Hybrid indoor location: Simultaneous zone and coordinates based location for AAL environments with 802.11 fingerprinting technology

Abstract

The continuous both indoor and outdoor location of subjects is an essential capability within AAL systems. It enables adaptive and context-aware behavior within the services implemented. In AAL systems, location information has been managed in a simplistic way until now: i.e. it refers either to specific rooms or concrete $(x, y)$ coordinates within the elder’s home. Understanding it as specific rooms is the most usual approach. In this paper we argue that managing both information granularity levels simultaneously is also interesting within AAL. And we restrict the discussion to 802.11 fingerprinting based indoor location technologies. We present here how to systematically build and evaluate 802.11 fingerprinting based indoor location systems for both approaches and how to integrate them within a unique service. In regard to fingerprinting methods, managing both information levels simultaneously allows reducing deployment and maintenance cost for such technology.

Keywords

8021.11 fingerprinting indoor location within AAL environments zone based location coordinates based location hybrid location

1. Introduction

Indoor location technology [25] has received a considerable attention from both academia and private companies during the last years. Position of devices and users is a capital piece of information in Context Aware Computing [14], thus it is crucial within the Ambient Assisted Living (AAL) domain. Technology like GPS (Global Location System) and A-GPS already provide a solution outdoors. However, due to the distortion of GPS signals indoors, location error significantly increases there. In consequence, alternative solutions are needed.

One of the most prominent technologies for indoor location is the so called 802.11 (WiFi) fingerprinting (WFP). The basic idea is simple yet powerful and it rests on the assumption that 802.11 signals are virtually ubiquitous nowadays and require that the element being located comes equipped with a 802.11 interface. Given that the 802.11 signal degrades with distance and obstacles, the vector compound by signal strength of the different 802.11 devices perceived at a given point can be used as a reference to locate devices at the same point the vector was created at. This technology needs a site survey phase in which a number of such vectors gathered in specific locations, labeled with the location, are stored in a data base. This data is then used to create a location function, mapping observed vectors to locations.

This technology presents two variants depending on which device measures signal strength. In the first variant there exists a set of 802.11 dedicated hardware elements deployed at the indoor space in charge of registering signal strength of the devices being located. Usually, such hardware gives wireless connection to a network as well meanwhile generating information for location. In the second variant the device being located measures signal strength from the 802.11 fixed devices (i.e. WiFi routers) acting as landmarks. For both variants, the location function produces the final location. We will refer to the later variant as infrastructureless WFP (IWFP) emphasizing the fact that no dedicated 802.11 infrastructure is needed to deploy a location service.

Regarding the application of WFP for positioning in AAL environments, it is of interest to state possible requirements and how WFP matches them. As this technology locates 802.11 devices, it is applicable to locate persons wearing a device (e.g. a smart watch) or mobile things of interest for the subject of AAL environment (e.g. door keys, smartphone, a companion robot). Also, AAL systems are not critical in the location dimension, i.e. they can tolerate location errors to some extent. And this is important as WFP’s output as a location service bears a measurable error. Cost is also important in AAL systems when it comes to systems installed at home. The first variant of WFP considered in the former paragraph required dedicated hardware. However, the IWFP variant can use WiFi routers distributed through the environment (e.g. from neighbours or companies nearby). Thus, in principle IWFP has more applicability in AAL. Thus, in this paper, we focus our attention on IWFP.

This work intends to make advances in three different directions. The first direction is related with the specific techniques used to create the mapping mechanism from signal strength vectors to device locations. As the error generated by the generalization process used when predicting the device’s position should be optimized, we support the argument that ensembles (i.e. a technique based on the combination of multiple basic location functions) are a powerful way to reduce it. The second one, we call it hybridization, refers directly to the integration of two different definitions of the location concept. WFP understand location information as a label (e.g. the living room, the kitchen) or as a pair of coordinates in $(x, y)$ but to the best of our knowledge no research has been done on how to combine both in the same scenario. AAL scenarios are specially suited for such combination. For example, in ADL (Activities of Daily Living) the subject’s location is important to recognize what is the activity (i.e. cooking, having a nap). The hybrid approach allows to deliver different location information granularity in regard to activities granularity. If instead of cooking, it is important if the subject is at the kitchen sink (i.e. for washing hands) or sitting at the kitchen table (i.e. for having dinner) and not so important if the subject is in whatever part of the corridor, the hybrid approach we propose if perfectly applicable. The third direction is the delivery in this paper of a detailed procedure to create and deploy the WFP service within AAL environments.

The rest of the paper is organized as follows. Some related works are revised in Section 2. Section 3 outlines the basic ideas behind the methodology followed in this work. Section 4 applies the different steps of the methodology to a specific AAL environment. Finally, Section 5 presents the conclusions and future works.

2. Related work

In WFP, the creation of a location model is the process of, starting from the data base of signal strength vectors obtained with a data survey, obtaining a location function as a mapping between signal strength vectors and locations. Let us denote such database as $SS = {({ss}_{1}, {ss}_{2}, \dots, {ss}_{n}, l) | {ss}_{i}, 1 ⩽ i ⩽ n, l \in L}$ where each ${ss}_{i}$ is the measured signal strength of the i-th access point available at the indoor space and L the set of possible locations the device could be located at. The data modeling process generates a function $f (ss) = l$ , where $ss$ is a tuple $({ss}_{1}, {ss}_{n} \dots, {ss}_{n})$ and $l \in L$ . The vast majority of papers included in this analysis use conventional Machine Learning techniques, e.g. neural networks [4], k Nearest Neighbours (k-NN) [35], decision trees [30] and others to generate $f ()$ . And depending on the form of L, regression techniques are used when L is defined in $R \times R$ . We refer to these systems as coordinate based indoor location systems (CBILS). Meanwhile, when L is simple a set of labels denoting zones of interest, we refer to these other systems as zone based indoor location systems (ZBILS). And the techniques used for providing $f ()$ are classification based.

