Floor identification is very important, not only to guide people to navigate easily in multistory buildings and malls but also to save people in case of emergency situations like fire, earthquake, and so forth. However, adding a dedicate setup for floor identification is expensive. Instead, it is more attractive to utilize the existing communication infrastructure for positioning and localization, which decreases the deployment cost. Nonetheless, it increases the dependence on the infrastructure. Such systems follow a fingerprinting technique where real-time/current features including RSS, magnetic data, and channel information are mapped against the pre-collected features of an indoor area.
The idea of fingerprinting refers to associating locations called RPs in an environment with unique characteristics, also known as
fingerprints. Typically signal features are used as fingerprints, whereas a mobile device can be used to gather single or multidimensional features at RPs. These features should exhibit uniqueness that can be used to discriminate different places [
65]. Among several possible candidates for fingerprints, multipath structure and the RSS have been regarded as the most common and accurate fingerprints for horizontal indoor positioning [
19,
98]. Depending upon the indoor structure of buildings,
power-delay profile can be different at different places, provided the carrier frequency is larger than 500 MHz. Similarly, due to the large deployment of WiFi APs for indoor environments, the RSS can be easily gathered and used as a fingerprint.
2.2.1 Fingerprinting Using WiFi APs.
WiFi fingerprinting solutions can be grouped into five categories with respect to the approach used for positioning: traditional fingerprinting, clustering approaches, solutions aiming at dynamic environments, heuristic-based solutions, and fingerprinting with pedestrian activity sensing. Traditional fingerprinting solutions using the RSS values received from WiFi APs become laborious and computationally complex when the size and number of floors grow large, resulting in poor performance. The size of the fingerprinting database grows larger with the increase in the size of indoor space and RPs. Matching real-time data with large databases requires a longer time and negatively affects flood detection accuracy. Clustering approaches focus on dividing the RSS into various groups based on the coordinate or RSS similarity, thereby reducing the computing time and enhancing accuracy. Dynamic solutions focus on leveraging real-time information to update radio maps and add additional people’s movement features in the positioning area. Heuristics-based approaches make use of empirical statistical characteristics of RSS values to enrich the feature vector. Deploying additional information on pedestrian activity through smartphone sensors such as accelerometer and gyroscope helps provide an accurate short-term position that enhances the positioning accuracy.
Fingerprinting solutions tend to be laborious and time-consuming. Expert surveyors wardrive the area to collect the fingerprints during the offline phase, making it a task of significant human resources for sizeable multifloor buildings. Wardrive is the process of collecting fingerprints at ground-truth points for the whole area intended for positioning. Wang et al. [
130] present an approach to reduce the labor involved in labeling the training data. Although the data are collected on all the building floors, labeling is carried out for only one floor. The authors use the information that despite exhibiting different APs and RSS vectors, the floor structure for university campuses is almost similar. AP visibility across several floors is used to find the correlation between the data from different floors. Results show that co-embedding the data from different floors can be used for automatic labeling. Two models named ”the nearest floor algorithm” and ”the group variance algorithm” have been proposed in [
7] to solve the floor identification problem using the WiFi signals. The former is a modified form of the famous
k nearest neighbor (k-NN); however, unlike traditional k-NN, which is used for horizontal positioning, the modified algorithm is used to find the neighbors for positions across different floors. For training and testing, WiFi fingerprints are collected containing
media access control (MAC) and RSS information. Due to the fluctuations in RSS due to dynamic factors, additional parameters are added to the group variance algorithm, such as range, variance, and RSS availability. Experimental results in a five-floor building show a floor detection accuracy of 86% and 72% with the nearest floor and group variance algorithms, respectively. A
linear discriminant analysis (LDA)-based floor identification approach is proposed in [
86] that leverages the
one-versus-one (OvO) rule combined with a majority voting to predict the final position. WiFi RSS values collected at RPs during the offline phase are used for training the classifier. Experiments are performed to analyze the influence of the number of deployed APs on the positioning accuracy using four and six APs for three-floor buildings. Using six evenly distributed APs, a 99% floor identification accuracy is obtained.
While the use of RSS values alone is the predominant approach, deploying additional features from the received signals tends to show better results. For example,
confidence interval sum-received signal strength (CIS-RSS) is used in [
89] for floor detection using the RSS from WiFi APs in a multifloor environment. The sum of RSS shows the different distributions for each floor and can be utilized for positioning. Online RSS values are used to estimate the CIS-RSS using the RSS values from installed APs on each floor, and a 100% accuracy can be achieved at a 95% interval. Similarly, Han et al. [
52] use a feature vector containing both RSS and MAC addresses of APs on each floor. In addition, the repetition of MAC addresses at each RP is estimated for different floors and recorded in the database as an additional feature. Later, clustering is performed based on MAC addresses for each floor to reduce the computational time and improve accuracy. The feature vector is compared with the already built database during the online positioning to find the most probable floor. Results show that using a cluster size of three, floor identification accuracy of approximately 90% can be achieved. The use of MAC address joined with its RSS across different floors helps to reduce the search space. In larger buildings with a few floors, an AP may only be visible to a few floors, and recording the MAC address for AP along with RSS and floor information can provide better results, as reported in [
52].
Fingerprinting is a time-consuming process and becomes tedious where large buildings with tens of floors are involved. One possible way to reduce the time and effort involved in fingerprinting is crowdsourcing. Crowdsourcing involves multiple users in the data collection process, which are later integrated into a single database. Since it uses non-expert surveyors, data collection is substantially reduced, compared to traditional fingerprinting. Khaoampai et al. [
66] introduce a similar concept, called ”fingerprint self-learning,” where a new fingerprint (one that is not present in the database) is added to the database with the pedestrian’s current floor. The user’s current floor is determined using the last known floor and activity trace, which is achieved using the accelerometer and barometer data from the smartphone. Using the proposed approach FloorLoc-SL, an 87% accuracy can be obtained on the smartphone.
