RECITE: A framework for user trajectory analysis in cultural sites

2.1.Underlying architecture

RECITE relies on an underlying operational architecture that enables the collection of the required visitors’ behavioural data. Figure 3 shows such an architecture. Bearing this figure in mind, we can derive the following general assumptions for the framework,

Fig. 3.

General operational architecture for RECITE.

2.2.The classification task

The objective of this work is to obtain an automatic mechanism in such a way that, every given trajectory is classified as one among a set of p predefined labels or classes. Let it be Ω={CL1,CL2,…,CLp}.

2.3.Individual trajectories extraction

The first stage of the framework pipeline focuses on collecting the individual trajectories of visitors using the BLE beacons installed in the museum as Fig. 2 shows. In that sense, a visitor’s trajectory can be defined as follows,

As we can see, RECITE defines a trajectory comprising two types of actions that a visitor can make during his stay,

In the following sections we describe how these two types of actions are collected.

2.3.1.Movement actions

To capture these actions, the guidance device of a visitor v regularly sends to the Location Server (LS) a frame ftv=⟨db1v,t,db2v,t,…,dbkv,t⟩ that contains the measured distances to all its detectable beacons ⟨b1,b2,…,bk⟩ at a certain instant t (see Fig. 3).

In order to process this information, the LS makes use of the museum’s floor-plan image, that is a 2-D Cartesian space where the coordinates of the visitor’s location will be defined. Therefore, the LS performs a mapping process to translate the set of beacon’s distances dbiv,t of a frame ftv to an estimated location (pixel) lest(v,t) within the map image where the visitor v is located at instant t. This point is defined by a particular X-Y coordinates within the map. To come up with a mapping process as accurate as possible, our Map-based location algorithm is implemented in two separate steps, depicted in Fig. 4

• 1. Dataset generation: This first step is done only once for any new museum where RECITE is going to be deployed. To build the map-like dataset we must associate a scanned frame fi for every reasonable location point li in our floor-plan. This might seem a rather time-consuming task, but in fact, we only need to consider the relevant locations where the user can stay (e.g. we do not consider restricted areas or empty spaces of the museum). As a result of this initial step, a set of location frames LF=⟨l1:f1,…,lnloc:fnloc⟩ is generated: one for each of the nloc reasonable locations li where a visitor might be at the museum.

• 2. Comparative inference: In this step, we compare the current frame ftv from any visitor v’s device with the frames in LF. The goal of this comparison is to obtain the target location lest(v,t)∈LF. This is defined as the one that contains the frame fi=⟨db1,db2,…,dbk⟩ whose beacons’ measured distances have the smallest differences with ftv. In order to optimize this search, we only compare ftv with a subset LFtv⊂LF of locations close to the previous location lest(t−1,v) of the visitor v. This is because the sampling rate of frames delivered by the guidance devices should be relatively short (in the range of seconds). Hence, we can assume that a visitor v could not have moved more than a few meters from lest(t−1,v). So we limit the search to this area.

Fig. 4.

Illustrative example of the map-based algorithm steps in an scenario with four deployed BLE beacons. The left image shows the data set generation, where the scanner samples the beacons’ signals for certain locations. Only location frames 1 to 3 (f1, f2, f3) are depicted, but this is repeated for every coordinate in the map. The right image represents the inference step, where a guidance device scans the nearby beacons to produce a frame ftv. Next, it is compared with the frames obtained in the previous step. In this case, the location associated to frame f2 would be the best match. That would be the lest(v,t) eventually inferred by the algorithm.

Lastly, in certain situations the guidance devices are not able to capture the signal of a nearby beacon bi due to several reasons (e.g. collisions or missed samples) [34]. As a result, its associate distance dbi is not included as part of the current frame ftv. This can lead into a significant error when compared with the frame f associated with the actual location lest(v,t)∈LF of v. To mitigate this effect, we pre-process each incoming frame ftv before comparing it with the location frames of LF by combining it with the previous sample ft−1v, calculating a weighted average for those beacons that were included in both frames.

