The widespread deployment of Wi-Fi access points (APs) provides an appealing candidate for indoor positioning. However, the major drawback to Wi-Fi-based positioning is that utilizing the signal faces several challenges in a dynamic environment. On the other hand, magnetic fields provide long-term stability in an indoor environment. Similarly, the available Wi-Fi APs can supplement the low number of magnetic field elements present in an indoor environment. Therefore, the hybrid use of Wi-Fi and magnetic field data provides several unique characteristics to compensate for the limitations encountered when each is used independently. In this paper, we propose applying the long short-term memory (LSTM) model to the spatial information from Wi-Fi and magnetic fields due to its advantages in time-series prediction and characterization for region classification. The results demonstrate that the proposed approach can perform indoor region classification with each of the values for precision, recall, and F1 scoring above 95.0%.
With the growth of mobile computing and Internet of Things technologies, location-based
services (LBS) are becoming an integral part of our lives both indoors and outdoors in fields such
as emergency response services, staf management, and vehicle tracking [
Creating a reliable and accurate LBS is often crucial for applications focused on indoor areas.
Further, the exact needs will depend on the requirements of a particular application. For example,
smart building applications often require the ability to distinguish diferent workplaces. In some
applications, such as emergency response services, locating a mobile device in a subregion [
To date, there has been extensive investigation conducted into indoor positioning systems (IPS) (e.g., Wi-Fi, Bluetooth Low Energy (BLE), ultra-wideband (UWB) and radio frequency identification (RFID)). These technologies can be categorized into infrastructure-based and infrastructure-free approaches. The former approach needs hardware and software to execute IPS, such as Wi-Fi, UWB, and BLE. The latter systems employ commonly available positioning technologies and can work without extra infrastructure (e.g., Pedestrian dead reckoning (PDR), magnetic fields). Infrastructure-based methods provide higher accuracy but are costly due to the infrastructure requirements. In contrast, an infrastructure-free solution provides pervasive positioning and is less expensive to implement and maintain.
Recently, localization methods using geomagnetic sensors in smartphone-based have piqued
the interest of many researchers owing to the pervasiveness of magnetic fields in indoor
environments and their independence from external infrastructure [
This paper proposes LBS based on indoor region classification utilizing Wi-Fi received signal strength (RSS) and magnetic data. We present the long short-term memory (LSTM) model by taking advantage of time-series characteristics for classification. The experimental result is evaluated with the classical machine learning (ML) and deep learning (DL) algorithms to compare the performance metrics such as precision, recall, and F1 score.
The remainder of the paper is organized as follows. In section 2, recent literature on smartphone general localization technologies and fusion method is introduced. The proposed system is presented in 3. In section 4, we present the experimental setup and the discussion of the results. Finally 5 concludes this paper.
Modern smartphones are embedded with various sensors, making them both communication
tools and sensing equipment [
In recent years, various smartphone-based IPS has been investigated. Among various
localization technologies, localization methods based on wireless signal (Wi-Fi and BLE) are the
most popular approach owing to low cost of infrastructure. The research [
The development of internal sensors in smartphones (e.g., acceleromete, gyroscope and
magnetometer) improves IPS accuracy. For example, JustWalk [
Many fusion-based localization technologies have been developed by combining two or more
technologies. For example, IndoorWaze [
The motivation for combining infrastructure-based (Wi-Fi) and infrastructure-less (magnetic
ifeld) technologies for indoor region classification is that combining the two can compensate
for the drawbacks of each and provide complementary information to improve localization
[
For example, Magicol [
Using a Wi-Fi based approach alone sufers from signal attenuation in a dynamic environment, especially when faced with human mobility and occupancy. However, human mobility and occupancy have little efect on magnetic fields. Similarly, because data from smartphone sensors include noise and variation, the magnetic field alone provides poor accuracy due to handshaking and the user’s movement. The drawbacks can be compensated by combining it with the Wi-Fi approach. Furthermore, there are almost always Wi-Fi access points (APs) and the magnetic ifeld generated by many sources in a public indoor environment. Therefore, we can exploit these two abundant resources in public places for localization. 135272251893510504 ....... ) .36 (ym .50 .......672756358554495544 0.45 0.9 1.35 1.8 2.25 2.7 3.15 x3(m.6) 4.05 4.5 4.95 5.4 5.85 6.3 6.75 (a) 44 .540 9 . 46 .530 811 . 48 .252 35172 . 50 ) ..36 52 (ym ..50 55449 . 54 .454 558 . 56 .635 76 . 58 .72 0.45 0.9 1.35 1.8 2.25 2.7 3.15 x3(m.6) 4.05 4.5 4.95 5.4 5.85 6.3 6.75 (b) 54 . 40 .900 315 . 45 .185 22 . 50 ..257
13 55 ()ym ..5360 454 . 60 .954 54 . 65 .585 6375 . 70 ..62 7 0.45 0.9 1.35 1.8 2.25 2.7 3.15 x3(m.6) 4.05 4.5 4.95 5.4 5.85 6.3 6.75 (c)
The rapid change in Wi-Fi RSS has a substantial impact on localization accuracy and is one of the challenges of Wi-Fi-based IPS. In contrast to Wi-Fi, magnetic fields in indoor environments exhibit long-term stability. When used in a dynamic environment with human mobility and occupancy, Wi-Fi-based techniques sufer from a significant loss in positioning accuracy because of variations in RSS over time. In contrast, human motion has a negligible efect on magnetic ifeld measurements due to the absence of significant ferromagnetic elements in humans.