Probably the most common regression technique employed in indoor location is neural networks [28]. One of the first proposals is [36] dated back to the year 2000. Since then, many other works appeared using the same technique. [3] uses a feedforward network as $f ()$ whose inputs are $ss$ tuples and the output are the coordinates $(x, y)$ , identical to the one we propose here. That work concludes that neural networks obtain results which are similar to the well known technique of k-NN. Other example with the same ANN structure is [29] but it argues that the direction of the 802.11 antenna is important when calibrating and using the system. The authors suggest that a more elaborated site survey must be done using the antenna pointing to different directions. Fahed et al. [16] presents a dynamic neural network (DNN) to try to alleviate the negative effects of 802.11 signal fluctuation. A dynamic database gathering new $ss$ tuples stored them for posterior ANN training. It effectively reduces the impact of non stationarity signal but not of significance for the AAL domain. Fang et al. [17] use a discriminant-adaptive neural network (DANN). This network is adapted to use only useful discriminative information, discarding the redundant information during the training process for a better performance. The results obtained by this DANN outperforms those achieved by a common multi-layer perceptron and a k-NN algorithm in the same environment. ANNs are also combined with other techniques. Derr et al. [13] combine a Counter Propagation Network with k-NN improving the results of other works using neural networks or k-NN algorithms alone. Xu et al. [39] is another example of the same combination. Lin et al. [27] compare the performance of neural networks and k-NN algorithm in terms of five variables: accuracy, precision, complexity, robustness and scalability. K-NN wins the comparison obtaining the best performance with high accuracy and precision. Another comparison between neural networks and k-NN is introduced in Zhou et al. [42], in this case the optimal configuration of k-NN obtains better results but authors expose that neural networks is a better solution due to the following: 1) the performance of k-NN will significantly deteriorate if some parameters are not properly selected, 2) ANNs achieve great location accuracy and also perform better in regard to storage cost and complexity of the inferencing process. Yubin et al. [41] combine ANNs with fuzzy c-mean clustering to select which access points the neural network must consider. Their results are compared with an ANN without preprocessing and k-NN, obtaining a better performance.

Table 1
Overview of related works results. The Machine Learning (ML) techniques are: (DT, decision trees), (BAY, bayesian networks), (FZZ, fuzzy algorithms), (k-NN, k nearest neighbours algorithms), (BAG, bagging), (SVM, support vector machine), (ANN, artificial neural networks), (RBF, radial basis networks)

Classification approaches Room accuracy ML techniques

Badawy et al. [1] 95% DT

Castro et al. [8] 97%* BAY

Garcia-Valverde et al. [21] 82.06% FZZ

Setiya et al. [33] 80% k-NN

Trawinsky et al. [37] 81% BAG+DT, BAG+FZZ

Brunato et al. [7] 96.8% SVM

Saha et al. [32] 97%* ANN, k-NN

Regression approaches Average error CDF ML techniques

Battiti et al. [3] 1.78 m – ANN

Mok et al. [29] – 44.8% within 2 m ANN

Fahed et al. [16] 1.86 m – ANN

Fang et al. [17] 2 m 70.4% within 2.5 m ANN

Yubin et al. [41] 1.51 m 60% within 2 m ANN

Derr et al. [13] – 96.4% within 1.5 m ANN+k-NN

Xu et al. [39] 1.26 m 81.3% within 2 m ANN+k-NN

Lin et al. [27] <1 m* 75% within 2 m k-NN

Zhou et al. [42] 1.91 m – k-NN

Bahl et al. [2] 2–3 m – k-NN

Ross et al. [31] 1.67 m 67% within 2.04 m k-NN

Honkavirta et al. [23] 5.6 m 95% within 13.7 m k-NN

Jekabsons et al. [24] 2.02 m 75% within 3.04 m k-NN

Gwon et al. [22] 3.8 m – k-NN

Farshad et al. [18] – 70% within 2 m k-NN

Silva et al. [34] 2.7 m 30% within 2 m k-NN

Yim [40] 2.3 m – DT

Chen et al. [9] – 83.4% within 1.5 m DT

Laoudias et al. [26] 2.5 m 52% within 2 m RBF

Brunato et al. [7] 3.06 m 50% within 2.75 m SVM

Saha et al. [32] – 95% within 1 m ANN, k-NN

Classification approaches	Room accuracy	ML techniques
Badawy et al. [1]	95%	DT
Castro et al. [8]	97%*	BAY
Garcia-Valverde et al. [21]	82.06%	FZZ
Setiya et al. [33]	80%	k-NN
Trawinsky et al. [37]	81%	BAG+DT, BAG+FZZ
Brunato et al. [7]	96.8%	SVM
Saha et al. [32]	97%*	ANN, k-NN