2.4.Trajectory segmentation

One of the problems of the previous gathering process is that the sequences of actions defining the trajectories may have varying lengths. Hence, the second step of the framework pipeline is to apply a segmentation process to all the trajectories in TR to normalize their lengths (see Fig. 2).

In that sense, trajectory segmentation is a well-known technique in the trajectory data mining field that divides a trajectory into fragments by several criteria, like time intervals or semantic meaning, for further processing [42].

In our setting, trajectories are split into slices of a predefined time length tseg. This way, a trajectory segment can be defined as follows,

We can see that nseg indicates the number of resulting segments of the whole trajectory trv whereas trv.tstart and trv.tend are the time instants at which the trajectory trv started and finished.

Once the different trajectory segments have been composed, they are distributed into different sets Sj,j∈[1,maxseg] so that Sj includes the j-th trajectory segments of all the trajectories in TR along with their associated label in Ω. In that sense, maxseg refers to the maximum number of segments of any trajectory trv∈TR.

For example, let us consider a trajectory trex=⟨⟨(5,9),3⟩→⟨(12,8),8⟩→⟨c23,10⟩→⟨(33,10),20⟩→⟨c9,25⟩→⟨(44,19),31⟩⟩ labelled with class CL1. As we can see, its sequence comprises 4 movement and 2 content actions. If we set tseg to 10 time units, the trajectory segmentation process will split the trajectory into 3 different segments (nseg=3), s1ex=⟨⟨(5,9),3⟩,⟨(12,8),8⟩,⟨c23,10⟩⟩,s2ex=⟨⟨(33,10),20⟩⟩ and s3ex=⟨⟨c9,25⟩,⟨(44,19),31⟩⟩.

2.5.Segments feature extraction

Once the trajectory segments have been generated, the next step is to extract a set of descriptive features of each of them. In that sense, depending on the type of action (movement or content display) under consideration, a different type of feature is calculated. As a result, each trajectory segment is compressed into a 3-dimension vector sf. A trajectory segment feature is defined in more detail as follows,

Besides, all the trajectory segment features are collected in their corresponding set Sfj∈X×Y×N×Ω.

Going back to our previous illustrative trajectory trex, its segments give raise to the following features sf1=⟨8.5,8.5,1⟩, sf2=⟨33,10,0⟩, sf3=⟨44,19,1⟩. We can see that, in the case of sf1 (the features of the first segment of the trajectory) its coordinates are calculated as 5+122=8.5 and 9+82=8.5. Finally, each of these features are included in their corresponding Sfj set. For instance, sf1 is stored in Sf1 with their corresponding class CL1.

2.6.Clusters generation and projection

The fourth step of the framework pipeline is to identify the general action trends per segment by means of the generation of different sets of clusters. Every cluster will become a fuzzy rule, and the fuzzy rules will be combined into fuzzy classifiers (see Fig. 2).

The approach is based on the generation of a set of FRCs for every CLr∈Ω={CL1,CL2,…,CLp}. Given a class CLr, the set of data corresponding to that class is clustered according to different spaces obtained increasing the number of segments one by one: Sf1×⋯×Sfj×CLr, ∀j∈[2,maxseg]. The goal of such an incremental clustering process is to generate FRCs able to classify trajectories composed of different number of segments. As we will see later, the identified clusters induce fuzzy rules when they are projected into each one of the dimensions Sfj∀j∈[1,maxseg].

In this phase, the fuzzy c-means (FCM) clustering algorithm [8] is applied. Before explaining that step, let us give some details about the clustering settings:

Our instance of the FCM algorithm generates ncjCLr clusters for each segment feature set Sfj and class CLr. As a result, a cluster is defined by its mean, so called prototype. Furthermore, the adopted fuzzy approach allows each datum in X×Y×N to belong at different degree, from 0 to 1, to each cluster. These membership degrees for each datum to each cluster compose a vector μ with z components (where z is the sheer number of trajectory segment features sjf for a particular number of segments j and the class CLr). The μ vectors are the rows of the so-called partition matrix Uj, associated to the data set, once the algorithm has ended.

