Indoor tracking, positioning, localization, and monitoring systems can be described as a collection of networked devices used for locating or tracking objects. This includes people inside buildings, train stations, airports, underground garages, and other locations sharing aspects of systems to detect violation of distancing measures. These systems are related to this study’s design aim.

Different types of indoor localization applications exist that utilize various sensor and communication technologies. For example, according to Jamil et al. [12], in the domain of e-health a localization system for monitoring elderly people was introduced. Two kinds of communication technologies (UWB and BLE) were used with two algorithms, hybrid localization and a TDOA-based algorithm (time difference of arrival), to enable the evaluation of elderly behaviors, monitoring of health status and wellbeing, and detection of emergency situations. Further, an economical design of wearable devices for indoor localization and motion analysis using data fusion algorithms is presented in zhang et al. [15] as another example. An interesting integration between an inertial measurement unit and UWB localization was introduced using a Mahony filter and quaternions as a solution that achieved high-accuracy, signal stability, and motion tracking capability.

Technologies and techniques related to outdoor monitoring and tracking approaches were omitted from the review conducted for this study. These approaches are considered more complex than required for the current study’s aim and normally involve more than one communication interface attached to the utilized IoT devices to cover both short- and long-range communication [11,18,20]. At this stage of development, introducing a detection system of distancing measures with wide coverage capability for highly movable objects is not part of our aim.

During the review, we found four main alternative technologies were used when designing wearable devices for indoor localization systems, namely, RFID, BLE and UWB, smart devices [19–21,24–26], and visual detection cameras. These alternatives strongly relate to the kinds of communication technologies used in the systems. As the basis of these technologies, a WSN was found a suitable communication technology adopted in the systems reviewed in this work. According to Nikodem et al. [11], for instance, detection cameras for monitoring road traffic and detecting and classifying vehicles were utilized and connected to a WSN. These cameras were installed on the sides of roads to track and monitor vehicles. In another study, wearable devices equipped with cameras and headsets and connected via a low-power Wi-Fi module could recognize paintings in a museum from the image captured by the camera using computer vision techniques and transfer the captured data to audio [14].

In addition, accuracy and energy consumption are critical factors that must be considered when designing wearable IoT devices and selecting their intercommunication technology. BLE is used to address energy consumption issues in many IoT applications [13,17,18,22]. However, UWB technology provides a high accuracy WSN for IoT systems [14,15]. In Kolakowski et al. [13], for instance, a balance between these two factors was discussed for the proposed system. UWB- and BLE-enabled devices were utilized in the proposed radio positioning system. UWB consumed higher energy during packet transmission than did BLE, although BLE provided more accurate detection results. A hybrid algorithm that offers the benefits of both technologies was designed and used for object detection and tracking [13].

In addition, a real-time intelligent crowd engineering analytical approach using mobility characterization and energy-efficient BLE to enhance mobility characterization and signal analysis (ICE-MoCha) was proposed in Jabbari et al. [27]. Crowd individuals wear the BLE tracker bracelets to manage the power consumption on the bracelet. Each tracker bracelet has a unique identification code for identifying each pedestrian in the crowd. The communication system of ICE-MoCha supports various messaging techniques for sending and receiving messages or probes from or to the tracker bracelets based on type: active or passive. Generally, the approach works to translate the BLE beacon signals from various transmitters into group semantics for tracking the status of the crowd and predicting potential accidents using a smart video surveillance technology.

Further, RFID-based wearable devices were utilized in several IoT indoor localization systems [5,16,22,23]. RFID technology can transmit wearable device IDs and data captured by the attached sensor to an RFID reader over a range of about 100–200 m with low power consumption because it works on radio frequency signals [25–28]. The RFID-based devices are normally attached by RFID tags. Each tag consists mainly of two major components: an electronic chip for storing data corresponding to the individual object attached to it and an electronic antenna for transmitting sensed data to the associated readers [29].

