A Review of Indoor Positioning Systems for UAV Localization with Machine Learning Algorithms

: The potential of indoor unmanned aerial vehicle (UAV) localization is paramount for diversiﬁed applications within large industrial sites, such as hangars, malls, warehouses, production lines, etc. In such real-time applications, autonomous UAV location is required constantly. This paper comprehensively reviews radio signal-based wireless technologies, machine learning (ML) algorithms and ranging techniques that are used for UAV indoor positioning systems. UAV indoor localization typically relies on vision-based techniques coupled with inertial sensing in indoor Global Positioning System (GPS)-denied situations, such as visual odometry or simultaneous localization and mapping employing 2D/3D cameras or laser rangeﬁnders. This work critically reviews the research and systems related to mini-UAV localization in indoor environments. It also provides a guide and technical comparison perspective of different technologies, presenting their main advantages and disadvantages. Finally, it discusses various open issues and highlights future directions for UAV indoor


Introduction
In order to enhance the connectivity of the wireless communication network, various network enhancement methodologies can be adopted, such as the millimeter-wave (mmwave) frequency band [1,2], massive multiple-input multiple-output (M-MIMO) [3], relay node (RN), Internet of Things (IoT) [4,5], heterogeneous network (HetNet) [6], mobile ad hoc networks (MANETs) [7,8], machine-to-machine (D2D) [9], power optimizations [10], handover processes [11], and interference cancellation [12].Some of the latest approaches, like Artificial Intelligence (AI), enabled micro base stations [13], machine learning [14], unmanned aerial vehicles (UAV) [15], blockchain [16], and human-centric communication [17] are potential concepts that can be used to design efficient net generation networks [18,19].However, some global areas require remote or temporary connectivity such as a terrestrial location where construction work is in progress, distant sports activity, indoor localization, remote health monitoring, or a war zone which requires communication devices for to-and-fro messaging, among others.For such scenarios, the utilization of UAVs can play an important role [20,21].The coverage and user capacity can be defined by the location, position, and altitude of the UAVs.
When discussing position tracking, we are usually asked why a Global Positioning System (GPS) cannot be used indoors [22].GPS is now widely used and recognized for outdoor location applications such as car navigation.In indoor settings, however, GPS technology has difficulties establishing a signal and maintaining accuracy [23].GPS cannot be utilized indoors due to poor signal strength and low accuracy.The GPS satellite's signal strength is weak, and after a long journey, the signal strength reaching the GPS receiver is significantly weaker, and barely strong enough to be useful [24].Any barrier in the line of sight between the antenna and the sky further weakens the signal.Walls usually reflect or obstruct GPS signals indoors, preventing them from penetrating the area.As a result, satellite signals cannot be properly picked up, and the room's poor signal makes it difficult to pinpoint one's location.While certain GPS devices can be put near a window to receive satellite signals, this is not always feasible or practicable in every building or indoor situation.In an open outdoor environment, GPS can reach 5−10 m precision, which is far from the half-meter accuracy required for many industrial use cases [25].Indoor precision has deteriorated even further.Both indoor location tracking and indoor navigation are in high demand for a wide range of applications.The UAV can surely provide a better solution for indoor localization.Figure 1 [26] shows the various UAV indoor location application scenarios.When discussing position tracking, we are usually asked why a Global Positioning System (GPS) cannot be used indoors [22].GPS is now widely used and recognized for outdoor location applications such as car navigation.In indoor settings, however, GPS technology has difficulties establishing a signal and maintaining accuracy [23].GPS cannot be utilized indoors due to poor signal strength and low accuracy.The GPS satellite's signal strength is weak, and after a long journey, the signal strength reaching the GPS receiver is significantly weaker, and barely strong enough to be useful [24].Any barrier in the line of sight between the antenna and the sky further weakens the signal.Walls usually reflect or obstruct GPS signals indoors, preventing them from penetrating the area.As a result, satellite signals cannot be properly picked up, and the room's poor signal makes it difficult to pinpoint one's location.While certain GPS devices can be put near a window to receive satellite signals, this is not always feasible or practicable in every building or indoor situation.In an open outdoor environment, GPS can reach 5−10 m precision, which is far from the half-meter accuracy required for many industrial use cases [25].Indoor precision has deteriorated even further.Both indoor location tracking and indoor navigation are in high demand for a wide range of applications.The UAV can surely provide a better solution for indoor localization.Figure 1 [26] shows the various UAV indoor location application scenarios.