For the sake of completeness, the pseudo-code of the FCM algorithm is included in Algorithm 1 where l indicates the iteration number, i the i-th cluster and k the k-th trajectory segment feature for a particular trajectory segment.

Algorithm 1:

Fuzzy c-means (FCM) algorithm

To obtain the fuzzy sets involved in a fuzzy classifier for class CLr, we project each cluster defined in the space X×Y×N into each one of the axis of this space. This is done ∀j∈[1,maxseg]∀CLr∈Ω. As a result of the projection, point-wise definitions of the fuzzy sets are obtained. Then, every point-wise fuzzy set is approximated by a Gaussian-bell in such a way that, to calculate the membership of a given input to the fuzzy set B with centre c and width a,

2.7.Fuzzy classifier composition

At this point we should recall that we carried out clustering tasks based on data of every particular class CLr separately and with incremental number of segments. To generate a FRC, rules based on the same number of segments but different classes in Ω, are combined. The resulting ensemble of FCRs will be used to classify the segment features of a visitors trajectory trsv={sf1,sf2,…,sfnseg}. For example, a FRC based on the maximum number of segments is shown in Table 1. The elements sfk.{x,y,nc} are the ⟨x,y⟩ coordinates and the nc reproductions of the k-th trajectory segment; Blxj, Blyj and Blncj are the j-th fuzzy set for the ⟨x,y⟩ coordinates and nc reproductions in the l-th rule; classl is the consequent of the l-th rule and l=1,…,r, being r the number of rules. Let us explain the classification mechanism.

2.7.1.Online classification mechanism

We should recall that the ultimate goal of the classifier is to label the ongoing trajectories of visitors during their stay at a museum. Due to the time-based segmentation described in Section 2.4, these target trajectories will have an increasing number of segments as time proceeds and visitors move across the museum.

In order to cope with this situation, instead of a monolithic FRC, we generate a palette of FRCs with different number fuzzy sets in their antecedents ⟨FRC1,FRC2,…,FRCmaxseg⟩. Each FRCj comprises the rules to classify trajectories with j segments. To generate each FRCj, we use the first j segment features of all the trajectories trv∈TR.

All in all, the classification mechanism to label the visitors’ ongoing trajectories involves the following steps,

• 1. Each time a new tuple ⟨ai,ti⟩ is appended to an ongoing trajectory trongoingv, its resulting number of segments nsegongoing and its set of trajectory features trsongoingv are calculated.

• 2. Then, the FRCnsegongoing is extracted from the palette of FRCs as it is the one targeting the current number of segments of the ongoing trajectory.

• 3. Finally, FRCnsegongoing is fed with trsongoingv to classify such trajectory with a label CLnseg∈Ω.

Figure 5 shows an illustrative example of the aforementioned mechanism where a trajectory covers different time-based segments and, thus, feeds different FRCs. As the figure depicts, the target trajectory is labelled with different classes ⟨CLs1,CLs2,CLs3⟩ as it evolves and comprises an increasing number of segments. In this case, a visitor’s trajectory grows from one to three segments covering six actions in total. As time proceeds, a different FRC classifier is used at each time. To begin with, the first two actions (a1,t1) and (a2,t2) will fed classifier FRC1 as they fit into the first trajectory segment generating the label CLs1 as outcome. Then, action (a3,t3) enlarges the trajectory until two segments, so the FRC2, targeting 2-segment trajectories, is used. Finally, when actions (a4,t4) and (a5,t5) are received, trv can be split into 3 different segments so FRC3 is used to generate a new classification outcome.

Fig. 5.

Example of the online classification procedure.

3.1.Place infrastructure

The exhibition place comprises two main rooms distributed along a corridor. In the rooms and the corridor, 15 iBeacons Wellcore W902 with chip TICC254111 were installed as BLE beacons.

Figure 6 depicts the floor-plan of the test-bed along with the spatial distribution of the beacons. Such area covered roughly 500 m² and the beacons are spatially distributed every 6 meters approximately. The size of this floor-plan image is 500 × 177 pixels and each pixel represents 0.1 m².

