1. Introduction
With the application and development of related technologies based on user location information, location-based services have become essential for daily work and life [
1]. This is particularly crucial in large and complex indoor environments, such as museums, airports, supermarkets, hospitals, underground mines, and other areas in which there is an urgent need for location-based services. Positioning technologies can be classified into two types: outdoor and indoor positioning technologies. In an outdoor environment, global-positioning systems, BeiDou-positioning systems, and other global navigation satellite systems (GNSSs) can provide users with meter-level location services widely utilized in daily activities, providing accurate positioning in outdoor spaces [
2]. The global-positioning system (GPS) has made navigation systems practical for many land vehicle applications. Abbott et al. [
3] introduced a method of integrating a GPS with a simplified inertial navigation system (INS) and provided a technique of using velocity aiding to improve positioning accuracy and reliability. With the increasing popularity of mobile devices with positioning capabilities, such as GPS phones, Zheng et al. [
4] applied a collective matrix factorization method to mine interesting locations and activities. They used them to recommend to users areas for performing specific exercises and activities to participate in when visiting particular sites. GPS-based location services provide more convenient and effective technical support for outdoor positioning.
However, in an indoor environment, where humans are located 80% of the time, GNSS positioning accuracy is drastically reduced owing to the obscuration of buildings and the multipath effect, hindering the satisfaction of the demand for accurate indoor location services [
5]. With the development of wireless indoor positioning technologies, such as Wi-Fi, Bluetooth, and ultra-wideband technology, various indoor positioning technologies and systems have been proposed for providing location services in large buildings [
6]. Bluetooth and Wi-Fi indoor localization are two standard wireless signal-based localization techniques. They can both utilize the features of wireless signals, such as received signal strength indicators (RSSIs) or channel state information (CSI), to estimate location information. They can also adopt the fingerprinting method, which builds a fingerprint database by collecting the signal features at different locations in advance and determines the optimal location by matching algorithms.
Indoor localization has many application scenarios and practical needs, such as emergency management, navigation services, logistics management, smart homes, etc. Filippoupolitis et al. [
7] proposed to use of Bluetooth low-energy (BLE) technology to address the occupancy problem in emergency management using beacons installed in buildings to provide the location information of users and combining machine-learning methods to determine whether there were occupants in specific areas. Moreover, in intelligent energy management, Tekler et al. [
8] proposed a novel plug load management system that also combined BLE and machine-learning methods to determine occupancy in specific areas and reduced the plug load energy consumption and user burden through intelligent plug load automation. Balaji et al. [
9] proposed leveraging existing Wi-Fi infrastructure in commercial buildings and smartphones carried by building occupants to provide occupancy-based fine-grained HVAC actuation in a smart home domain. Tekler et al. [
10] used a feature selection algorithm to select the most essential features from sensor data. Then, they used different deep-learning models to predict occupancy based on these features.
Owing to the large availability of existing infrastructure, Wi-Fi is widely used in homes, hotels, cafes, airports, shopping malls, and other large or small buildings, making Wi-Fi one of the most compelling wireless technologies for location services [
11]. Considering the ubiquity of mobile devices and routers in the experimental site and the comprehensive coverage of Wi-Fi signals, this paper chooses to use Wi-Fi technology for this research. Typically, a Wi-Fi system consists of several fixed access points (APs) deployed in locations known by the system or network administrator that provide easy accessibility and installation. Mobile devices that can connect to Wi-Fi (e.g., laptops and cell phones) can communicate with each other directly or indirectly (through APs), permitting the implementation of a location function in addition to a communication function [
12]. This Wi-Fi positioning system with fingerprinting technology is becoming increasingly popular, and using the ubiquitous received signal strength intensity (RSSI) signal received by a Wi-Fi device for positioning is an effective way to identify a user’s location in indoor environments. To measure the distance between nodes, the RSSI (received signal strength indicator)-ranging technique utilizes the principle of regular signal attenuation with increasing distance for wireless signals [
13]. The signal strength of a transmitting node can be obtained from an RF chip register. Based on the received signal strength, the receiving node calculates the transmission loss of the signal and converts it to distance using a theoretical or empirical model [
14]. This ranging technique only requires a wireless transceiver at a node; no additional hardware is needed, keeping the application cost low.
With the advent of artificial intelligence, the challenges faced in indoor positioning have provoked the use of deep learning to improve the efficiency of positioning frameworks, ushering in cross-era changes. Chen et al. [
15] proposed multisource information fusion positioning technology to effectively utilize Wi-Fi fingerprint data and the geomagnetic field for positioning, addressing the problem that Wi-Fi signals are unstable in complex indoor environments and buildings distort the local geomagnetic field, resulting in low positioning accuracy at a single location source. Liu et al. [
16] proposed a joint convolutional neural network (CNN)-based channel state information (CSI) fingerprint indoor localization method to obtain average positioning errors of 24.7 cm and 48.1 cm in two positioning scenarios in a gallery and a laboratory, respectively, and verified that the joint localization algorithm was effective.
Considering the advantages of Wi-Fi fingerprint-based indoor localization methods combined with neural network methods, this study proposes a CNN-based fingerprint indoor localization model consisting of two stages. In the offline stage, a fingerprint database containing all the reference points in the localization area is constructed for offline training. Meanwhile, in the online scene, an algorithm is applied to match real-time fingerprint information from a user with the offline fingerprint database and estimate the user’s location. This algorithm addresses the problem of the limitation of resources in front-end data collection while providing a low-cost and high-accuracy indoor positioning solution. The main contributions of this study are as follows:
1. 3D ray-tracing technology is proposed to generate RSSI signals in the localization area, as simulated location information can avoid the inherent noise and instability of actual wireless signals, which can cause instability in localization performance.