Background
UAVs exist in many different forms and sizes.The sizes can range from several meters long to a few centimeters, and in terms of forms, they range from fixed-wing aircraft to blimps and multi-rotor UAVs [27,28].Fixed-wing aircrafts and blimps offer longer flight times, but multi-rotor UAVs have been popular in research because they offer higher maneuverability and control.Among the multi-rotor UAVs, a quadrotor configuration has been relatively more popular, with hexacopters and octocopters being other popular configurations [29].A study conducted by [30] surveys recent trends in UAV research in terms of the types of UAV systems, their applications in different research areas, trends in UAV research over the years, flight time, degree of autonomy, etc. UAVs can also be employed in the form of multi-UAV systems or swarms.For ease of deployment in most applications, some degree of autonomous operation of an UAV system is often desirable [31].Autonomy, in terms of robotics, is the ability of a system to operate without the intervention of a human operator, and the same definition applies in the context of

Background
UAVs exist in many different forms and sizes.The sizes can range from several meters long to a few centimeters, and in terms of forms, they range from fixed-wing aircraft to blimps and multi-rotor UAVs [27,28].Fixed-wing aircrafts and blimps offer longer flight times, but multi-rotor UAVs have been popular in research because they offer higher maneuverability and control.Among the multi-rotor UAVs, a quadrotor configuration has been relatively more popular, with hexacopters and octocopters being other popular configurations [29].A study conducted by [30] surveys recent trends in UAV research in terms of the types of UAV systems, their applications in different research areas, trends in UAV research over the years, flight time, degree of autonomy, etc. UAVs can also be employed in the form of multi-UAV systems or swarms.For ease of deployment in most applications, some degree of autonomous operation of an UAV system is often desirable [31].Autonomy, in terms of robotics, is the ability of a system to operate without the intervention of a human operator, and the same definition applies in the context of UAVs.Vision-based autonomous navigation and control strategies for autonomous operation are some very important aspects of UAV implementation [32].
UAVs have grown in popularity in various military and civilian sectors in recent years.They may work in hazardous settings where entry would be dangerous or impossible.Mini-UAV is a novel technology with the potential to be used for a range of activities such as precision farming, building maintenance, and surveillance missions [33].They are particularly intriguing to academics due to their tiny size, great agility, and inexpensive cost, in addition to being excellent for indoor use.Most UAVs employ a GPS module to establish their location; however, GPS does not work well indoors [34].If the mini-UAV cannot get location data, it will struggle to fly independently.
Signal processing has provided a set of tools that have been refined and utilized to great advantage over the last fifty years for UAVs [35].Different tools are used to solve diverse problems, and these tools are periodically combined to construct signal processing systems.Speech and audio, autonomous driving, picture processing, wearable technologies, and communication systems are all powered by signal processing.Currently, signal processing is influenced by deep learning [36].Deep learning for signal data requires extra steps compared to applying deep learning or machine learning (ML) to other data sets.Obtaining good-quality signal data is challenging because noise and fluctuation are used for UAV communication, especially in indoor environments.Most signal data include undesirable elements such as wideband noise, jitters, and distortions.
From the above-outlined introduction, this paper deals with a review of UAV localization in indoor environments.The summary of the contribution of this work is as follows:

•
It discusses the various existing surveys which work on indoor localization.

•
It examines several ML-based indoor localization approaches.

•
It explores various open issues and suggests future directions for ML-based indoor localization approaches.

An Indoor Localization Strategy for a Mini-UAV in the Presence of Obstacles
Long Cheng et al. [37] propose a new approach to mini-UAV localization in a wireless sensor network.To solve the problem of localization in non-line of sight (NLOS) environments, they present NLOS identification and a maximum joint probability algorithm.The proposed method requires only the RSS estimation model parameters to detect the propagation requirement and a particle swarm optimization (PSO)-based maximum joint probability procedure to assess the position of the mini-UAV.The outcome evidences the higher output rate for the NLOS environment.Furthermore, the highest joint probability approach based on PSO surpasses other techniques.

3.
2. An IMU/UWB/Vision-Based Extended Kalman Filter for Mini-UAV Localization in Indoor Environments Using 802.15.4aWireless Sensor Network Alessandro Benini et al. [38] describe a method for indoor localization of a mini-UAV using Ultra-Wide Band technology, a low-cost inertial measurement unit (IMU), and visionbased sensors.An extended Kalman filter (MA) is presented in this paper as a potential method for improving localization.The suggested method permits using a low-cost IMU in the estimated measure and built-in visual odometry to detect markers near the touchdown area.Ranging measurements allow for the reduction in inertial sensor errors caused by the limited performance of accelerometers and gyros.

Classification of Indoor Environments for IoT Applications: A ML Approach
Mohamad Ibrahim Alhajri et al. [39] present a ML approach for indoor environment classification based on real-time radio frequency (RF) signal measurements in a realistic setting.Various ML techniques like decision trees (DT)s, Support Vector Machine (SVM), and k-nearest neighbor (k-NN), were investigated using various RF features.The findings prove that a ML approach based on weighted k-NN, channel transfer function (CTF), and frequency coherence function (FCF) surpasses other techniques in detecting further indoor environment types with a 99.3% classification accuracy.The estimated time is set to less than 10 s, showing that the applied method is feasible for real-time implementation developments.Their study's goal was to outline the process and underline the advantages of ML as cutting-edge technology and a useful tool for classifying indoor settings.