Fig. 1 and Fig. 2 present the heatmap for data collected in the selected scenario. The RSS and magnetic field information provide a distinctive feature and unique pattern in an indoor environment, as shown in Fig. 1 and 2. Fig. 3 shows the magnetometer values from the , , and -axis collected by a smartphone along the 100-meter corridor. We can see from Fig. 3 that the magnetic field varies from location to location, but the changes are not significant. The magnetic field’s and components are pointed towards the north and east, while the component is pointed vertically towards the earth. Therefore, the normalized magnitude of the magnetometer is calculated as follows:
Given the widespread use of Wi-Fi and the pervasiveness of magnetic fields in an indoor environment, we may be able to acquire both signals concurrently in many places. Anomalies in the geomagnetic field induced by local disturbances caused by ferromagnetic construction materials can be employed to enable pervasive positioning technology for an indoor environment that is not reliant on infrastructure.
The architecture of the proposed system is depicted in Fig. 4. First, the corridor area is classified into 15 diferent regions. Then, RSS values from numerous APs and magnetic field data for each region are collected to create a dataset for training the model.
During the data collection, we recorded the RSS and magnetic data simultaneously for
each subregion in the corridor. The corresponding RSS reading is set to -200 dBm in the
Android program if the RSS for a specific AP is not detected. We utilize linear [
This paper uses LSTMs for Wi-Fi and magnetic datasets for indoor subarea classification based on time series data. LSTM is a form of recurrent neural network (RNN) designed to avoid
the problem of long-term dependency. Unlike RNN, which exhibits gradient exploding when dealing with a long-term time series, LSTM employs memory cells and a gated approach to solve the problem.
Each LSTM unit manages the learning and forgetting of timing information via a memory cell and many non-linear gating units. Fig. 5 depicts the structure of an LSTM unit. The three gates in LSTM (i.e., forget, input and output) manage the memory and performing the update for cell state information. Given an input of forget gate is ℎ− 1 and , a standard LSTM can be formulated as follows:
= ( · [ℎ− 1, ] + ) where , denotes the output, weight and bias of forget gate. Sigmoid is the activation function which responsible for determining which values to let through (0 or 1) and can be given as follows: =
1 1 + − The second layer is divided into two sections (i.e., sigmoid and ℎ function). During the first part, an input gate is updated as follows: (2) (3)
c t Next cell state
h t Next hidden state
In this paper, the LSTM networks are employed for fusing Wi-Fi and magnetic fields based on the time-series feature. Our final model integrates elements from LSTM and dense neural networks based on hyperparameter selection. We use a three-layer network comprising two The second part generates a candidate state value by ℎ activation layer as The ℎ activation assigns weightage to the values, determining their level of relevance (-1 to 1), and is formulated as follows: The next cell state is determined as follows: The last stage is to determine the output. First, a sigmoid layer is run to identify which elements of the cell state are sent to the output. The cell state is then passed via the ℎ function to shift the values between -1 and 1 multiplied by the sigmoid gate output. The process can be summarized as follows:
= ( · [ℎ− 1, ] + ) = ℎ( · [ℎ− 1, ] + ) ℎ() = − − + − = * − 1 + * = ( · [ℎ− 1, ] + ) ℎ = * ℎ() (4) (5) (6) (7) (8) (9) Min-max normalization
One hot encoder
LSTM(50) LSTM(50)
Dense(15)
LSTM layers with 50 number of hidden states and one dense node for output classification as
shown in Fig. 6. We use min-max normalization to normalize the data [