For instance, the research discussed in Zhang et al. [22] introduced an active RFID-based indoor locating system that uses a frequency-hopping technique to reduce the noise of the captured RSSI fingerprinting data. Two kinds of low-cost tags, with a tag–tag communication protocol, are utilized: virtual reference tags and object tags. The reference tags are distributed in the covered area to improve accuracy, whereas the object tags present RFID coordinates to RFID readers. Further, an RFID-based 3D indoor tracking system for eHealth was presented in Paolini et al. [23]. The study introduced a customized low-cost reader that is remotely controllable and supported by real-time 3D scanning features to detect the dynamic location of a patient within an indoor area. The algorithm developed retrieves the reader–tag distance and its height. The approach demonstrated high accuracy detection in the tags’ height, with an error rate below 20 cm in all the simulation experiments conducted, in which the average error of 10 experiments performed was less than 18 cm for distances up to 4.50 m [23].

Various sensors and existing communication technologies for indoor tracking and positioning systems were reviewed and investigated based on four main factors: energy-saving capability, accuracy, costs, and coverage range. The following table (Tab. 2) summarizes each technology considered in the review and its features regarding these four factors.

Graph theory played a significant role in developing collective adaptive systems (CASs). CASs can be defined as complex systems with a massive number of heterogeneous entities that interact with each other without a central control to achieve an individual or shared goal. In [30], a model of pedestrian movement was introduced as a CAS using an expressive process calculus language (CARMA). The model is based on graph structure representing a network of pedestrian paths. Each individual entity, which represents a pedestrian, acts as a decision-making element when it produces probabilistic choices about the next movement in the future and how that affects other pedestrian movements. Pedestrians are expressed as graph nodes in the model, whereas the current paths of each pedestrian are expressed via graph edges and determined by the pedestrian’s current and next position.

In Allam et al. [5], the design of a wearable wireless sensor is presented to ensure crowd safety in high density situations. The device communicates with other devices in its sensing range to send simple warning and safety messages or status notifications across devices. This approach is based on the theory of graph density, in which each device in a crowd counts the number of surrounding devices and then calculates the density around it. Consequently, the previously determined threshold is examined to decide whether the device is in a safe situation. If not, an appropriate warning message is propagated to other devices within range, and guidance and instructions through messages are provided to prevent overcrowding and dangerous situations. The technique is based on an existing agent-based crowd simulation model called Multi-Agent Simulator of Neighborhoods (MASON), which is written in Java and used in 2D environment simulation tools.

Additionally, graph theory applications have a long history in the domain of crowd analysis, monitoring and control. In Matthias et al. [29], for instance, a graph-based framework for detecting and tracking pedestrians, dense crowds, and events is proposed using hidden Markov models. The framework requires a short sequence of images for effective pedestrian tracking and location of possible danger areas in a crowd. All results are implemented and demonstrated using a simulation model.

Further, in Angela et al. [7], a graph-based approach obtains indoor dense locations from tracking data using the semantic location. The dense location queries, processing, and indexing are performed using a proposed technique called the Dense Location Time Index, which combines the information of locations with time points. In another approach [8], a decision graph finds the cluster centers in the proposed pre-screening method. This method is presented to the density-based clustering algorithm to obtain possible shapes of the clusters and the internal density between each point in the graph [8].

Moreover, Xu et al. [9] introduces an approach for analyzing and recognizing crowd dynamics based on a static proximity graph. The proximity graph data structure represents crowd textures as graph vertices. Spatial closeness relationships between persons in the crowd are recognized in terms of their texture over time. The graph is supported by annotation extension attached to the vertices to add further semantics representing crowd context. Thereby, it provides a precise picture for the decision maker. Wearable devices on graph individuals collect information and are connected via a Bluetooth-based wireless communication technology and utilized as a network infrastructure.

In previous research, a substantial number of real-world problems are represented mathematically and formalized using either directed or undirected graph theory. The cyber security knowledge (directed) graph and route planning (directed) graph presented in Jia et al. [31] and [32], respectively, are examples of the variety of applications of graph theory. In this study, it is assumed that the direction of each individual is not vital information in the model. Thus, the (undirected) complete weighted graph model is considered a theoretical base of this work.