Mini-UAVs Detection by Radar
Miroslav Kratky and L. Fuxa [40] present possible methods for UAV recognition, especially within the radar frequency spectrum.The detection range of sensors will only gradually and insufficiently expand in the future, necessitating the development of alternative solutions.According to this paper, one option is to link them into a linked, complex surveillance system.Their common interconnectedness, interoperability, and modularity should result in synergistic effects such as increased detection probability and decreased false alarms.When using radar to detect Low, Small, and Slow (LSS) targets, the target's radar cross-section (RCS) is the limiting factor.The carrier frequency f, or wavelength of radar, is its limitation.Additional internal and external influences include specific radar technical solutions, tactical employment within the terrain and combat formation, atmospheric conditions, and crew proficiency, among others.The capacity of radar systems to identify small, sluggish, and low-flying targets has been verified, which is the main contribution of this research to military science.

Indoor UAV Localization Using a Tether
Xuesu Xiao et al. [41] present an approach to localizing an UAV in indoor environments using only a quasi-taut tether.They propose a new sensor modality for tethered UAV indoor localization that uses tether-based feedback instead of GPS, inertia, and vision-based sensing.This localizer uses tether sensory information, including tether length, elevation, and azimuth angles, and is based on the transformation of polar to Cartesian coordinates.The authors show in Figure 2 that, when the tether is long and dragged down by gravity, it forms an arc instead of an ideal straight line.A mechanics model is built to quantify the inevitable tether deformation.This model can correct the calculated altitude angle and tether length, enhancing localization precision.Tests demonstrate enhanced localization precision on the Fotokite Pro, a physically tethered UAV.The findings demonstrate that the model can successfully reduce the detrimental effect of increased tether length on localization outcomes and boost localization accuracy by 31.12%.The floating strength tolerance of the particular UAV platform determines how much this proposed strategy can tolerate hovering stability mistakes.
prove that a ML approach based on weighted k-NN, channel transfer function (CTF), and frequency coherence function (FCF) surpasses other techniques in detecting further indoor environment types with a 99.3% classification accuracy.The estimated time is set to less than 10 s, showing that the applied method is feasible for real-time implementation developments.Their study's goal was to outline the process and underline the advantages of ML as cutting-edge technology and a useful tool for classifying indoor settings.

Mini-UAVs Detection by Radar
Miroslav Kratky and L. Fuxa [40] present possible methods for UAV recognition, especially within the radar frequency spectrum.The detection range of sensors will only gradually and insufficiently expand in the future, necessitating the development of alternative solutions.According to this paper, one option is to link them into a linked, complex surveillance system.Their common interconnectedness, interoperability, and modularity should result in synergistic effects such as increased detection probability and decreased false alarms.When using radar to detect Low, Small, and Slow (LSS) targets, the target's radar cross-section (RCS) is the limiting factor.The carrier frequency f, or wavelength of radar, is its limitation.Additional internal and external influences include specific radar technical solutions, tactical employment within the terrain and combat formation, atmospheric conditions, and crew proficiency, among others.The capacity of radar systems to identify small, sluggish, and low-flying targets has been verified, which is the main contribution of this research to military science.

Indoor UAV Localization Using a Tether
Xuesu Xiao et al. [41] present an approach to localizing an UAV in indoor environments using only a quasi-taut tether.They propose a new sensor modality for tethered UAV indoor localization that uses tether-based feedback instead of GPS, inertia, and vision-based sensing.This localizer uses tether sensory information, including tether length, elevation, and azimuth angles, and is based on the transformation of polar to Cartesian coordinates.The authors show in Figure 2 that, when the tether is long and dragged down by gravity, it forms an arc instead of an ideal straight line.A mechanics model is built to quantify the inevitable tether deformation.This model can correct the calculated altitude angle and tether length, enhancing localization precision.Tests demonstrate enhanced localization precision on the Fotokite Pro, a physically tethered UAV.The findings demonstrate that the model can successfully reduce the detrimental effect of increased tether length on localization outcomes and boost localization accuracy by 31.12%.The floating strength tolerance of the particular UAV platform determines how much this proposed strategy can tolerate hovering stability mistakes.

Multi-Ray Modeling of Ultrasonic Sensors and Application for Micro-UAV Localization in Indoor Environments
Lingyu Yang et al. [42] suggest a method that is based on an IMU and four Ultrasonic sensors.It is suitable for use in a light Micro Air Vehicle (MAV) because it accurately approximates a beam pattern while maintaining low computational complexity.An EKF is utilized and the IMU, after a quick approach, is described for constructing the Jacobian matrix of the measurement function.The model's accuracy is tested using a MaxSonar MB1222 sensor, and a simulation and experiment are run using the Thales II MAV platform.To achieve higher precision positioning, sonar and IMU sensor measurements are fused.The jump filter is used to suppress abnormal and significant differences between estimates and measurements.The proposed methods are validated using simulations, and the findings show that the model has a localization precision of about 20 cm.The findings demonstrate that its positioning precision exceeds 20 cm, and its computing complexity is sufficiently low to run on the stm32 platform.An investigation conducted with an unmodeled obstacle indicates that the suggested method's great robustness does not affect the findings.