In order to perform the experiment in the test-bed exhibition, we installed the ad-hoc mobile application infoArt in the different guidance devices of the exhibition area.22 This application has two main features,

Fig. 6.

Use case floor plan. The red dots indicate the location of the BLE beacons. The black line depicts the ideal route to visit the exhibition area.

3.2.Location algorithm study

As an alternative to the algorithm to detect the movement actions described in Section 2.3.1, we developed a simple mapping approach along with a calibration mechanism to improve its accuracy in our testbed. This was coined as the MinMax location algorithm. In the next two subsections, we state the main findings of our study with respect the algorithm (Section 3.2.1) and the calibration technique (Section 3.2.2).

3.2.2.Calibration of the beacons

In order to perform the required calibration of the guidance devices, as pointed out above, we took the following steps,

• 1. First of all, we gathered the signal intensity of each of the 15 deployed beacons in the setting with each guidance device at different displacements covering 1, 2.5, 5, 7.5, 10 and 15 meters. For each of those displacements, we gathered a total of 20 samples.

• 2. Next, a third-party BLE library,44 integrated in our infoArt application, determined the distance to each beacon based on the collected signal intensities.

• 3. We averaged the resulting distances per beacon after 5, 10 and 20 samples. This was done because the signal intensity tends to stabilize after several samples [30].

• 4. On the basis of the previously-calculated distances, a polynomial regression model is generated. This model was used by the application to estimate more precisely the distance to each beacon.

Table 2 shows an example of this calibration process for beacon b2 and one of the guidance devices in the exhibition area. The columns are the distance mean of the first 5, 10 and 20 sampled signals received from the beacon by the device and the rows represent the physical displacement from where the samples were taken. This way, at 7.5 meters (the real distance between a guidance device and beacon b2), the distance estimated by the device was 5.15 meters using the first 5 signal intensities from b2 and 6 meters using 10 and 20 samples respectively.

Table 2

Example of the mean distance calculated for beacon b2 for different number of samples and real distances for a particular guidance device

Beacon b2	Num. of measurements
Real distance	5	10	20
1 m	0.66	0.6	0.62
2.5 m	2.67	2.88	2.98
5 m	4	4	4
7.5 m	5.15	6	6
10 m	6.1	6.5	6.5
15 m	10.65	9.65	9

Fig. 7.

Interpolation curve for beacon b2. Where the x-axis represents the sampled distance and the y-axis indicates the real displacement.

In Fig. 7 we plot the measurements of Table 2 as 3 curves (one for each column). Whilst the x-axis represents the measured distance by the application, the y-axis indicates the real distance between the beacon and the device. From these 3 series we generated a fourth-grade polynomial regression model. This model was used by the application to more accurately detect the distance between the guidance system and beacon b2. Lastly, the outcome from this model was integrated in the frames ftv delivered by the guidance device to the LS.

3.2.3.Comparative between map-based and MinMax location algorithms

In order to compare the two indoor location algorithms, we performed a visual analysis of the resulting trajectories with each approach. In that sense, given the ideal route of the use case depicted in Fig. 6, the combination of the MinMax algorithm and the calibration mechanism, produced the trajectory depicted in Fig. 8(a). Alternatively, the Map-Based algorithm generated the trajectory shown in Fig. 8(b). We can clearly see that this second trace is much more similar to the actual route followed by the visitor than the one obtained by the MinMax algorithm.

Table 3

Parameters settings for the FRCs generation

Parameter	Involved step	Value
tseg	Trajectory segmentation	1 m
tend	Trajectory segmentation	5 m
m	Trajectory clustering	2.0
ϵ	Trajectory clustering	0.0001

Fig. 8.

Examples of two captured trajectories. The trace of each one is depicted as a black line. The BLE beacons are shown as small red squares.

The main problem of the MinMax approach was that the distances that the BLE scanner library55 provided us were neither precise nor always repeatable. This made it rather difficult to accurately obtain the location of the devices. Therefore, the MinMax approach gave us poor results, even with the calibration refinement. On the contrary, our two-step Map-Based algorithm (described in Section 2.3.1) performed much better. This is because it does not rely on the measured distances to the beacons, but rather compare the measurements provided at any given time instant t with the ones stored in the previously-generated set LF.