Regression approaches	Average error	CDF	ML techniques
Battiti et al. [3]	1.78 m	–	ANN
Mok et al. [29]	–	44.8% within 2 m	ANN
Fahed et al. [16]	1.86 m	–	ANN
Fang et al. [17]	2 m	70.4% within 2.5 m	ANN
Yubin et al. [41]	1.51 m	60% within 2 m	ANN
Derr et al. [13]	–	96.4% within 1.5 m	ANN+k-NN
Xu et al. [39]	1.26 m	81.3% within 2 m	ANN+k-NN
Lin et al. [27]	<1 m*	75% within 2 m	k-NN
Zhou et al. [42]	1.91 m	–	k-NN
Bahl et al. [2]	2–3 m	–	k-NN
Ross et al. [31]	1.67 m	67% within 2.04 m	k-NN
Honkavirta et al. [23]	5.6 m	95% within 13.7 m	k-NN
Jekabsons et al. [24]	2.02 m	75% within 3.04 m	k-NN
Gwon et al. [22]	3.8 m	–	k-NN
Farshad et al. [18]	–	70% within 2 m	k-NN
Silva et al. [34]	2.7 m	30% within 2 m	k-NN
Yim [40]	2.3 m	–	DT
Chen et al. [9]	–	83.4% within 1.5 m	DT
Laoudias et al. [26]	2.5 m	52% within 2 m	RBF
Brunato et al. [7]	3.06 m	50% within 2.75 m	SVM
Saha et al. [32]	–	95% within 1 m	ANN, k-NN

The k-NN technique is widely used in this domain, probably for its conceptual simplicity and immediate implementation. Bahl et al. [2] proposed k-NN to improve the results of signal propagation models. Many others have followed the same path [18,24,34]. The works of Ross et al. [31] and Honkavirta et al. [23] compare the performance of the k-NN algorithm with some probabilistic approaches. Gown et al. combine WiFi with Bluetooth by sensor fusion [22], using k-NN as localization algorithm. Examples of other Machine Learning techniques for CBILS are decision trees [1,9,40] and Radial Basis Function Networks [26].

Classification techniques in the field of indoor location are less common. Castro et al. [8] use a Bayesian network (BN) for ZBILS. Depending on room size, number of adjacent rooms and other factors, the system achieves different levels of accuracy. Elnahrawy et al. [15] also use a BN and compare it with a number of well known indoor location systems, e.g. RADAR [2]. Saha et al. [32] use ANNs and k-NN as classifiers, concluding that the achieved accuracy depends on the relative distance among calibration points and on the number of APs used. Other examples of ZBILS are [33] for k-NN, [7,38] for Support Vector Machines and Garcia-Valverde et al. [21] for a fuzzy rules based classifier. This last one presents the novelty of online learning and adaptation through time allowing the system to maintain the accuracy when the environmental conditions change.

Regarding the combination of ZBILS and CBILS, we have found only the work of Brunato et al. [7] proposing Support Vector Machines [12] for CBILS and its classification variant for ZBILS. However, they do not consider their combination in a single service merging different location information granularities.

Most of the techniques reviewed so far were evaluated in terms of the error incurred when locating mobile entities. However, as it can be seen in Table 1, the error’s difference for the best techniques is not significant, at least within the AAL context.

In our opinion, to make the difference over other indoor location systems is necessary to propose different approaches beyond the machine learning employed. These approaches could affect the pre-process or post-process of the techniques. The main novelty of our work is the proposal of a hybrid schema of classification and regression in the same building. Depending of the characteristics of every room or zone, our indoor location system can only predict the room where the user is, or whether this room meets certain requisites, the system is able to detect the concrete location of the user inside the room. Another novelty presented by our work is the use of ensemble techniques to improve the results obtained by machine learning techniques. The application of ensembles to improve classification techniques in the field of indoor location is a relatively new topic in the literature, some recent works from 2013 are beginning to apply them, e.g. the work of Trawinsky et al. [37]. They apply bagging to the outputs of two different classifiers, a decision tree and a fuzzy rule-based classification system. The results with bagging outperforms those achieved by the classifiers individually.

3. System creation and deployment

This section briefly presents the steps to create and deploy the Hybrid indoor location system. The nine steps of the methodology appear graphically represented in Fig. 1.

Fig. 1.

Steps to create and deploy an indoor location service. Notice that the graphical representation for steps 2, 3 and 4 abuses notation to briefly represent that the three steps can be performed through only one of the two threads (i.e. classification or regression).

Firstly, a site survey (step 1) is performed in the building where the indoor location system it is going to be deployed. A site survey phase is needed to obtain the $SS$ data defined in the first paragraph of Section 2. In order to collect the $SS$ data, the site survey is done by an operator who walks around the building with an Android mobile device, running a scanning software, gathering tuples for $SS$ at different locations of interest and for a predefined time.

Regarding CBILS site survey, an imaginary grid of calibration points must be designed over the indoor zone. Calibration points are appropriately located within the center of 2×2 metre tiles. A distance of 2 meters was chosen after reviewing the literature as most of the related works reach mean distance errors around 2 meters. Data gathering must be, and every point during 2 minutes and 7 days.

The case for ZBILS is simpler as no grid of calibration points is required. Instead, the operator should move freely through the space, taking into account that it must spend 2 minutes per zone.