Indoor Positioning Using Bluetooth Technology
By utilizing the RSSI, the authors have suggested a low-cost indoor positioning system (IPS) for UAVs that is based on Bluetooth low energy (BLE) beacons [43].BLE is a lowpower technology designed to send sparse quantities of data.The relationship between the RSSI and the distance between two or three Bluetooth devices, an onboard receiver, and transmitters placed in the interior operating field is examined using a mathematical model.This project's experimental findings and system performance analysis demonstrate the viability of its methods.The initial examination is to validate the device's attributes and operation and the correctness of the approaches used.

UAV Localization Using Ultra-Wideband (UWB)
The design and assessment of a realistic and collaborative UWB positioning system by employing released integrating radio frequency devices and antennas are covered in the study of [44].The UAV equipment uses GNSS emulation's signals to maintain its location.In addition to other aspects like antenna characteristics, constellation-aware parameters have been considered.A non-line-of-sight rejection has been implemented based on the relationship between the initial path and the strength of the aggregated channel impulse output.In order to get a large sample set of findings to evaluate the system's accuracy in actual usage, an experiment using a variety of locations and orientations is carried out.In a preliminary experiment, the function of the model achieves a root-mean-square error with a probability of 95% of less than 10 cm in the horizontal plane and less than 20 cm in 3D space.A GNSS emulation system is installed on an exploratory UAV carrier to evaluate the real-time inflight deployment of the UWB locating mechanism.It is a proof of concept that the GNSS emulation might be used with commercially available UAV devices to provide those systems with the ability to navigate indoors.In order to enhance flying performance, additional study is needed to enhance the processing for the UAV-specific navigational controller and to address magnetometer problems indoors, maybe with the use of integrated measurement unit (IMU) sensor fusion.

Magnetic Field Measurements Based Indoor Positioning
The authors in [45] offer an interior location system for an UAV in their article.Magnetic field observations are the primary source of data needed to determine where the platform is located.This is a parameter that is both cheap to measure and simple to get.A versatile measuring system has emerged as a result of this approach.The measuring system comprises three-axis electronic magnetometers, a battery-powered microprocessor system with an SD memory card, and an LCD display.The validation experiments were carried out in a dedicated room, and permanent magnets were used to modify the local magnetic field, which could act as beacons.

Interference and Energy-Based Approach for UAV Localization Learning Techniques
The authors in [46] design a prediction-based proactive drone management framework to decrease network interference and enhance energy efficiency in the multiple drone small cells (DSC) scenarios.The system proactively determines whether a DSC should be awake or asleep due to the predicted user positions at the next timeslot.The RF-based mobility prediction model with a high accuracy of 93.14% is built from a small sample data in the offline phase.In the online phase, the wake/sleeping schemes of the DSCs are proactively determined according to the predicted user positions.This study only addresses DSC-touser access links' interference and energy consumption problems.
Table 1 summarizes the comparison, with advantages, disadvantages and limitations, of the current existing studies on indoor localization.The faster UAV can be detected by using high-frequency-based radar [40] Indoor UAV Localization Using a Tether predication time has been reduced in the real environment Power optimization approach required to be in more sophisticated indoor radio channel [41] Multi-Ray Modelling of Ultrasonic Sensors and Application for Micro-UAV Localization in Indoor Environments The proposed approach is efficient for localizing the UAV in a moving frame Tracking accuracy can be enhanced by using the ML approach [42] Indoor Positioning using Bluetooth Technology Kalman filtering is applied to improve collected data from noise, drift, and bias errors It must focus on outdoor tests for a safe landing area determination system [43] UAV localization using Ultra-wideband (UWB) Better probability positioning accuracy using the global navigation satellite system (GNSS) It required optimizing the filtering for the UAV-specific navigation controller and magnetometer issues indoors [44] Magnetic field measurements based on Indoor Positioning Higher attainable accuracy results as compared to infrared or sound waves based positioning systems The impact examination of powered and operating electronic devices needs to be explored [45] Interference and energy-based approach for UAV localization Learning Techniques Interference and enhance energy efficiency in the multiple drone small cells (DSC) scenarios Computational complexity is increasing, which needs to mitigate for higher interference scenarios [46]

Machine Learning-Based Indoor Localization
Several studies provide a thorough examination of ML-enabled localization methods using the most popular wireless technologies [47].In addition, several authors discuss various ML techniques (supervised and unsupervised) that could effectively address indoor localization challenges, such as NLOS, device heterogeneity, and environmental variations, with sufficient difficulty [48].These authors also discuss various ML techniques, supervised and unsupervised, that could relieve various indoor localization challenges to achieve a comprehensive indoor positioning system (IPS) [49].Therefore, the following sub-sections will discuss various ML models focusing on several learning-based approaches.