All in all, the MinMax algorithm achieved poorer results in terms of trajectory perception, because of the low reliability of the distances measured from the beacons. Even though we tried to tackle this problem with a calibration mechanism, we still did not achieve suitable results. Therefore, we eventually use our initial Map-based approach to capture the visitors’ movement actions for the rest of the use case analysis.

3.3.Trajectories classification

Once the indoor positioning system was calibrated, the FRCs for the collected trajectories were generated and tested.

3.3.2.Collection of visitors’ trajectories

We collected 64 different trajectories during a 1-week period ranging from 19/04/17 to 25/04/17. Each one was manually labeled by the exhibition operators with one of the four classes previously mentioned. This gave rise to 16 trajectories of each class.

Figure 9 shows the heat map of every captured trajectories composing the TR set. From this figure, we can clearly see the main mobility trends of visitors. In that sense, visitors followed quite similar paths across all the exhibition area. However, in the first room (Room 1 at Fig. 6) we can also observe a little more scattering in the collected trajectories due to exhibition characteristics.

Fig. 9.

Heat map of all the collected trajectories. Pixels with color green have been more frequented by users, followed by yellow tones and finally in red we see unexplored places.

In Fig. 8, we can see two examples of collected trajectories. From these traces, we clearly observe the noisy and imprecise nature of such trajectories. This justifies the trajectory segmentation and feature extraction process to normalize the trajectories representation and the fuzzy logic approach to classify them.

3.3.3.Classifier configuration

Table 3 shows the key parameters used to generate the ensemble of FRCs. We can see that tseg was set to 1 minute. This value was configured that way after observing the average time length of the trajectories per class. As we can see in Fig. 10, all the trajectories range from 5 to 9 minutes on average so splitting trajectories into 1-minute granularity was a reasonable option. In the case of m and ϵ, both FCM parameters were set to common values in the literature [20,31].

Fig. 10.

Average time length of trajectories in TR per class.

3.3.5.Classifier accuracy

Once the FRCs were generated on the basis of the aforementioned parameters, we studied its accuracy. For that goal, we used the F1 score as measurement. This score is calculated following the next formula:

F1=2×precision×recallprecision+recall

where

recall=TruepositivesTruepositives+Falsenegativesprecision=TruepositivesTruepositives+Falsepositives.

Table 5 shows the recall, precision and F1 scores of RECITE. In order to evaluate the suitability of the segmentation approach, the table also shows the results of a FRC (FRCnoseg) that just takes as input a single trajectory segment feature comprising the whole actions of an ongoing trajectory without segmentation. In that sense, we used a 3-fold cross-validation approach to train and test the FRCs.

Table 5

Accuracy scores of the palette of FRCs of RECITE with and without velocity features and a monolithic FRC without including a trajectory segmentation step

Model	Precision	Recall	F1
RECITE FRCs	0.95	0.94	0.94
RECITE FRCs (+velocity feat.)	0.92	0.93	0.92
FRCnoseg	0.84	0.79	0.79

From these results we can see that the trajectory segmentation step clearly helps to improve the classification accuracy. This is because RECITE generates a FRC for each possible sub-trajectory achieving a higher level of specialization and, thus, accuracy.

Furthermore, Table 6 shows the confusion matrix of the evaluation. In that sense, this matrix includes all the intermediate trajectories that users generate during their visits. In general, the main errors occur in adjacent labels. However, 4% of hurry sub-trajectories are classified as lost ones when these two labels are not very similar. In that sense, we have observed that, sometimes, some hurry visitors return to places already seen in their attempt to quickly cross the exhibition area. This makes such visitors mimic the behaviour of lost visitors.