In this section, the notation of first-order predicate logic (FOPL), with extensions for equality ( = ), and membership ( ∈ ), is employed to express the mathematical foundation of the graph concepts used in the proposed system design. A syntactic sugar notation is used for a unary predicate to assign a type to a variable, for example, Graph(g) is re-expressed as g:Graph . Further, the notation ( , ) is utilized to shorten and replace the logical AND(∧) operator.

In the proposed FOPL formalization system, there are four basic types of variables, namely Graph,Vertex,Edge,Attribute, and Value . Each variable is utilized to represent concepts of graph theory in the system. The following formal definition describes the notation of the computational graph model adopted in this work:

(1) AGraphisatupleG=(V,E,P,ðŒ,t)

V :A finite set of all vertices in graph G . Vertices correspond to individuals within a specific place who form a crowd. A vertex is expressed using the following formula:

(2) ∀v,u:Vertex⋅(v,u∈V)

E :A finite set of all edges in graph G . Edges are undirected and correspond to the actual distances between the graph vertices. This is expressed in the formula:

(3) ∀e1,e2:Edge⋅(e1,e2∈E)

The notation evu is used to represent an edge between two vertices v and u , where v and u are called the endpoints of evu .

Pv : A set of all attributes that belong to the value of a vertex v expressed as a pair of keys. Similarly, some edges may also have attributes in the model. The notation Pevu is used to represent the set of all attributes of the edge evu . The pair of keys’ values is described by the following formula:

∀i,j∈{1,2,…,n},i≠j

(4) ∀v:Vertex,∀keyi,keyj:Attribute,vali,valj:Value

∃Pv⋅(Pv={ { keyi,vali},{keyj,valj}})

w : When modelling a distance between individuals in the crowd, it is important to determine a weight function 𝌠, where R is a real number associated with each edge e in the graph. This is represented via the following expression:

ðŒ:E→R

(5) ∀v,u:Vertex,evu:Edge,g:Graph⋅(v,u,evu∈g)∧(evu∈E)

∃(x∈R)⋅ðŒ(evu)=x

The weight is computed using the traditional distance formula as:

(6) DISTANCE=u2−v2

where v and u are two points that have x and y coordinates v(x1,y1)andu(x2,y2) .

Unlike other crowd modelling approaches that consider including all edges in the model, only edges between individuals that reflect the social distancing rules for reducing the transmission rate of COVID-19 are included in the current study. Thus, the range of distance considered in the approach falls between 0 m and 4 m. Thus, it is unnecessary to represent all of the connections between individuals in the crowd that have distances (weights) over 4 m. Function t is used to indicate that two individuals (vertices), v and u in the model, fall within the range of interest for evaluation of COVID-19 physical distancing measures. Therefore, edge e links the pair v and u . This is expressed using the formula:

t:E→(N×N)

(7) ∀g:Graph,(v,u):Vertex,e:Edge⋅(v,u,e∈g)∧t(e,v,u)

The convention evu is adopted to represent each edge that connects the pair of vertices (v,u) and appears in E for simplicity, where v and u are said to be the end vertices of an edge evu . The following formula describes this concept of the model:

∀v,u:Vertex,

(8) ∃evu:Edge⋅t(evu,v,u)∧(evu∈E)

The graph model of the crowd in a determined place is created by locating each individual as a graph point (vertex) within a specific area at an exact time using appropriate technology, for example, BLE. Each graph vertex, in the real-world set, has attributes to store the intersection of latitude and longitude lines, which represent the coordinates of that vertex. The coordinates are used to identify the exact location of the vertex in the place.

The proposed model focuses on modelling the distance between crowd individuals to observe the violation of social distancing rules. The distance (in meters) between each pair of vertices is calculated to be the weight of the undirected edge between that pair. Times and dates are other possible pieces of information (attributes) that might be assigned to each edge.

The type of graph model based on a related technology for capturing vertex locations is a complete undirected weighted graph. Graph G is considered a complete graph if there is a unique edge connecting every pair of its distinct vertices. If G has n vertices the graph is called Kn and the number of its edges is n(n−1)/2 .