k-Nearest Neighbor (k-NN)
The k-nearest neighbor algorithm, also called the k-NN, uses distance to classify or predict the grouping of individual data points using supervised learning methods [50].With k-NN, all computation is delayed until the function is evaluated and locally approximated.If the features reflect distinct physical units or have mostly changed sizes, regulating the training data can drastically improve the precision of this method which utilizes distance for categorization [51].Weighting neighbor contributions might be a beneficial strategy for both classification and regression, enabling nearby neighbors to contribute more to the average than remote neighbors.The k-NN is a fundamental ML technique based on a supervised learning approach that assumes that the new and current instances are comparable and allocates the new instance to the category closest to the existing categories [52].After saving all prior data, new data points are classified using the k-NN algorithm based on similarity.This suggests that when new data is received, it may be swiftly sorted using the k-NN approach into the proper suite category.Although the k-NN approach may be used for both classification and regression problems, classification jobs are where it is very regularly employed [53].The k-NN does not make assertions about the distribution data because it is non-parametric.Since it does not immediately begin learning from the training dataset, this method is also referred to as a lazy learner algorithm [54].Instead, it makes use of the dataset to carry out an action when categorizing data.The k-NN technique keeps the data throughout the training phase and classifies fresh data into a category that is very near to the new data received.

Support Vector Machine (SVM)
The SVM is an intelligent, fast, and highly adaptable ML algorithm that can be used for regressions and classifications (linear and nonlinear ML) and for detecting outliers [55].It is one of the most well-liked ML models due to these features [56].SVMs are divided into support vector classification and support vector regression based on supervised ML.Face recognition, text and hypertext categorization, image classification, bioinformatics, protein fold, distant homology, and handwriting detection are just a few applications for SVMs [57].The SVM algorithm finds an N-dimensional hyperplane that splits the input points into distinct categories.In addition to regression, classification, and outlier detection, SVM is a supervised learning algorithm that performs effectively in high-dimensional environments [58].When there are more dimensions than samples, this method is also effective.It is also memory-efficient since the decision function only uses a subset of the training points.Regardless, SVMs employing kernel functions avoid over-fitting if the number of features is significantly more than the number of samples and the regularization term is crucial [59].Furthermore, SVMs do not yield probability estimates immediately; instead, they need a time-consuming five-fold cross-validation procedure.

Decision Tree
A decision tree model is an approach used in ML that presents possibilities based on the characteristics of the input [60].It follows the "branch node theory", according to which each branch stands for both a decision and a variable.Decision tree model methods come in different varieties.Some of these algorithms have been applied to categorize radiological pictures, and as a result, they may be found in radiology or radiology-related computer science articles.Decision tree models, unlike most other ML algorithms, apply principles and are hence understandable [61].Instead of diagrams of linked nodes, decision tree models can be shown using partitioned graphs and perspective plots.A decision tree is a tree-like diagram in which the tree trunk represents the internal nodes that describe a test of a certain feature, the branches reflect the test results, and the leaves indicate the nodes' categorization marks.In the decision tree approach, categorization aims to organize the data in a way that includes both the root node and the leaf node [62].Decision trees can analyze data to find key system features that point to potentially dangerous behavior.Consequently, evaluating the arrangement of intrusion identification information increases the value of distinct security frameworks.The growth of attack signatures, various checking activities, and patterns and instances that stimulate checking may all be recognized [63].The usage of decision trees differs from other methods in that it offers a complete set of rules that are straightforward to put into practice and that are easily connected with real-time technology.
A DT represents a set of conditions which are hierarchically organized and successively applied from the root to the leaf of the tree.DTs are easy to interpret, and their structure is transparent [64].DTs produce a trained model that can represent logical rules, and the model is used to predict new datasets through the repetitive process of splitting.In a decision tree method, features of data are referred to as predictor variables, whereas the class to be mapped is the target variable.For regression problems, the target variables are continuous.For an UAV localization application, a simple example of a decision tree to predict 3D location is depicted in Figure 3.
immediately; instead, they need a time-consuming five-fold cross-validation procedure.

Decision Tree
A decision tree model is an approach used in ML that presents possibilities based on the characteristics of the input [60].It follows the "branch node theory", according to which each branch stands for both a decision and a variable.Decision tree model methods come in different varieties.Some of these algorithms have been applied to categorize radiological pictures, and as a result, they may be found in radiology or radiology-related computer science articles.Decision tree models, unlike most other ML algorithms, apply principles and are hence understandable [61].Instead of diagrams of linked nodes, decision tree models can be shown using partitioned graphs and perspective plots.A decision tree is a tree-like diagram in which the tree trunk represents the internal nodes that describe a test of a certain feature, the branches reflect the test results, and the leaves indicate the nodes' categorization marks.In the decision tree approach, categorization aims to organize the data in a way that includes both the root node and the leaf node [62].Decision trees can analyze data to find key system features that point to potentially dangerous behavior.Consequently, evaluating the arrangement of intrusion identification information increases the value of distinct security frameworks.The growth of attack signatures, various checking activities, and patterns and instances that stimulate checking may all be recognized [63].The usage of decision trees differs from other methods in that it offers a complete set of rules that are straightforward to put into practice and that are easily connected with real-time technology.
A DT represents a set of conditions which are hierarchically organized and successively applied from the root to the leaf of the tree.DTs are easy to interpret, and their structure is transparent [64].DTs produce a trained model that can represent logical rules, and the model is used to predict new datasets through the repetitive process of splitting.In a decision tree method, features of data are referred to as predictor variables, whereas the class to be mapped is the target variable.For regression problems, the target variables are continuous.For an UAV localization application, a simple example of a decision tree to predict 3D location is depicted in Figure 3.