Once step 1 is done, what we get is $SS = {({ss}_{1}, {ss}_{2}, \dots, {ss}_{n}, l) | {ss}_{i}, 1 ⩽ i ⩽ n, l \in L}$ where each ${ss}_{i}$ is the measured signal strength of an access point available at the indoor space and L the set of possible locations the device could be located at. Step 2 is in charge of preprocessing data and separate it into a training data part and a test data part. In the approach we use, data preprocessing is in charge of labeling ${ss}_{i}$ values from readings in $SS$ corresponding to the i-th AP which was not seen in the corresponding lecture. The RSSI values range is $[- 100, 0]$ , so to distinguish an access point with very low signal strength from an unseen AP, we use −125 in these cases. To split the dataset into a training and a validation data sets, we usually use the last day of the site survey for validation purposes and all the previous days for training the algorithm, e.g. when a site survey of 4 days is performed, the first three days constitute the training data set while the last day is used for validation. As this process is iterative, the whole pipeline can be performed since the second day, using data from the 1st one for training and data from the 2nd one for evaluating.

Step 3 consists on training the machine learning model. Using the training data set obtained in the previous step, a supervised machine learning model is trained to be used as location function. This step is generic so any supervised machine learning technique could be used (see Section 2).

Fig. 2.

Layout of the two stories home.

Once the obtained the best possible model given the Machine Learning technique at hand, it must be tested in step 4 to prove if the results obtained with the model are sufficiently accurate. Then using the validation dataset as input for the trained model, the model generates a list of inputs which must be compared with the original outputs of the validation dataset. In case of a classification technique, a confusion matrix is generated. A confusion matrix M allows to visualize the performance of a classification algorithm. The value $m_{ij}$ of M represent how many times the algorithm predicted the room i while the user was really located in the room j. By this way, the accuracy of the algorithm can be calculated as a hit percentage, i.e. $Ac = \frac{h}{p}$ being h the number of prediction hits and p the total number of predictions.

Meanwhile, for a regression technique, the error e is a distance in meters calculated as follows. $e = \sqrt{{(e_{x})}^{2} + {(e_{y})}^{2} + {(e_{z})}^{2}},$ where $e_{x} = | P_{x} - O_{x} |$ , $P_{x}$ represents the variable x predicted by the location function and $O_{x}$ represents the original variable x of the validation set.

After Step 4 is finished, it is necessary to check if the accuracy of the algorithm is sufficiently good to be proven in a real environment. The developer must consider if the results achieved in step 4 could cover the necessities of the application which the indoor location system will be applied. In the results are good enough, then the machine learning model is selected. Then the hybridization process to combine both ZBILS and CBILS follows. Again, the steps are similar as those in steps 2, 3 and 4. See Section 4.3 for a detailed explanation.

4. Location at home: AAL scenario

This section illustrates how to apply and validate the technology proposed in this paper to a typical AAL scenario: a real two stories house located in a residential area of Murcia. Figure 2 shows the house’s layout.

This section’s purpose is threefold. Firstly, it delivers empirical evidence about the dominance of the Random Forest [6] machine learning technique over BNs (see Section 4.1) when it comes to ZBILS. This conclusion is important as Bayesian Networks was the technique of choice in many of the papers reviewed and it actually produced good results.

Secondly, it will show how to create a CIBLS with neural networks through a careful treatment of the learning process (see Section 4.2). In the experiments performed, the model obtained outperforms the rest of CBILS systems reviewed. And last but not least, we will illustrate how to combine a CIBLS with a ZIBLS to construct a hybrid solution capable of mixing coordinates and zone location depending on user requirements (see Section 4.3).

4.1. ZBILS within an AAL scenario

A BN [10,11] is a model of probabilistic dependency between variables. In regard to WFP location, such variables are all continuous (RSSI values) but one (labels representing the rooms of the building). The BN learning process is supervised. During the training phase, the model can learn the probabilistic dependencies which exist between the input and output variables. Among all the possible probabilistic dependencies, it is necessary to find a network structure that minimize the classification error (i.e. in this case maximizes the accuracy of the ZBILS). Once the network is trained and validated, it can be used to generate the probability for a mobile device to be a specific room, depending of the RSSI values obtained from it. Each time the learnt BN receives a set of RSSI values from the mobile device, it returns a probability for each possible room. Locating the mobile device is a matter of choosing the room whose probability is the max one.

Fig. 3.

(a) Evolution of Bagging based classifier’s accuracy with the number of iterations. (b) Evolution of Random Forest based classifier’s accuracy depending of the number of trees generated.

In this paper, we wanted to go further than using stand alone Machine Learning techniques by applying ensembles. Ensembles are meta-techniques, in the sense that they work using stand alone Machine Learning techniques as basic building blocks to create more complex models, the ensembles. The most well known ensemble based techniques are Boosting [19], Bagging [5] and Random Forests [6]. Regarding boosting, we have used AdaBoost [20]. This algorithm focuses on examples that are the hardest to classify correctly (i.e. RSSI lectures which can lead to different locations). As a consequence, the overall accuracy of the algorithm will be improved. The results achieved with AdaBoost in our experiment over the Bayesian Network with Tabu Search was slightly better, a 0.14% of improvement after 5 iterations. Bagging (Bootstrap aggregating) is an ensemble based technique. Bagging is a meta-algorithm able to improve the results of weak classification algorithms, combining the classifications of randomly generated training sets. After using the Bagging algorithm with the BN used in this paper, the results obtained in this experiment has not improved the results of the Bayesian Network with Tabu Search. It actually behaved as expected as BayesNet is not precisely a weak algorithm. Random Forests work generating a set of decision trees at training time, using to construct each tree a random subsec of the predictor variables (i.e. a subset of the APs in this case). The different outputs of each tree for the same unlabeled example are aggregated using a voting strategy. The most repeated location is the one selected. This experiment tries all of them.