Mathematically, graphs can be represented in several ways, such as via adjacency lists and adjacency matrices. The adjacency matrix is considered one of the graph representation alternatives adopted in this work. To create a corresponding adjacency matrix A of the complete graph Kn , suppose the total number of vertices is n ; then, the definition of the adjacency matrix A can be expressed with the following formula:

∀i,j∈{1,2,…,n},i≠j

∀v,u:Vertex,evu:Edge⋅((v,u,evu∈Kn)∧(w(evu)=x)

(10) →

[A]i,j=w(evu)

Managing a massive amount of stored data (big data) requires a type of data model that can support its characteristics, namely, volume, velocity, variety, and veracity. These characteristics are not fully supported in the traditional relational data models. However, owing to the dynamic nature of the NoSQL data model, it is suitable to deal with critical issues of big data that are related to its characteristics, such as efficiency, scalability, and availability [33,34]. A graph database, as a form of the data model, has demonstrated the ability to express and manage connected sensor-based data in various IoT systems [35,36], and is an example of a dynamic big data schema.

This study introduces a Neo4j graph database implementation of the complete weighted graph data model, discussed previously in Section 3, to model the violation of distancing regulations within a particular place. Neo4j is one of the most popular Java-based open-source native graph database systems and is used to implement the property graph data model efficiently. In the Noe4j graph model, data are structured as nodes, relationships, and properties attached to the nodes or relationships. It provides robust transactional storage of big data and supports full ACID features (atomicity, consistency, isolation, durability) to maintain the integrity of stored data. Neo4j also supports deep graph analysis and querying using a query language named “Cypher” [37]. It is a reliable graph database system for large and rich graph modelling of crowds. Consequently, Neo4j is considered an appropriate implementation platform for the proposed weighted graph model presented in this paper.

Noe4j’s declarative query language, Cypher, is similar to SQL but specialized for graphs. The Cypher language is considered a simple and easy to learn language that allows users to store, retrieve, and manipulate data stored in a graph database. The Cypher syntax is declarative and is expressed in a logical way for matching patterns of nodes and relationships in graphs. For example, the keywords “MATCH/ RETURN” work similarly to the “SELECT” statement in SQL to project nodes or relationships that satisfy given constraints. Additionally, the keywords “CREATE/ MERGE” are used to create new nodes in the graph, similarly to the “INSERT” statement in SQL. The following figure (Fig. 2) provides an example of a complete directed weighted graph consisting of 11 nodes and 54 relationships.

Because the output of the proposed transformational system in the middleware layer, discussed in the following subsections, is an executable Neo4j Cypher query script file, the Noe4j database engine can run Cypher script files to construct, or update, the graph representation of the real-world group in the system. Then, several predefined queries that support decision makers might be executed on the stored graph to retrieve insight results. The following snapshot of the Cypher script (Listing 1) illustrates the executable code used to generate three graph nodes, A, B, and C, using CREATE statements, in addition to three undirected and weighted relationships with a randomly computed distance weight between them, using MERGE statements.

Nodes of the Neo4j crowd graph represent people (individuals) in a crowd, whereas undirected and weighted edges represent the actual distances between individuals at a corresponding time point. Thus, the graph created at a specific time interval represents the status of the social distancing rule application at that given point in time. In the real world, the person node holds properties to represent the location of an individual who wears a sensor or is attached to a communication device.

Each person node in the crowd graph is located near another person node within a particular distance. The distance between them is represented in the graph by a relationship between two person nodes. Each distance relationship has a calculated weight property to represent the measure between each pair of individuals. The weight of each distance is computed using Formula 7, as mentioned previously in Section 3.1 and Section 3.2.

This layer consists of wearable devices connected to a wireless network. In this layer, a WSN (e.g., Wi-Fi and Bluetooth, Zonal Intercommunication Global-standard [ZigBee], and fifth generation cellular wireless [5G]) and a suitable tag-based detection technology were combined to collect required information about physical distancing between individuals during a corresponding time interval. Thus, the physical distancing was tracked using wearable devices or tracker badges and transmitted to the nearest edge point for analysis and processing.