Table 6

Confusion matrix of the RECITE FRCs. A cell in row R and column C with R ≠ C, contains the sheer number and percentage of sub-trajectories whose real label was R but it was wrongly classified as C

Real	Inferred
	Observer	Interested	Hurry	Lost
Observer	610 (89%)	52 (7.6%)	1 (0.1%)	20 (2.9%)
Interested	19 (2.8%)	661 (97.2%)	0	0
Hurry	0	0	415 (96.5%)	15 (3.5%)
Lost	25 (3.5%)	33 (4.6%)	0	659 (91.9%)

The next aspect we evaluated was the impact of the length of an ongoing trajectory on the classifier. The results of this evaluation are shown in Fig. 11.

Fig. 11.

F1 score of the classifier per trajectory number of segments. Each bar depicts the F1 score of the FRCs for trajectories with number of segments equal to its x-axis value.

According to this figure, we can see the F1 score of the proposal was above 0.9 regardless of the number of segments of the target trajectory. However, the accuracy of our classifier was slightly higher for trajectories with segments ranging between 3 and 9. This was because this range of segments comprised most of the trajectories whereas the number of trajectories comprising more than 10 segments was much lower. This resulted in an accuracy drop of the solution.

3.4.New segment features sensitivity study

As it is put forward in Section 2.5, the present version of the framework extracts only three different features from each segment, namely the average x-y coordinates and the average number of content displays. Consequently, we also evaluated the effect in RECITE of adding new segment features as part of the framework, namely, the average segment speed and the average segment bearing.66 This way, we include the velocity of the trajectories so as to perceive in a more detailed manner the actual movement of the visitors in the museum.

Bearing in mind the definition 3, each trajectory segment feature sjv is now a vector ⟨xs,ys,ncs,ss,bs⟩ where ss and bs are the mean speed and bearing of the segment sjv.

As Table 5 shows, including these features in the solution did not actually improve the accuracy of RECITE with respect to the baseline approach. This way, the F1 score was slightly lower (0.92 vs 0.94). Regarding the confusion matrix depicted in Table 7, we can see the velocity features slightly improved the detection of lost trajectories as their classification rate increased from 91.9% (see Table 6) to 93.7%. This type of trajectories are defined by many changes of direction due to the roaming behaviour of the visitors. As a result, the bearing features help to better identify these types of movements. Nevertheless, this rate actually dropped for the other three types of trajectories.

The reason of this lack of actual improvement of RECITE is due to the well-known curse of dimensionality problem [27]. Adding more features to the target input space makes the trajectories segments more distant among them. As a result, the approach generated much more clusters and, thus, fuzzy rules as we can see in the Table 8. According to this table, the system generated, for example, 8 different clusters given the first segments of the trajectories labelled as CLlost. Nonetheless, when these segments are processed by the FCM algorithm without considering the velocity-related features only one cluster is detected (see Table 4).

Finally, FRCs are easily interpretable as they are composed of IF-THEN sentences. In this context, a drawback of enlarging the framework with new input features is that the complexity of the FRC increases and the aforementioned descriptive capability of this type of fuzzy systems is lost.

4.1.Indoor positioning systems

To begin with, RFID readers have been widely used in this context so as to locate visitors holding different types of devices [41]. In that sense, approaches can be distinguished depending on the requirement of an explicit user check-in [22] or not [23,24].

Some works have also proposed WiFi access points as indoor positioning enablers. For example, [14] combines the signal strengths of WiFi routers and the outcome of an image-processing engine able to detect the presence of visitors to come up with a precise indoor positioning solution. Ambitrack [9] also proposed tracking through cameras and image-processing. However, this approach is more expensive to install and maintain than the inexpensive BLE beacons. An interesting approach to locate visitors in an art gallery is introduced in [12] where a combination of BLE beacons along with image recognition mechanism integrated in a wearable device allows to detect whether a visitor is in front of any art keypoint of a gallery. However, in most cases BLE is used to detect presence of visitors at much more large scale (e.g. galleries or corridors) [13,26,32,38,39].

Our work also makes use of an ad-hoc indoor positioning system relying on BLE. However, unlike other solutions we consider a careful calibration process performed at device and global level based on polynomial regression. Thus, we are able to accurately locate the visitor making use of affordable beacons as described in Section 2.3.