The models produced by all the techniques mentioned above are validated by using always the last day of the calibration process for testing. As we mentioned, the outcome of such a test can be represented as a confusion matrix Based on a CM structure, the accuracy, $Ac$ , of a ZBILS is calculated as follows: $Ac = \frac{\sum_{i = 1}^{n} m_{ii}}{\sum_{i = 1}^{n} \sum_{j = 1}^{n} m_{ij}},$ and the main diagonal represent hits of the model when locating. A CM for location can be even more interesting if contiguous rooms appear as contiguous columns in the CM.

To work with BNs we use Weka software, and the BayesNet classifier. This algorithm, combined with K2 as the search heuristic gave us a room accuracy of 73.06%. It was the best result over the rest of search heuristics (i.e. Tabu Search, Hill Climbing, etc.). The room accuracy after applying bagging to the Bayesian Network is 76.55%, i.e. an improvement of 3.49%. Figure 3(a) shows the evolution of Bagging with different number of iterations.

However, Random Forests clearly outperformed the rest of approaches. Figure 3(b) shows the accuracy evolution for different number of trees. With 170 trees, the accuracy is 78.45%, i.e. 5.39% better than the original Bayesian Network approach and 1.9% better than the Bagging approach.

4.2. CBILS at home

In this step, we are concerned with location users at any point $(x, y)$ within a building. The typical Machine Learning task for this problem is regression, i.e. learning a target function generating predictions in $R^{n}$ (i.e. $n = 2$ in this case). The motivation for this experiment is illustrating to what extent, it is possible to achieve a reasonable distance error when it comes to CBILS based on WiFi fingerprinting.

Fig. 4.

Situation of the calibration points grid over the two stories house.

Notice that creating a regression model to locate in $(x, y)$ is considerably harder than a classification model to locate in zones. Now, the location service is required to distinguish between contiguous $(x, y)$ positions, usually with no obstacles between them. Thus, RSSI values gathered from both positions an used for learning are actually very similar (i.e. highly noisy examples). Moreover, the service must be capable of generalization in points where there was no real data gathered from them (see next paragraph). Another interesting point is that the analysis of the related works (see Section 2) shows that most of the techniques achieve a similar level of performance when building a CBILS system. In coherence with this, the purpose of this section is not arguing about the generation of the best fitted model but a detailed illustration on how to employ the procedure described in Section 3 to get a reasonably good $f ()$ .

We relay on the generalization power of the Machine Learning models when we assume that calibrating in concrete points of the room is sufficient to provide a position estimate for any point of the room. The first step of the experiment consists of the subdivision of a concrete room in calibration points. The calibration process is very similar to that performed for the entire house for classification purposes. But in this case, rooms must be atomized into subzones. The accuracy we will obtain is inversely proportional to the size of the subzones. The smaller the zones are, a greater granularity will be achieved. However, atomizing a room into very small zones has a disadvantage: more time has to be devoted to calibration.

Fig. 5.

(a) Evolution of training and test error during learning for each MLP configuration (i.e. 5, 10, 15 and 20 hidden neurons). (b) CDF of the MLP whose learning process were stopped at the best test error (5, 10, 15 and 20 hidden neurons).

We subdivide the room in zones of 2×2 meters. We have selected 2 meters as a reference value because after reviewing the literature, the majority of related works obtain error values which are very close to 2 meters. It is actually very difficult to achieve a better error than that. Thus, such subdivision is a good trade-off between size and error. Thus, an imaginary grid of calibration points is layered on the indoor space at equal constant distance between contiguous points (see Fig. 4). The rest of the process for creating the location service remains the same: data must be gathered for each calibration point, and labeled with the corresponding two $(x, y)$ values (note that in this case $(x, y)$ will not be labels but real values. The calibration of a zone to deploy a CBILS is considerably more expensive than calibrating the same zone to deploy a ZBILS. However, such cost can be justified when the need for a CBILS instead of a ZBILS is a must.

We have subdivided the house in two zones, see Fig 4. All the rooms inside this zone were also subdivided in 34 calibration points. In this case, the site survey performed at every calibration point lasted one minute, a total of 34 minutes per day. Notice that we are considering two floors, i.e. the service will actually locate in $(x, y, {1, 2})$ .

We decide to use an ANN for creating the regression model as they show good results. This particular ANN uses is a single hidden layer feed forward network of sigmoidal perceptrons (i.e. a multi-layer perceptron MLP). It has 7 inputs, as only 7 APs are detected at this building. The network has 3 output variables, representing the coordinates x, y and z. Note that the z output will be a numeric value in meters indicating the floor where the user is (i.e. a value of 1 in the ground floor and 4 in the first floor). The calibration data is composed of 5 days, subdivided in the first 4 days for training (8195 examples) and the last day for validation purposes (2095 examples).

Figure 5(a) shows the curves of training and generalization of the ANN. The best average error was achieved with the 20 hidden neurons approach, 1.83 meters. All the results of average, maximum and minimum error are shown in Table 2. Figure 5(b) shows the average error CDF for all the configurations, i.e. 5, 10, 15 and 20 hidden neurons.