As a design choice, Bluetooth smart badges were considered IoT wearable devices. These cost-effective badges are based on BLE technology, which facilitates short-range communication and offers an adequate level of indoor tracking accuracy. Additionally, beacon devices were used to expand the Bluetooth coverage range of the closed area because the standard beacons have an approximate range of 70 m. Thus, based on the dimensions of the closed area, the number of required beacons could be determined. Fig. 3 shows the Bluetooth network infrastructure using multiple BLE beacons and cloud beacons that were connected to the proposed IoT edge-based cloud system.

Because one of the objectives of the proposed design was to introduce a social distancing tracking solution, cloud beacons, which have the capability of data backup, configuration, and application of updated settings to all other connected BLE beacons, were employed to monitor other BLE beacons and send their collected data over the Internet to the cloud backend data storage. They were also used to increase the collection of distancing data for more than one closed area. Owing to the Wi-Fi connectivity feature available in cloud beacons, they can upload data to the edge server and then to the centralized cloud storage over Wi-Fi. It is worth mentioning that the range of Wi-Fi networks can be up to 200 m greater than the standard Bluetooth range.

As a result, each person approaching the area covered by the proposed physical distancing monitoring system is asked to wear the Bluetooth smart badge assigned by a previously configured unique ID to sense and report distance changes within the group to the decision maker via the wireless network infrastructure. The assigned ID of each badge represents a specific location and direction (if required) of an individual within the group inside the closed area. The approximate distances within the group of individuals and a timestamp are detected and transferred via the Wi-Fi connection to the edge point and then stored in the backend data storage.

This approach also considers information security issues and individual privacy since the distributed Bluetooth badges and the cloud beacons only monitor and record anonymous data at each point in the area, excluding any information that can identify an individual. It should be noted that detailed specification regarding algorithms and techniques used for detecting distances between individuals is outside the scope of this paper.

The collected and processed information in the middleware layer is stored in the system using a centralized cloud database that supports big data storage, future processing, and retrieval. The stored data might be used for real-time or post processing, archiving, querying, analysis, or prediction.

The graph-oriented database system was selected as a backend data model to gain the benefits of the graph data structure. This is because of the nature of the sensed data stored in the database and the capability of graph databases to perform efficient query processing and deep graph analytics. Importantly, the data storage is physically located in the cloud storage as a centralized cloud database.

To enable the system to calculate the degree to which the existing group is violating the social distancing regulation, the number of edges, with distance values < = 2 m, is counted and compared with the total number of edges between group individuals (with distance > 2 m and < = 4 m). Additionally, the number of individuals involved in the graph is considered to examine the size of the graph. In the proposed proof of concept, this can be achieved using several simple Neo4j Cypher query statements:

Q1: What is the number of attached nodes in the graph?

MATCH (n:person)-[r:DISTANCE]->(m:person) WHERE toFloat(r.weight) < 4.0

RETURN COUNT(n)

Q2: What is the number of the considered edges in the graph?

MATCH (n:person)-[r:DISTANCE]->(m:person) WHERE toFloat(r.weight) < 4.0

RETURN COUNT(r)

Q3: What is the number of edges with no violation of distancing measure in the graph?

MATCH (n:person)-[r:DISTANCE]->(m:person) WHERE toFloat(r.weight) >= 2

AND toFloat(r.weight) < 4.0

RETURN COUNT(r)

Q4: What is the number of edges with mild violation of distancing measure in the graph?

MATCH (n:person)-[r:DISTANCE]->(m:person) WHERE toFloat(r.weight) > 1.40

AND toFloat(r.weight) < 2.0

RETURN COUNT(r)

Q5: What is the number of edges with high violation of distancing measure in the graph?