4.2.Trajectory modelling

Regarding a building-based approach, many solutions just perform a simple mapping step considering the concrete building space where the sensors is actually located to construct the visitors’ paths. Examples of this are [38,39] that model trajectories as sequences of visited galleries. However, [15] relies on a museum graph model that accurately represents its premises, like doors, rooms or vitrines. Then, a mapping approach is performed to convert the collected spatio-temporal data from users to the graph-based representation. However, the proposal does not define any particular indoor positioning system to capture the visitors’ movement. A similar approach is followed by [1,26] where a semantic indoor trajectory model is proposed or [14] with a more holistic ontology.

POI-based approaches have also been widely used in the literature [12,22,23]. In some cases, this type of models are enriched with additional information. For example, [22] merges the visited POIs along with the digital content consulted by the user in the museum displays to compose the set of features that defines the user’s movement.

Furthermore, it is interesting to mention the work in [13] where the building bricks to compose user trajectories are the named coverage areas of the Bluetooth beacons. Thus, it provides an intermediate solution between POIs and building elements to represent the visitors’ trajectories. Similarly, [32] considers both POIs (artworks) and building elements (rooms) to compose the trajectories.

Our work is enclosed on an alternative way to model trajectories based on a two-dimensional Cartesian space. In this domain, only a few works have actually relied on this representation. For instance, [24] makes use of this technology to calculate the probability of a visitor’s position in a map image representing a gallery floor. The RECITE model leverages the accuracy of the underlying positioning system. As a result, we have been able to perform certain computations from the outdoor trajectory mining field like the segmentation step discussed in Section 2.4. Moreover, our work enriches this model by labelling certain segments of the trajectory with the content displayed by the user in his guidance device.

4.3.Trajectory analysis

Concerning the analysis of the collected trajectories, there is a wide range of solutions considering both the applied methods and their purpose. To begin with, some examples can be found where trajectory pattern mining has been carried out by means of simple frequency counting methods to discover the most frequent sequences of visited galleries [39]. In [13], authors pursue the same goal but, in this case, sequence alignment methods are applied over trajectory data. In [38] trajectories are characterized using the random walk model.

As far as trajectory clustering is concerned, many different algorithms has been tested. For instance, [24,32] make use of the k-Means clustering algorithm to aggregate visitors’ trajectories and, thus, uncover general movement trends. However, [24] performs a spatial partitioning by means of the state chain model before applying the clustering algorithm. In [23], authors perform a clustering task over the set of trajectories and map each cluster to a set of predefined visitor profiles. Later on, two sequence miners are applied to each group in order to reveal interesting patterns defining visitor behaviors.

Other works have used optimization algorithms like A∗ to provide navigation services to visitors in order to find the best route between two rooms in the museum [14].

In addition to that, some works just provide a preliminary design of the trajectory analysis that might be carried out. For example, [15] proposes to identify stay-points of visitors’ routes and then performs an incremental clustering over such points using time windows to detect behaviour changes during consecutive time intervals. Another worth mentioning work is [26]. It provides a semantic reasoning mechanism that, by means of a set of domain-dependent predicates, it is able to infer the purpose of visitors’ movements (e.g. visit a particular gallery or buy something at the gift shop).

In the trajectory classification field, we have noticed a scarcity of proposals within the cultural domain. Thus, [22] makes use of two well-established classification algorithms like logistic regression and multi-layer perceptron to rate the suitability of an unseen POI recommendation based on the previous movement of a user. For that goal, some profiling information of the visitor like age or gender is considered.

In this context, the work at hand provides a set of FRCs that smoothly merges the spatio-temporal data from visitors displacements and data related to the content displayed in the guidance devices to timely detect the visitors behaviour. To do so, it does not rely on any sensitive profiling data from users.

We should also mention that our work share some similarities with the proposal stated in [11]. In that work, a FRC-based mechanism to classify trajectories based on Volunteer Geographic Information (VGI) is described. However, VGI provides a more coarse-grained representation of the target users’ movement than in our setting. Consequently, a segmentation step is not necessary in that work. Furthermore, in such a work, only the spatio-temporal features of the trajectories are considered for the classification task whereas RECITE also considers other factors like the usage of the multimedia content made by visitors.