Extra-Tree
Extra-Trees, or extensively randomized trees, is an ensemble learning approach.This approach creates a collection of decision trees [65].When constructing the tree, the

Extra-Tree
Extra-Trees, or extensively randomized trees, is an ensemble learning approach.This approach creates a collection of decision trees [65].When constructing the tree, the decision rule is chosen at random.This technique and random forest are quite similar, except for the randomly selected split values.In order to produce a combination of the unpruned decision or regression trees, the Extra-Trees methodology uses the usual top-down construction process [66].It divides nodes by choosing cut-points randomly and builds the trees using the whole learning sample, both of which are key differences from earlier treebased ensemble techniques.In addition, the Extra-Trees splitting method can be used for numerical features.This method has two parameters: the minimum sample size for splitting a node and the number of randomly selected attributes for each node.These parameters are repeatedly mixed with the initial learning sample to create an ensemble model [67].In classification problems, the final prediction is produced by the majority vote, whereas in regression problems, it is produced by the arithmetic average.

Random Forest
Using decision tree methods, a supervised ML methodology known as a random forest is developed [68].This strategy is used to anticipate behaviors and outcomes in a wide range of industries, including banking and e-commerce [69].This method employs ensemble learning, which is a method for mixing several classifiers to tackle challenging issues.There are numerous possible decision trees in a random forest algorithm.The random forest algorithm "forest" is trained using either bagging or bootstrap aggregation [70].The ensemble meta-algorithm bagging improves the ML systems' precision.Based on the predictions provided by the decision trees, the algorithm chooses the result.By averaging or averaging the output of multiple trees, it generates predictions.As the number of trees increases, the output's accuracy increases.Accuracy is increased while dataset overfitting is decreased.This method produces predictions without needing a complex set of package settings.Compared to the decision tree technique, it is more accurate and handles missing data well [71].Without the requirement for hyper-parameter adjusting, this method can deliver a decent forecast.Moreover, random forest addresses the decision tree generalization issue.Each random forest tree selects a sample of features at random at the node's splitting point.Before training, three critical hyperparameters for random forest algorithms must be set [72].Among these are node size, tree count, and sampled feature count.The random forest classifier may then be used to tackle classification or regression problems.

Neural Network (NN)
A neural network is a circuit or network of artificial neural networks (ANNs) made up of artificial neurons or nodes [73].ANNs are also used to address difficulties in AI [74].ANNs are a collection of algorithms that use a technique inspired by the way the human brain functions to find undiscovered connections in a batch of data.In order to uncover relationships among huge amounts of data, neural networks are a group of algorithms that simulate the functioning of a brain [75].As a result, they are frequently able to mimic synapses and neural connections seen in the brain.Applications utilized in the financial sector include forecasting, market research, fraud detection, and risk assessment.Deep learning methods use multiple-process-layer neural networks, sometimes referred to as "deep" networks [76].To produce the final output (last layer), the input data (first layer) is processed against a hidden layer (middle layer).It is possible to have several hidden layers.Figure 4 [77] shows a simplified view of a feed-forward artificial neural network.

Feed-Forward Neural Network (FFNN)
A feed-forward neural network (FFNN) is a kind of artificial neural network in which there are no cycles in the connections between the nodes [78].Numerous linked neurons make up an ANN.Each neuron receives a set of floating-point numbers and multiplies them by a set of weights, which are also floating-point numbers [79].The weights serve as a method to emphasize or disregard particular inputs.
Table 2 shows the summary along with the advantages and disadvantages of various above studies wherein ML algorithms are used for UAV indoor localization.

Feed-Forward Neural Network (FFNN)
A feed-forward neural network (FFNN) is a kind of artificial neural network in which there are no cycles in the connections between the nodes [78].Numerous linked neurons make up an ANN.Each neuron receives a set of floating-point numbers and multiplies them by a set of weights, which are also floating-point numbers [79].The weights serve as a method to emphasize or disregard particular inputs.
Table 2 shows the summary along with the advantages and disadvantages of various above studies wherein ML algorithms are used for UAV indoor localization.Missing data in the collection do not have any impact.

•
A small change in the data can significantly alter the structure of the decision tree.

•
Calculations can eventually get very complicated.

•
It takes longer to train the model.

• Expensive
Extra Tree

•
Able to monitor variables that are both continuous and categorical

•
Crediting the missing data is not essential.

•
Less coding and comparison are needed during the pre-processing processes.