MATCH (n:person)-[r:DISTANCE]->(m:person) WHERE toFloat(r.weight) > 0.01

AND toFloat(r.weight) <= 1.39

RETURN COUNT(r)

Based on the results returned from these queries, simple calculations are required to allow the decision support system to generate an appropriate message to send to the decision maker:

The percentage of individuals who are involved in the group:

% involved = (total nodes - no of nodes out of the group) / total nodes)*100

Two main steps are considered for creating the source model. The first step is to construct a complete graph structure on the fly by gathering all the sensed data, captured at a specific time from IoT sensors, and manipulated as a density matrix in memory based on edges. The second step is to create an eXtensible Markup Language (XML) memory model of the captured complete graph. During the process of building the XML memory model, all nodes that are outside the range of distancing violations are eliminated. This XML memory model implementation [37] represents the simplified matrix and is considered the source of the translation operation in the form of a graph structure.

A common method to represent and parse XML memory models is to use a Document Object Model (DOM) or Java Document Object Model (JDOM) package. Moreover, options other than XML might be used for data formatting and serialization standards, namely JavaScript Object Notation (JSON) and YAML Ain’t Markup Language (YAML). XML has the advantage of describing the structure of data using a human-readable and friendly format and it has a flat learning curve. In addition, the XML format relies on a standard means of representing models using the XML Metadata Interchange (XMI) format in the model transformation strategy adopted in this work, as explained in the following subsection. The following algorithm (Listing 2) describes the rules of translating the complete in-memory graph model into the simplified intermediate source graph model of the core transformation step.

This output source model is considered an intermediate textual domain specific language (DSL) for describing the structure of a graph using XML and conforming to a defined metamodel of a source model (Fig. 5). Listing 3 provides a snapshot of the graph model produced. It is important to note that constructing an intermediate representation, using a language similar to the structure of graph databases, is considered a critical step in the design because it leads to simplifying the process of generating the executable Neo4j Cypher in the transformation chain.

To present transformation rules applied to the final executable Neo4j Cypher queries for the intermediate XML graph model, the ATL language was used to express each matching rule of this transformation component in the middleware layer [38]. ATL is a hybrid model transformation language that contains a mix of declarative and imperative constructs to describe rules of transformation in a simple manner. One of the main reasons for adopting the hybrid approach of model transformation is to simplify mapping rules to be easier to read and understand, which is crucial in larger transformation modules from a software engineering perspective.

The transformational approach of ATL is based on pattern matching, in which a pattern of a source model, expressed using XMI standards, is evaluated and transformed using a set of declarative ATL matching rules into a pattern of a target model, also expressed using XMI format. The approach maintains traceability links between elements in a source model, rules, and the created elements in the target model. The imperative side of ATL can be seen when the developer needs to encode or customize control flows explicitly for complex parts of transformational algorithms. These imperative definitions of mapping rules are used to support the declarative mapping rules in the transformation module. This is implemented in ATL using helper functions, as shown in Listing 4, which demonstrates the definition of two ATL declarative matching rules, Vertex2CreateNodeStmt and Edge2MergeStmt, used to transform the intermediate graph XML model into an executable Neo4j Cypher script. In each rule, one-to-one mapping statements are declared. For instance, each vertex v is mapped into a generated CREATE Cypher script statement for creating a new Neo4j graph node with the same properties of v. Additionally, to specify the concrete syntax of the generated script, helper functions are utilized to express the detailed syntax of the target script. A detailed description of the syntax and semantics of the ATL language are outside the scope of this work but can be found in Jouault et al. [38].

As proof of concept, a discrete-time simulation model is adopted. It is assumed that the changes in distances between individuals occur only at each discrete time tick. Consequently, a simple algorithm for generating multiple weighted graphs with random distances between their nodes is implemented using Noe4j Cypher script. The Neo4j built-in random function rand() is used to generate random weights for all edges in each complete graph to represent distances between individuals rather than to detect real-world locations of individuals inside the store at a specific time. In the simulation, 21 weighted graphs are generated randomly. After inspecting the weights on their edges, all edges with distances of > 4 m are eliminated from the graph and their end-point nodes are considered unattached. Listing 4 illustrates the Cypher script used to return the updated graph after eliminating edges with weights > 4 m.