Fig. 6.

(a) Resulting decision tree. (b) Whisker-Plot diagram showing the error of the 2 output variables. (c) CDF of average error of the best ANN and the best regression tree.

The regression tree model created in this experiment has 7 input variables and 3 output variables. The best average error achieved with this model is 2.31 meters. Table 3 contains the average, maximum and minimum error for two different configurations of regression tree. The default tree and the best one. The resulting best decision tree is shown in Fig. 6(a). Figure 6(b) compares the outputs of the best regression tree and the best ANN by Whisker-Plot diagrams. Meanwhile, Fig. 6(c) characterizes their average errors in terms of Cumulative Distribution Function (CDF).

The technique which obtains better location results is the MLP.

4.3. Hybridization of classifiers and regression models

This section proposes an hybrid ensemble to merge classification and regression techniques to generate a single

f ()

capable of generating both zone and coordinates based location predictions. Let us direct again our attention to the house we have used extensively in the former experiments. Suppose also that the AAL system in use employs ADL techniques which require different levels of information granularity (i.e. coordinates and zones) depending on how critical is the activity that must be recognized at rooms (e.g. the system needs a precise elder location at the kitchen and bathroom but only zone based location at the corridor, living and bedroom). Let us denote those rooms requiring coordinates granularity as

{r_{1}, r_{2}, \dots, r_{n}}

and those requiring zone granularity as

{l_{1}, l_{2}, \dots, l_{m}}

. Given the technology we have already shown, the direct method to construct a location service would be to implement a CBILS system for the whole house and use postprocessing to convert those coordinates falling into

{l_{1}, l_{2}, \dots, l_{m}}

zones to the right label. However, the site survey of an indoor location system based only on regression will require higher calibration times. Performing the right calibration for each zone and then integrating the output of both CBILS and ZBILS systems is cheaper.

Table 2
Results on Resilient Backpropagation over a single hidden layer feed forward sigmoidal perceptron

	Average error	Maximum error	Minimum error
5 hidden neurons	1.96 m	6.68 m	0.15 m
10 hidden neurons	1.86 m	6.62 m	0.06 m
15 hidden neurons	1.88 m	7.48 m	0.07 m
20 hidden neurons	1.83 m	6.24 m	0.11 m

Table 3

Distance error for decision trees

	Average error	Maximum error	Minimum error
Default tree	2.35 m	6.71 m	0.39 m

Best tree	2.31 m	6.42 m	0.39 m

Fig. 7.

Layout for testing the two hybridization strategies. In dark, rooms with $(x, y)$ granularity.

We have defined two different strategies for merging. The first one is based on the generation of a ZBILS, L, for the whole house, i.e. all rooms in ${l_{1}, l_{2}, \dots, l_{m}, r_{1}, r_{2}, \dots, r_{n}}$ and the required regression models (i.e. neural networks) ${R_{1}, R_{2}, \dots, R_{n}}$ , corresponding to the rooms $r_{i}$ . If, for a given input, L predicts that the user is located in a room $l_{j}$ , $1 ⩽ j ⩽ m$ , then $l_{j}$ is the location output. Otherwise, if L predicts that the user is located in a room $r_{i}$ , the regression model $R_{i}$ , $1 ⩽ i ⩽ n$ returns a $(x, y)$ location output.

The second approach consists on a meta-classifier $Mc$ which predicts whether the device is located at any room into the set ${r_{1}, r_{2}, \dots, r_{n}}$ or at any room of the set ${l_{1}, l_{2}, \dots, l_{m}}$ . $Mc$ ’s output has only two possible labels, R or C. If the output is R, a regression model R trained with lectures of all the rooms of the set ${r_{1}, r_{2}, \dots, r_{n}}$ makes a new prediction returning the required $(x, y, z)$ position. Otherwise, if the output of the meta-classifier is C, a classification model C trained with lectures of all the rooms of the set ${l_{1}, l_{2}, \dots, l_{m}}$ classifies the concrete room where the user is located.

Table 4

Accuracy and average, maximum and minimum errors for the two hybridization strategies

	Room accuracy	Average error (meters)	Maximum error (m)	Minimum error (m)
Approach 1

Classifier	78.45%	–	–	–
Regression Model Area 1	–	1.76	3.36	0.03
Regression Model Area 6	–	0.99	3.7	0.01

Approach 2

Meta-Classifier	96.67%	–	–	–
Classifier	76.78%	–	–	–
Regression Model	–	1.77	3.97	0.06

The main disadvantage of the first strategy arises when there is a high number of rooms requiring $(x, y)$ location granularity. The higher the number of rooms requiring $(x, y)$ location granularity, the higher the number of different regression models are required. On the other hand, having one regression model per room has an advantage. The number of outputs of the models can be reduced from $(x, y, z)$ to $(x, y)$ when there is more than one floor. Simpler models, i.e. with less output variables and covering smaller areas will result in a better location performance. Another advantage is that only one classifier is required, so when a room $l_{j}$ , $1 ⩽ j ⩽ m$ is predicted, no more actions are required. Notwithstanding, the second strategy always requires two predictions. Besides, it requires two classifiers and one regression model generating outputs in $(x, y, z)$ . The strong point here is that only a single regression model is required, thus it is better for dwellings with $(x, y, z)$ granularity in a large number of rooms.