• It takes up memory •
Takes more time • A small modification in the data can significantly impact the tree structure.

•
The complexity of space and time is more complicated.
Random Forest • Provide extra time for convergence • The gradient issue of vanishing and explosion.

Performance Evaluation Metrics for ML Models
This section briefly describes the performance evaluation metric for machine learning models in indoor UAV localization systems [80].The performance evaluation metric of machine learning is divided into regression and classification, as described below: Error localization refers to the process of (automatically) identifying the fields in an edit-failed record that need to be imputed.To ensure the final (corrected) record will not fail to edit, a minimum set of fields is imputed using an optimization algorithm [81].

Mean Squared Error
The mean squared error, or MSE, measures the accuracy of statistical models.It assesses the average squared variation between the values anticipated and observed.When a model is error-free, the MSE is equal to 0. The worth increases as model inaccuracy grows.The mean squared error is also known as the mean squared deviation (MSD).The MSE, or the second moment of the error, includes both the variance of the estimator, which shows how widely spread guesses are from one data sample to the next, and the difference between the approximate average value and the true value.The units of measurement for MSE and the square of the value being assessed are the same [82].

Root Mean Square Error
Root mean square error (RMSE) is the residuals' standard deviation.RMSE is an assessment of how to spread out the residuals, which are a measure of how distant the data points are from the regression line.RMSE illustrates the Euclidean distance between measured true values and forecasts.Determine the difference between the prediction and the actual value for each data point, the residual norm for every data point, the mean of the residuals, and the square root of those means to determine the RMSE.It is very beneficial to have a single number when evaluating a model's performance in machine learning, whether at training, cross-validation, or maintaining after deployment.One of the most used metrics for this is the root mean square error [83].

R-Squared
The R-Squared statistic indicates how much variation in a dependent variable is described by one or several independent variables within a regression model.In the context of investing, R-squared is commonly regarded as the proportion of a fund or security's motions that changes in a benchmark index can explain.An R-squared of 100% means that changes in the index accurately describe all changes in security [84].

Accuracy
In science and engineering, accuracy refers to the degree to which the dimensions of a quantity are near the actual value of that number.Accuracy is the difference between a set of measurements (observations) and the actual value.Therefore, accuracy has a value between 0 and 1.The accuracy measure falls short when working with unbalanced data and models that provide a probability score.Inadequate precision results in a discrepancy between the result and the real value.High precision and trueness are required for greater accuracy [85].

Precision and Recall
Precision is the degree to which the measurements of corresponding objects agree with one another.The degree to which repeated data points under the same conditions yield the same findings represents the accuracy of a measuring system, which is linked to reproducibility and repeatability.Accuracy is not necessary for precision.In other words, scientific observations can be appropriate without being exact, and they can be highly precise but not remarkably accurate.The highest caliber scientific observations are exact and accurate [86].

F1-Score
The F1-score metric employs a combination of precision and recall.In other words, the F1 score is the harmonic mean of the precision and recall values [87].

Confusion Matrix
A confusion matrix is a tabular visualization of the ground-truth labels versus model predictions.Each row of the confusion matrix describes the instances in a predicted class and each column describes the instances in an actual class [88].

AUROC (Area under Receiver Operating Characteristics Curve)
AUROC is a combined measure of sensitivity and specificity.The AUROC is a measure of the overall performance of a diagnostic test and is interpreted as the average sensitivity value for all possible specificity values [89].
Table 3 describes the performance evaluation metric for machine learning models.Localization algorithms can be classified as range-free or range-based.In range-based localization, TDOA, TOA, TOF, TOA, RSSI, and CSI are used as the distance measurement technologies [90][91][92].Though TOA, TDOA, and AOA provide high accuracy, they require complex hardware arrangements to measure, while CSI and RSSI require simplified hardware setup with good accuracy [93].In addition, the performance may suffer from strong multipath and NLOS propagation in urban scenarios.Further, RSSI has been recommended for many localization systems [94].Proposed UAV indoor localization algorithms can also be categorized as deterministic, probabilistic, filter-based, or ML-based algorithms [95,96].Proposed location estimation algorithms based on filters are highly mathematical and mostly impractical in implementation on real hardware devices.Related works prove that ML-based algorithms provide a promising localization performance in terms of localization accuracy [97].Moreover, ML/DL-based systems are easy to deploy on edge devices with clouds.Furthermore, for massive data volumes of multi-story buildings, the deep learning neural network (DNN) emerges in a modernist way [98,99].DNN can function well with fewer training data dimensions and draw out more useful features from subsequent samples.Statistical and empirical methodologies will provide extremely useful guidance on the various discovering techniques for indoor location study [100].Localization strategies will face challenges with protocols, delays, and radio waves.Additionally, the performance of algorithms and the variety of applications both affect accuracy.The research trend for location-aware computing and navigation path prediction is still reaffirmed [101,102].