The simulation then answers the five queries, presented in Subsection 4.3.1, for each of the generated graphs (Tab. 3). The Neo4j 3.5.14 database system is used to run the script, thus producing 21 graphs with a total of 176 nodes, 1040 properties, and 864 relationships, completed after 103 ms. Fig. 6 provides a snapshot of the creation of the weighted graphs used in the simulation.

A second round of inspection is applied, and it is noted that the simulation results showed three different cases. These cases can be categorized based on the number of nodes falling in the distancing range as no violation, mild violation, and high violation cases. Generated graphs belonging to the no violation case represent the possible potential ideal situation, captured by the system at a specific time tick, when there is no detected violation of social distancing measures inside the store; that is, the majority of distances between individuals range between 2 m and 4 m.

Graphs belonging to the mild violation cases demonstrate a situation where some minor or moderate violations of the distancing measures in the store are detected. The group of persons has a small number of nodes falling in the range of 1.4 m and 2 m, whereas the remainder fall in the greater range. However, graphs belonging to the high violation cases show major violations of distancing measures captured by the system. A larger number of nodes violates the social distancing rules inside the store, with distancing ranges starting from 0 m to about 1.39 m.

According to the simulation results demonstrated in Tab. 3, it is possible at this stage of system development to detect the distancing measure violations for a group of individuals using simple graph database queries, rather than machine learning and computer vision techniques. From the 21 randomly generated graphs, the system detects the class of violation and its degree, as demonstrated in the following chart. Fig. 7 summarizes the results showing the difference in violation of social distancing measures for each case. The simulation produces results of 10 high violation, 8 mild violation, and 3 no violation cases. For simplicity, a selected graph from each category, represented in Tab. 3, is considered in the following subsections to demonstrate in more detail the simulation results of Graphs 4, 11, and 21.

Before discussing the scalability details of the proposed design, it should be recalled that the main objective of the system is to detect violations of COVID-19 physical distancing measures. To evaluate scalability, a larger test case to detect distancing violations of a crowd of people (121 persons) was designed. The Neo4j 3.5.14 database system was utilized to run the Cypher script, shown in Listing 5, to produce a larger graph with 121 nodes and 2020 relationships (edges) in 176 ms.

Tab. 7 shows the number of generated elements of the graph, including the number of created nodes and the number of created or matched edges. It also lists the time the system took to assess the graph structure and edges with weights > 4 m.

The concept of connected graphs was considered in orchestrating the generation of edges between nodes to make this case more realistic. The 121 nodes were distributed into five groups of nodes, each of which represented a subgraph. Then, edges were defined between nodes in every pair of neighboring graphs. This strategy is illustrated in Fig. 11.

Like the strategy of assessing graphs in the original cases discussed in Section 6.2, the predefined queries Q1–Q5 were considered to determine the level of distancing violation. For simplicity, Tab. 8 shows the triples of Group1 and Group2, including their connected nodes and edges. These triples are presented following the format: (NODE EDGE[weight] NODE).

Based on the results shown in Tab. 7 and Tab. 8, no problems are apparent with the horizontal scaling of the proposed system design. As with any expansion of the graph size, the system is able to record all of the triples to be assessed using the predefined queries Q1Q5 and provide the percentage of violations to output an appropriate message to the decision maker, as shown in Tab. 9.

Overall, the proposed approach employed various utilizations of IoT devices, edge/cloud computing, and BLE beacons for indoor localization, detection, and monitoring. Consequently, the main objective of the system design presented is achievable. An advantage of the proposed design is the ability to use the recorded data in the closed area to analyze and understand group behavior inside that area using only simple graph database query processing. For deeper analysis, the Neo4j Graph Data Science library introduced by Neo4j enables data scientists to make valuable predictions.

Another advantage is that the proposed design can be considered vertically scalable by increasing the number of IoT devices, including BLE beacons and Bluetooth badges, to accommodate larger crowds and coverage areas. A potential extension of the proposed system architecture is the adoption of smartphones or smart watches as alternatives to Bluetooth badges because all smart devices are currently equipped with Bluetooth connectivity. This would enable the design presented to be expanded to include, for instance, a mobile application installed in smart devices to notify individuals about the status of their current location, which could also apply to outdoor areas.