There are two types of machine learning algorithms (i.e., parametric and nonparametric models). In parametric models, we estimate parameters from the training dataset to learn a function that can classify new data points without needing the original training dataset. Logistic regression and linear Support Vector Machine (SVM) are two examples of parametric models. Nonparametric models, on the other hand, cannot be defined by a fixed set of parameters, and the number of parameters increases proportionally as the training data. Examples of nonparametric models in ML are K-Nearest Neighbor (KNN), random forest, and decision tree. 3.4.1. K-Nearest Neighbor KNN is a nonparametric model used for classification by estimating the likelihood that a data point will become a member of another group based on the similarity of the patterns. It is a simple yet eficient algorithm and provides a benchmark for comparing against other ML algorithms. The KNN algorithm can be summarized as follows. First, it chooses a distance metric and a number of . Second, it locates the k-nearest neighbors from the data to be classified. It achieves this by calculating the Euclidean distance between the training and test samples in the dataset and then ranks the distance by increasing order. Finally, it tallies votes to assign class labels to the category with the highest number of neighbors. 3.4.2. Support Vector Machine An SVM is an ML model to solve linear and nonlinear problems for classification. In SVM, the algorithm first draws a line (hyperplane) that provides a decision boundary to segregate the data into classes. Then, the SVM algorithm finds the points from both classes that lie closest to the hyperline, and the process is known as "support vectors." The goal is to maximize profit margins (distance between the support vectors and the hyperline), and the best hyperplane is the one with the largest margin. To deal with multi-class classification issues ( > 2), binary SVM classifiers are used. The th SVM is trained so that samples in the th class are labeled as positive and the rest as negative. During the classification stage, a test sample is obtained from each of the SVMs and labeled based on the classifier with the highest output among the classifiers. We adopted the radial basis function as the kernel, considering the relationship between classes and features is nonlinear. 3.4.3. Random Forest Random forest is also a supervised ML algorithm for classification. It comprises several decision trees, with the number of trees increasing the robustness and accuracy of the algorithm. The random forest delivers a class prediction for each tree, and the class that receives the most votes becomes the model’s prediction.
The experiment was conducted on the second floor of the university’s building which has an
area measuring 110 m × 16.9 m, as shown in Fig. 7. The environment consists of a number of
main corridors that connect to diferent rooms. A total of 142 APs were detected during the
experiment. To construct the database, the user utilizes the indoor APs’ information (e.g., MAC
address, RSS values, and timestamp) and surrounding magnetic fields. The data was collected by
a single user using a Samsung Note S8 model. The user walked randomly inside each region of
the building for over the course of two hours during the data collection. The sampling frequency
was set between 0.25 Hz and 1 Hz for Wi-Fi localization [
The mobile phone is held in front of the user’s body, and the user can change any direction during the data collection. To minimize magnetic fluctuations during walking, we fixed the user’s altitude by having them hold the phone (i.e., hand-held horizontally with the Y-axis towards the R8 R15 R4 R7
R6
R5
R4
R3
R2
R1 R14
R13 110 m (a) (b) R5
R10
R13
R14
R15 heading direction). Because data is automatically collected while the user is walking, the method of database construction in this study ofers considerable cost-savings. Furthermore, it can be extended as a crowdsourcing approach in the future. The walking method of collecting the Wi-Fi and magnetic field data is proposed to overcome the drawbacks of the static point-to-point method in the conventional fingerprinting approach.
Given the true and predicted labels, we utilize the metrics of precision, recall, and F1 score for model evaluation and comparing with the benchmark algorithm. These metrics are widely used in evaluating ML settings and our proposed algorithm.
Since we examine the multiclass classification model with an =15 region, we used a 15 class confusion matrix to map one region to all classifications. We define the parameters for each region as follows: 1. ( ) = Number of region is correctly labeled to region . The classifier predicts a positive sample in a true positive event, and the actual value is positive. 2. ( ) = Number of region incorrectly labeled to region . In a false positive event, the classifier makes a mistake by predicting a positive sample with a negative model. 3. ( ) = Number of region not classified to region . The classifier predicts a negative result while the actual sample is positive during a false negative event.