Wireless Technologies and Hardware Design for Future UAV Localization Systems
Designing concerns for hardware setups for ML algorithm-based UAV localization systems for disaster management applications is critical [103].The system's wireless technology should be carefully deiced considering the sensing range, and power consumption, among others [104].Wireless technologies UWB, LTE, Bluetooth, LoRA, or ZigBee can be proposed along with ML approaches [105,106].Since the wide bandwidth makes detecting the time-delayed versions of the transmitted signal easier, UWB achieves excellent multipath resistance and good material penetrability [107].Since the introduction of the Bluetooth 4.0 standard protocol, Bluetooth, another wireless technology standard for sharing data over short distances, has grown in popularity.A version of Bluetooth called Bluetooth low energy (BLE) is designed for low-power applications and enables some applications to run continuously for several months [108].However, BLE is not recommended for UAV localization due to its short-range communication.Further, when deploying many nodes in the network multipath, channel fading and co-channel interferences may occur [108].Therefore, applying co-channel interference mitigating techniques is essential [109].
Since the nodes are being deployed in a robust environment, power management is critical in these hardware systems when applying ML approaches [110,111].This ensures that the localization system with a deep reinforcement learning approach continues to function normally even if some measuring units may be destroyed or malfunction due to the harsh environment, or run out of energy [112].Computing the location placement technique with high robustness could work even when some signals are not accessible; positioning approaches must employ this incomplete information.To prolong the battery life, power management techniques such as deep sleep modes could apply [113].

Signal Conditioning Techniques for Future UAV Localization Systems
Ranging measurements, including RSSI, fluctuate highly due to the multipath environment [114].Therefore, strong signal processing techniques should be applied before using ML algorithms to train [115].Related works have proposed many linear and nonlinear filters, such as moving average, gaussian, particle, and Kalman filters for smoothing the signals [116].Similar to the moving average, the filters are easy to implement on hardware setups, however, the Kalman filter shows impractical implementation due to its complexity [117].The cost function is proposed for the linearization maximum likelihood algorithm with a mean square error [118].

Privacy Concerns and Security in Future UAV Localization Systems
In some applications, UAV calls for recording user preferences, activity history, present location, and prior moves [119].The development, adoption, and expansion of UAV applications may be severely constrained by the risks connected with the violation of location privacy when applying ML approaches [120].Additionally, UAVs require that the user reveals their location to enable personalization [121].Location data may be kept, used, and sold by service providers.These possible hazards may deter users.Unrestricted access to a person's location information could result in dangerous interactions.
Network attack is a high threat in UAV localization systems [122].Even if secure communication is carried out between anchor nodes or UAVs and user devices, attacks may still be possible when users' devices communicate with service servers [123].The attacker will be able to determine UAV location information if the attacker receives communication from the user and service server.Therefore, the service provider should implement a man-in-the-middle attack protection method (MITM) [124].MITM implements protection measures using mutual authentication methods like public key infrastructures (PKI) or stronger mutual authentication methods like secret keys or passwords [125].

Conclusions
The detailed description of radio wave signals for indoor positioning, based on widely used technologies and efficient positioning techniques, was reviewed in this study.The introduction outlined how non-radio wave transmissions behave.As previously indicated, the current positioning algorithms have dealt with the inaccurate positioning problem brought on by the signal variation caused by multipath propagation, the hardware, and software complexity, real-time processing, and the dynamically changing environment.Several methods for localization of mini-UAVs in indoor environments are based on ML approaches, where each technology comes with its own unique set of limitations.Compared to the traditional localization algorithms proposed, ML algorithms are more accurate, less complex, and easier to deploy on real hardware devices (edge computing).Consequently, choosing a suitable remedy in a unique circumstance is the best course of action.In certain ways, developing low-cost positioning technologies to improve positioning accuracy is severely constrained due to the high expense of high-precision indoor positioning technology, and necessitates additional auxiliary equipment or a great deal of simulated processing.Localization strategies will face challenges with protocols, delays, and radio waves.Additionally, the performance of algorithms and the variety of applications both affect accuracy.The research trend for location-aware computing and navigation path prediction is still reaffirmed.Integrating and using various technologies will be the future trend to achieve complementary advantages.Furthermore, channel fading and co-channel interferences may occur when deploying many nodes in the network multipath.Therefore, applying co-channel interference mitigating techniques is essential.In general, the less expensive the technology, the more accurate it is, and the easier it is to popularize.Signal core features are extracted from signals using classifier algorithms in localization.ML for signal processing is a better technology that can be used to detect the position of an UAV.Integration of a range finder could significantly improve overall performance in indoor environments.

Figure 2 .
Figure 2. UAV flying in a motion capture studio with a tether dragged down by gravity [41].

Figure 4 .
Figure 4.A simplified view of a feedforward artificial neural network [77].

Figure 4 .
Figure 4.A simplified view of a feedforward artificial neural network [77].

Table 1 .
Summary of existing studies on indoor localization.

Table 2 .
ML algorithms used for UAV indoor localization.

Table 3 .
Performance Evaluation of Machine Learning models.