Clustering is another potential solution to improve positioning accuracy in floor identification schemes. Clustering involves dividing the data into different groups based on the similarity of intra-cluster samples and differences of inter-cluster samples. For large indoor buildings with tens of floors, the size of the fingerprint database can grow exponentially, increasing the processing time and causing latency for real-time positioning systems. Clustering can play an important role in clustering fingerprints and reducing the search space, thus improving the response time and enhancing the positioning performance. Clustering for fingerprinting solutions can be based on 3D coordinates or AP based [
18]. A clustering approach for floor identification is proposed in [
34], which utilizes the concept of
penalized logarithmic Gaussian distance (PLGD). The concept of penalizing involves various penalties for clusters showing poor performance. Experiments are performed in two buildings to analyze 3D and AP-based clustering performance. Results demonstrate that both 3D and AP-based clustering reduce the response time. However, 3D coordinate clustering enhances the positioning performance. Using the proposed approach, an accuracy of 97% can be obtained for floor detection. Along the same directions, a clustering-based fingerprinting approach is presented in [
107], where a modified K-means clustering approach is used along with the
weighted centroid localization (WCL) algorithm. In the weighted approach, the position of a device is computed as the weighted average of visible AP positions. Clustering is carried out for RSS fingerprints collected for each floor to use only the
cluster heads (CHs) for localization. An average floor identification accuracy of 88.00% can be achieved involving two university buildings, one shopping mall, and one office environment. Despite the lower accuracy, fingerprint database size and consequent computational time of position estimation are reduced substantially.
Tiwari and Jain [
120] propose using Fuzzy means clustering for WiFi data and statistical features from a barometer to perform floor detection. Initially, different buildings are identified using the WiFi majority rule, followed by floor identification through a barometer. Several statistical features from barometer data are used, including
root mean square (RMS), Kurtosis, Skewness, peak-to-peak, Crest factor, shape factor, margin factor, and impulse factor. Floor identification accuracy of 98.21% is achieved with a complexity of
\(O(1)\). An unsupervised clustering (K-medians and Kohonen layer) is adopted by Campos et al. [
28] to group the collected RSS by imposing architectural constraints. Initially, the online collected RSS values are processed using
principal component analysis (PCA) to reduce data dimensionality. A backpropagation
artificial neural network (ANN) is used to find the floor of the mobile device. Using a window length of 25 with a combination of unsupervised and supervised clustering and
database correlation methods (DCMs), 97% vertical positioning accuracy can be achieved.
Besides clustering, employing probabilistic approaches helps to increase the robustness of fingerprinting solutions. Owing to the lower processing and storage available on mobile devices, robust and lower computational complexity solutions are desired. Razavi et al. [
108] analyze the performance of four robust algorithms with gathered RSS, including weighted centroid localization, nonlinear joint parameter estimation and trilateration, linear joint parameter estimation and multilateration, and deconvolution-based path loss estimator. Experiments are carried out in two university buildings, a six-floor mall, and a shopping center. Results reveal that a probability of 85.80% is achievable with a weighted centralized localization algorithm.
The major drawback of the fingerprinting solutions is updating the fingerprint as the APs may be replaced, be added, or experience presence failure. AP presence failure during the online RSS collection has a significant influence on floor detection accuracy. Maneerat et al. [
88] present a solution to AP presence failure using a
robust mean of sum (RMoS), which is based on the mean of the summation of the strongest RSS and uses the confidence interval comparison [
59]. Within a five-floor environment, 95% accuracy is achieved with 40% RN failure. Despite the low cost and high accuracy of fingerprinting solutions, fingerprint databases become outdated due to changes in the indoor infrastructure including the addition of furniture, vending machines, computer equipment, and so forth. Similarly, RSS varies with different types of mobile devices. In addition, the influence of human mobility also affects RSS values. To counter the effects of human mobility and device heterogeneity, an adaptive indoor positioning system, DIPS, is presented in [
9]. It comprises a dynamic radio map generator, RSS certainty, and people location component. Considering these elements, it integrates the people’s presence into the RSS map. Results suggest that incorporating this information into the radio map helps to obtain a 99% floor identification accuracy.
A heuristic approach based on the collected RSS for WiFi signals is presented in [
23] to estimate the floor of a pedestrian. Four characteristics from the WiFi signals for a multifloor environment are used to determine the floor information: floors with a maximum count of signals, floor with maximum signal strength, floor with maximum average signal strength, and floor with maximum signal strength variance. Such observations are attributed to the attenuation of wireless signals through ceilings and floors. A maximum of 99.97% accuracy can be achieved using the proposed Locus system. Table
4 (given in Appendix A) provides characteristic features of WiFi and cellular-based fingerprinting solutions.
Owing to the complexity of indoor infrastructure and dynamic environments, positioning accuracy suffers from fingerprinting solutions. One way to enhance the performance of fingerprint-based approaches is combining the activity recognition from inertial sensors of mobile devices [
105]. Providing information on an indoor structure like the position of stairs and elevators, main entrance, and height of floors further elevates the floor identification performance. Sun et al. [
118] present a multistage floor identification approach. It involves a discriminative floor model using
Fisher’s linear discriminant (FLD), pedestrian state recognition, and WiFi fingerprinting matching. A floor identification accuracy of 94.3% can be achieved with the proposed model.