To perform experiments with both approaches, we have employed the house environment of Fig. 7. It shows in dark the two rooms of the set ${r_{1}, r_{2}, \dots, r_{n}}$ , in other words, the two rooms which need $(x, y, z)$ granularity. These rooms are Area 1 and Area 6. The other rooms only need to be classified. Looking at the results obtained in Section 4, we have use Random Forests classifiers and Artificial Neural Networks as regression models. Table 4 shows the results obtained of both hybrid approaches. Figure 8 shows the average error CDF for all the regression models, i.e. the models of area 1 and area 6 in the first approach, and the single model that covers these two areas in the second approach.

Fig. 8.

CDF of the MLPs of the two hybrid approaches.

Considering the results obtained and the advantages and disadvantages of every approach previously exposed, our suggestion is that depending of the necessities of the application which is going to use the location engine, the developer must choose one of the two approaches. The first approach could be preferred in buildings where a few rooms need $(x, y)$ granularity as the complexity of the engine could be high with a large number of regression models. The developer could prefer to train one regression model in these cases, even if this one obtain slightly worse results of average error. Another reason to choose the first approach in a building with few rooms with $(x, y)$ granularity is that the classifier only needs one prediction. So every time the user is located in a room of the set ${l_{1}, l_{2}, \dots, l_{m}}$ , one prediction in comparison with the meta-classifier approach.

5. Conclusions and future work

This paper presents an approach to deal with different location service types in the same location. We have seen that this can perfectly be the case within the AAL domain. The paper presents how to set up the CBILS and the ZBILS depending on the particular necessities at home and how to produce the hybrid model in charge of merging both services. It includes different criteria that can be used in order to select the two different strategies for the hybrid location service. But most important, it shows how to assess the performance of the service before deciding whether to deploy it in production mode. As a way of example, ANNs and Random Forest have been used as the techniques to construct the location services but this work does not put the emphasis in producing models to outperform other former works but in how to use them.

Future works will advance in the direction of measuring the level of degradation of the service in real homes in order to more precisely quantify the cost of maintaining the service. In consequence, business models capable of delivering a profitable and quality AAL service must be defined.

References

[1]

Mohamed Badawy and

M.A.

Bani Hasan, Decision tree approach to estimate user location in wlan based on location fingerprinting, in: National Radio Science Conference, NRSC 2007, IEEE, 2007, pp. 1–10.

[2]

Bahl and

V.N.

Padmanabhan, Radar: An in-building rf-based user location and tracking system, in: Proc. of the Nineteenth Annual Joint Conference of the IEEE Computer and Communications Societies, INFOCOM 2000, Vol. 2, IEEE, 2000, pp. 775–784.

[3]

Battiti,

Villani and

LeNhat, Neural network models for intelligent networks: Deriving the location from signal patterns.

[4]

C.M.

Bishop, Neural Networks for Pattern Recognition, Clarendon Press, Oxford, 1995.

[5]

Breiman, Bagging predictors, Technical report, Department of Statistics, University of California, Berkeley, California 94720, September 1994.

[6]

Breiman, Random forests, Machine Learning45(1) (2001), 5–32.

[7]

Brunato and

Battiti, Statistical learning theory for location fingerprinting in wireless lans, Computer Networks47(6) (2005), 825–845.

[8]

Castro,

Chiu,

Kremenek and

Muntz, A probabilistic room location service for wireless networked environments, in: Ubicomp 2001: Ubiquitous Computing, Springer, 2001, pp. 18–34.

[9]

Chen,

Yang,

Yin and

Chai, Power-efficient access-point selection for indoor location estimation, IEEE Transactions on Knowledge and Data Engineering18(7) (2006), 877–888.

10.

[10]

G.F.

Cooper and

Herskovits, A Bayesian method for constructing Bayesian belief networks from databases, in: Proc. of the Seventh Conference on Uncertainty in Artificial Intelligence, Morgan Kaufmann Publishers Inc., 1991, pp. 86–94.

11.

[11]

G.F.

Cooper and

Herskovits, A Bayesian method for the induction of probabilistic networks from data, Machine Learning9(4) (1992), 309–347.

12.

[12]

Cortes and

Vapnik, Support-vector networks, Machine Learning20(3) (1995), 273–297.

13.

[13]

Derr and

Manic, Wireless based object tracking based on neural networks, in: 3rd IEEE Conference on Industrial Electronics and Applications, ICIEA 2008, IEEE, 2008, pp. 308–313.

14.

[14]

A.K.

Dey, Providing architectural support for building context-aware applications, PhD thesis, College of Computing, Georgia Institute of Technology, December 2000.

15.

[15]

Elnahrawy,

Li and

R.P.

Martin, The limits of localization using signal strength: A comparative study, in: 2004 First Annual IEEE Communications Society Conference on Sensor and Ad Hoc Communications and Networks, IEEE SECON 2004, IEEE, 2004, pp. 406–414.

16.

[16]

Fahed and

Liu, Wi-fi-based localization in dynamic indoor environment using a dynamic neural network, International Journal of Machine Learning and Computing3(1) (2013).

17.

[17]

S.-H.

Fang and

T.-N.

Lin, Indoor location system based on discriminant-adaptive neural network in ieee 802.11 environments, IEEE Transactions on Neural Networks19(11) (2008), 1973–1978.

