference on Indoor Positioning and Indoor Navigation (IPIN), Oct. 2014, pp. 639-645, doi: 10.1109/IPIN.2014.7275542. (Note: Updated citation based on [9] likely referring to Yelkovan et al.) Original citation [9] pointed to an IEEE Xplore page without full details.

[10] C.-H. Kao, R.-S. Hsiao, T.-X. Chen, P.-S. Chen, and M.-J. Pan, “A hybrid indoor positioning for asset tracking using Bluetooth low energy and Wi-Fi,” in 2017 IEEE International Conference on Consumer Electronics - Taiwan (ICCE-TW), Jun. 2017, pp. 63–64. doi: 10.1109/ICCE-China.2017.7990996.

[11] X. Gu, W. Wu, X. Gu, Z. Ling, M. Yang, and A. Song, “Probe Request Based Device Identification Attack and Defense,” Sensors (Basel), vol. 20, no. 16, p. 4620, Aug. 2020, doi: 10.3390/s20164620.

[12] “Wi-Fi signal attenuation coefficients when passing through different materials,” Keenetic. Accessed: May 29, 2024. [Online]. Available: https://help.keenetic.com/hc/en-us/articles/213968869-Wi-Fi-signal-attenuation-coefficients-when-passing-through-different-materials

[13] E. Grossi and M. Buscema, “Introduction to artificial neural networks,” European Journal of Gastroenterology & Hepatology, vol. 19, no. 12, p. 1046, Dec. 2007, doi: 10.1097/MEG.0b013e3282f198a0.

[14] C. Cortes and V. Vapnik, “Support-vector networks,” Mach Learn, vol. 20, no. 3, pp. 273–297, Sep. 1995, doi: 10.1007/BF00994018.

[15] K. Taunk, S. De, S. Verma, and A. Swetapadma, “A Brief Review of Nearest Neighbor Algorithm for Learning and Classification,” in 2019 International Conference on Intelligent Computing and Control Systems (ICCS), May 2019, pp. 1255–1260. doi: 10.1109/ICCS45141.2019.9065747.

[16] L. Rokach and O. Maimon, “Decision Trees,” in The Data Mining and Knowledge Discovery Handbook, vol. 6, 2005, pp. 165–192. doi: 10.1007/0-387-25465-X*9.

[17] D. Maulud and A. Mohsin Abdulazeez, “A Review on Linear Regression Comprehensive in Machine Learning,” _Journal of Applied Science and Technology Trends*, vol. 1, pp. 140–147, Dec. 2020, doi: 10.38094/jastt1457.

[18] I. Rish, “An Empirical Study of the Naïve Bayes Classifier,” IJCAI 2001 Work Empir Methods Artif Intell, vol. 3, Jan. 2001.

[19] D. Berrar, “Cross-Validation,” in Encyclopedia of Bioinformatics and Computational Biology, 2018, pp. 542-545. doi: 10.1016/B978-0-12-809633-8.20349-X.

[20] T. Chai and R. R. Draxler, “Root mean square error (RMSE) or mean absolute error (MAE)?– Arguments against avoiding RMSE in the literature,” Geoscientific Model Development, vol. 7, pp. 1247–1250, Jun. 2014, doi: 10.5194/gmd-7-1247-2014.

[21] G. Kul, T. Özyer, and B. Tavli, “IEEE 802.11 WLAN based Real Time Indoor Positioning: Literature Survey and Experimental Investigations,” Procedia Computer Science, vol. 34, pp. 157–164, Jan. 2014, doi: 10.1016/j.procs.2014.07.078.

[22] M. Nabati and S. A. Ghorashi, “A real-time fingerprint-based indoor positioning using deep learning and preceding states,” Expert Systems with Applications, vol. 213, p. 118889, Mar. 2023, doi: 10.1016/j.eswa.2022.118889.

[23] S. Li et al., “TDOA-based passive localization of standard WiFi devices,” in 2018 Ubiquitous Positioning, Indoor Navigation and Location-Based Services (UPINLBS), Mar. 2018, pp. 1–5. doi: 10.1109/UPINLBS.2018.8559705.