2. To address small-scale intensive localization needs and tackle the problem of minor differences in RSSI signal characteristics among APs in localization, the construction of a Wi-Fi fingerprint heatmap set is proposed, which can better characterize the differences in intensity characteristics at different reception points.
3. The lack of localization accuracy demonstrated by traditional CNN models is improved upon in this study, providing a framework with excellent localization performances for areas with different depths.
4. Experiments are conducted with the synthetically created and UJIIndoorLoc indoor localization datasets [
17]. The simulated and actual measurement results verify the effectiveness of our proposed localization method.
2. Related Work
An indoor positioning system effectively uses Wi-Fi APs and radio signal strength (RSS) to facilitate localization [
18]. However, implementing a fingerprint-based approach requires time-consuming radio surveys and data acquisition to construct a database for each building. The task of front-end data collection is costly in both time and labor. RSS values are incredibly dependent on the environment, making front-end data collection exceptionally difficult, resulting in meager qualification rates of data collected in the field, which do not meet current survey standards [
19]. In some large-scale scenarios of localization, Zhang et al. [
20] proposed to convert collected Bluetooth RSS into fingerprint images required for calculation and establish a CNN for classification training. However, tedious collection work is often needed before research is carried out. In addition, different devices have different signal sensitivities, and data elimination is an important step. Liu et al. [
21] proposed constructing a ratio fingerprint by calculating the ratios of different RSSIs from important contribution access points, which somewhat alleviated the collection work. However, in small-scale scenarios, the percentages of different RSSIs from important contribution access points were also reduced, and the method of constructing ratio fingerprints was no longer suitable for this scenario.
To address the above issues, Li et al. [
22] uploaded the RSSI fluctuations of detected Bluetooth nodes to the cloud and performed real-time correction of the RSSI values. Sinha et al. [
23] simulated constantly varying RSSI values based on reference RSSI values to achieve data augmentation. Both methods processed the data at the front end, significantly saving resources in front-end data collection. Sun et al. [
24] verified that the deployment of radio mapping could dramatically reduce the front-end data collection effort. However, the inherent fluctuation in RSSs generally does not guarantee that the position containing the highest probability predicted by each classifier is actual, resulting in a severe barrier to desired performance in existing fusion methods. To overcome these drawbacks, Hashem et al. [
25] proposed the design and implementation of WiNar, a Wi-Fi indoor location determination system based on the round-trip time that combines the advantages of fingerprinting and range-based techniques to overcome the various challenges of indoor environments. A localization model based on CNNs and extended short-term memory networks was also proposed [
26]. Guo et al. [
27] used the k-nearest neighbors (KNN) algorithm and outlier detection methods to construct an indoor localization framework for simple fingerprints. The existence of just a single evaluation index and poor adaptability to outlier detection hindered the ability of this framework to achieve significant improvement in localization performance. Xie et al. [
28] proposed using a back-propagation (BP) neural network and a weighted KNN algorithm to obtain higher localization accuracy. Wi-Fi indoor localization is highly environment-dependent; however, the BP algorithm was susceptible to initial weights.
These studies highlight the difficulties in providing accurate indoor localization, described as follows. Due to a lack of front-end resources, data collection is labor-intensive and costly under resource constraints. Uneven data cause unsatisfactory localization performances. Finally, improvement in the localization capability of models is hindered by an imbalance in training sample features.
In this study, we use a 3D ray-tracing technique to construct fingerprint datasets of a localization area [
29], addressing the problems of data contamination and consumption in field collection. Furthermore, considering that the differences in data features in self-constructed localization areas are typically too small, we propose the construction of a fingerprint heatmap to characterize the uniqueness of sample features.
To validate the performance of the proposed model in localization, experiments are conducted using the indoor positioning database UJIIndoorLoc and the simulation database, Remcom Wireless InSite 3.3.0 [
30]. The results show that the proposed model performs well regarding localization accuracy while ensuring low power consumption.
6. Conclusions
In this study, we proposed a neural-network-based localization method for Wi-Fi fingerprint indoor localization. Our approach considered the needs of localization areas of different scales. We processed the acquired raw data to construct a grayscale fingerprint map for large-scale scenarios and a thermal fingerprint map for small-scale settings, which could better fit the training requirements of the corresponding scenario data. We conducted experiments on both simulated and real datasets, and the results showed that our proposed method could achieve over 99% validation accuracy for both. Our approach could reduce the workload of front-end data collection to some extent and also provided data support for algorithm validation in localization research.
Similarly, this study still needs to be continuously optimized in future work.
1. The existing method only considered RSSI values in ideal environments when simulating and measuring the positioning data. Since RSSI values are very sensitive to devices and environments, future work should aim to reduce the impacts of uncertain factors on measurement, improve the model’s compatibility with RSSI value fluctuations, and further optimize its performance in positioning.
2. The existing method was based on convolutional neural networks for positioning result classification and discussed the location results in small-scale and large-scale scenarios. Future work should be more fine-grained, consider point-level location output, and optimize the model for more accurate location positioning.