Based upon those definitions, we develop our performance metrics as follows: 4.2.1. Precision It is characterized as the measure of true-positive to predictive positive. It indicates the fraction subregion in an indoor environment that was predicted correctly.
(100%) =
∑︀=1 ∑︀=1 + Fi score is a metric which takes into account both precision (P) and recall (R) and is defined as follows: It denotes the ratio of true-positive to the total number of positive classifications. It indicates the fraction of subregions in an indoor environment that can be accurately classified. ∑︀=1 +
The proposed LSTM-Sigmoid-5500 method outperforms all ML and DL benchmarks methods with a classification accuracy of 95.7%. The classification accuracy of the 15 regions in the indoor environment is displayed in the confusion matrix in Fig 8. As shown in Fig. 8, each indoor region is correctly identified with the prediction accuracy not going below 90%. It means that the proposed LSTM-Sigmoid-5500 functions consistently well across diferent areas.
We tested the performance of the proposed LSTM-Sigmoid-5500 on various numbers of LSTM units and hidden layers, as shown in Fig. 9. Increasing the number of LSTM units resulted in a corresponding improvement in accuracy. However, performance can degrade in some cases when the network goes deeper. Therefore, the best accuracy result is obtained with two hidden layers and 50 LSTM units per hidden layer, as shown in Fig. 9. Hence, we chose a 2-layer with 50 LSTM units as our model.
We compare the validation results of a 2-layer LSTM using sigmoid with the softmax activation function. We selected the same dimensions of the hidden states for all LSTM architectures. Fig. 10 shows the curves of cross-entropy loss and the accuracies and the cross-entropy loss of LSTM-Sigmoid and LSTM-Softmax with 5500 epochs. From Fig. 10, it can be observed that LSTM-Sigmoid always has a lower loss than LSTM-Softmax on the training data. This pattern is also seen in the test data, especially at the tails of the loss graphs.
It can be observed that sigmoid provides a lower loss and consequently produces better accuracy than softmax. In softmax, increasing the probability of one class reduces the overall (10) (11) (12) 2 0.00% 92.42% 7.58% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 3 0.00% 3.39% 96.61% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% probability of all other classes because the activation function requires the sum of the probabilities of the output classes to be one. Conversely, increasing the probability of one class does not afect the total probability of the other classes when using a sigmoid. This is why sigmoid outperforms softmax in multi-label classification. As shown in Table 1, the average precision was 93.9% for KNN, 90.1% for SVM, 88.1% for the random forest, 72.4% for Dense-Softmax-5500, and 66.2% for Dense-Sigmoid-5500, respectively. Meanwhile, LSTM-Softmax-5500 produces the lowest precision at 48.7%. The proposed method outperformed the baseline KNN by 1.8%, 2.4%, and 2.4% in terms of the average precision, recall, and F1 score, respectively. However, compared to LSTM-Sigmoid-5500, LSTM-Sigmoid-5500 increased the average precision, recall, and F1 score by 47.0%, 50.1%, and 53.5%, respectively. LSTM-Sigmoid-5500 is designed to perform indoor classification by retrieving the features based on the Wi-Fi and magnetic field dataset. The accuracy measures of the training and test data are close to each other indicating that the proposed model is not overfitted. As shown in Fig. 8, the lowest prediction happens at region 8, which gives the prediction accuracy of 90.8%, where it wrongly predicts regions 4, 6, 12, 13, and 14, respectively. Meanwhile, the highest percentage of misprediction occurs in region 4, with 7.1% predicted as region 9.
This study proposed the creation of an indoor regional classification method through fusing Wi-Fi and magnetic data collected by smartphone users. An LSTM network was used for data fusion and region classification based on Wi-Fi and magnetic field time series data. We compared our proposed LSTM-Sigmoid-5500 system with the other six benchmark schemes (i.e., KNN, SVM, random forest, LSTM-Softmax-5500, Dense-Sigmoid-5500, and Dense-Softmax-5500). Performance was analyzed in terms of average precision, recall, and F1 score for the diferent classification models. The proposed LSTM-Sigmoid-5500 method outperforms other benchmark methods, addresses the underfitting problem, and works well with time-series data. In the future, we plan to provide fine-grained localization instead of coarse-grained localization and improve the localization performance by using landmarks and more advanced ML and clustering methods. Moreover, we also plan to perform an extensive experiment in diverse scenarios with diferent dimensions and interference to verify the results.
This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT).(No. NRF-2022R1A2B5B01002385).