18.

[18]

Farshad,

Li,

M.K.

Marina and

F.J.

Garcia, A microscopic look at wifi fingerprinting for indoor mobile phone localization in diverse environments, in: International Conference on Indoor Positioning and Indoor Navigation, Vol. 28, 2013, p. 31.

19.

[19]

Freund and

R.E.

Schapire, A short introduction to boosting, Journal of Japanese Society for Artificial Intelligence14(5) (September 1999), 771–780.

20.

[20]

Freund,

R.E.

Schapireet al., Experiments with a new boosting algorithm, in: ICML, Vol. 96, 1996, pp. 148–156.

21.

[21]

Garcia-Valverde,

Garcia-Sola,

Gomez-Skarmeta,

J.A.

Botia,

Hagras,

Dooley and

Callaghan, An adaptive learning fuzzy logic system for indoor localisation using wi-fi in ambient intelligent environments, in: 2012 IEEE International Conference on Fuzzy Systems (FUZZ-IEEE), IEEE, 2012, pp. 1–8.

22.

[22]

Gwon,

Jain and

Kawahara, Robust indoor location estimation of stationary and mobile users, in: Twenty-Third Annual Joint Conference of the IEEE Computer and Communications Societies, INFOCOM 2004, Vol. 2, IEEE, 2004, pp. 1032–1043.

23.

[23]

Honkavirta,

Perala,

Ali-Loytty and

Piché, A comparative survey of wlan location fingerprinting methods, in: 6th Workshop on Positioning, Navigation and Communication, WPNC 2009, IEEE, 2009, pp. 243–251.

24.

[24]

Jekabsons,

Kairish and

Zuravlyov, An analysis of wi-fi based indoor positioning accuracy, Scientific Journal of Riga Technical University44(1) (2011), 131–137, Computer Sciences.

25.

[25]

LaMarca and

DeLara, Location systems: An introduction to the technology behind location awareness, Synthesis Lectures on Mobile and Pervasive Computing3(1) (2008), 1–122.

26.

[26]

Laoudias,

C.G.

Panayiotou and

Kemppi, On the rbf-based positioning using wlan signal strength fingerprints, in: 2010 7th Workshop on Positioning Navigation and Communication (WPNC), IEEE, 2010, pp. 93–98.

27.

[27]

T.-N.

Lin and

P.-C.

Lin, Performance comparison of indoor positioning techniques based on location fingerprinting in wireless networks, in: 2005 International Conference on Wireless Networks, Communications and Mobile Computing, Vol. 2, IEEE, 2005, pp. 1569–1574.

28.

[28]

Mitchell, Machine Learning (Mcgraw-Hill International Edit), McGraw Hill Higher Education, 1997.

29.

[29]

Mok and

B.K.S.

Cheung, An improved neural network training algorithm for wi-fi fingerprinting positioning, ISPRS International Journal of Geo-Information2(3) (2013), 854–868.

30.

[30]

J.R.

Quinlan, C4.5: Programs for Machine Learning, The Morgan Kaufmann Series in Machine Learning, Morgan-Kauffman, San Mateo, California, 1993.

31.

[31]

Roos,

Myllymäki,

Tirri,

Misikangas and

Sievänen, A probabilistic approach to wlan user location estimation, International Journal of Wireless Information Networks9(3) (2002), 155–164.

32.

[32]

Saha,

Chaudhuri,

Sanghi and

Bhagwat, Location determination of a mobile device using ieee 802.11 b access point signals, in: 2003 IEEE Wireless Communications and Networking, WCNC 2003, Vol. 3, IEEE, 2003, pp. 1987–1992.

33.

[33]

Setiya and

Gaur, Fingerprinting based localization of mobile terminals using ieee802. 11, World Journal of Science and Technology2(3) (2012).

34.

[34]

J.A.

Silva,

M.J.

Nicolau and

Costa, Wifi localization as a network service, 2011.

35.

[35]

Skalak, Prototype selection for composite nearest neighbor classifiers, PhD thesis, University of Massachusetts Amherst, 1997.

36.

[36]

Smailagic,

Small and

D.P.

Siewiorek, Determining user location for context aware computing through the use of a wireless lan infrastructure, Institute for Complex Engineered Systems Carnegie Mellon University, Pittsburgh, PA, 15213, 2000.

37.

[37]

Trawiński,

J.M.

Alonso and

Hernández, A multiclassifier approach for topology-based wifi indoor localization, Soft Computing, 1–15.

38.

[38]

C.-L.

Wu,

L.-C.

Fu and

F.-L.

Lian, Wlan location determination in e-home via support vector classification, in: 2004 IEEE International Conference on Networking, Sensing and Control, Vol. 2, IEEE, 2004, pp. 1026–1031.

39.

[39]

Xu and

Sun, Neural network-based accuracy enhancement method for wlan indoor positioning, in: 2012 IEEE Vehicular Technology Conference (VTC Fall), IEEE, 2012, pp. 1–5.

40.

[40]

Yim, Introducing a decision tree-based indoor positioning technique, Expert Systems with Applications34(2) (2008), 1296–1302.

41.

[41]

Yubin,

Mu and

Lin, Hybrid fcm/ann indoor location method in wlan environment, in: IEEE Youth Conference on Information, Computing and Telecommunication, YC-ICT’09, IEEE, 2009, pp. 475–478.

42.

[42]

Zhou,