4. Experimental Design
This section discusses the experimental design related to a cleaning risk use case as the most important prevention activity in post COVID-19 workplace reopening with the use of the IoCT as a multi-sensor platform. To effectively integrate the multi-sensor system for cleaning risk analysis, a multi-criteria evaluation [50] was applied to identify, and rank COVID-19 risky behaviors based on an available multi-sensor system in the CCIT building at the University of Calgary campus. Since 80 percent of the data used by the proposed IoCT system was geospatially related, Spatial Multi-Criteria Decision Analysis (SMCDA) provided a superior framework for a variety of decision-making situations [51,52,53].

4.1. COVID-19 Risk Assessment Using IndoorGML
The SMCDA simultaneously represents decision spaces as well as criteria values based on attribute and geographic topology [50]. For this research, topological relationships from the OGC IndoorGML dual graph were used for risk aggregation for the multi-sensor system. A scientific SMCDA process can be put in place using the different steps shown Figure 6. In order to initialize the decision-making process for this paper, equal weights for various risk criteria map layers were considered. This helped ensure fast implementation and quick proof of concept.
The main step for risk criteria assessment was determining the factors affecting the risk of COVID-19 spread based on information from existing studies from both the World Health Organization (WHO) [27] and the Government of Canada website [28]. The number of active virus particles present in a place was considered the most important factor for determining the risk of infection [27]. Various transmission ways of SARS-CoV-2 transmission include airborne transmission caused by small droplets, and larger droplet transmission (droplets can survive up to several days on different surfaces) [54]. The term “viral load” will also be used to refer to the number of active virus particles present in a space. Virus particles live for different lengths of time, depending on a number of factors, the most significant one being surface material. Risk of infection for any particular IndoorGML cell space was modelled as the viral load within the space.
Assuming that a proportion of any average group of people is infected, the viral load within a space increases along with the number of people occupying it, the amount of time the people spend in the space, and the actions of the people within the space. Talking loudly, exercising, and coughing expel more droplets into the environment than other activities, and thus increase the viral load within the space. The virus also passes from surface to surface through touch, so touching surfaces without cleaning hands in between also increases the viral load within the space. The viral load was broken down into a hierarchy of smaller cause similar to a root cause analysis in order to evaluate the different factors. The following layers represent the respective criterion maps. Effective parameters were identified based on available sensors and data according to the implemented IoCT multi-sensor system. The viral load risk criteria are listed as follow:
C1: Risk from Cleaning: Cleaning schedule reported on a smartphone app based on the time that had elapsed from the previous cleaning. For this paper, the cleaning frequency for each room was every 6 h, meaning that after six hours the risk is maximized at one whereas immediately after cleaning it is at 0. C1 is a spatiotemporal map layer comprised of OGC IndoorGML cells with values between 0 and 1.
C2: Risk from Contact Tracing: Proximity tracing map extracted from beacons which includes a trajectory map of traced people on an OGC IndoorGML graph. These trajectories show the location of the cleaner and the number of people in a place. If a person identifies himself or herself as a COVID-19 infected person, the historical trajectories can be used for the contact tracing map layer calculation.
C3: Risk from People Density: Gathering restriction map from smart cameras which includes the number of people over each IndoorGML node. This value changes over a range of 0–1 based on the number of people divided by the capacity of the room (which can be assigned or generated from the area property of an IndoorGML cell node). This information is reported online and aggregated once the room is cleaned.
C4: COVID-19 Risky Behaviors: Risky behavior violation map which includes the number of incidents or violations (number of people violating social distancing, hugging, people touching common surfaces and objects, talking loudly, exercising, coughing, and sneezing). This value is a weighted average of the risky behavior factors (number of violations) over the frequency of cleaning (normally six hours for each room). This layer was generated using smart cameras and audio sensors based on the number of detected risky events using deep learning algorithms as described in the following sections.
As we progress with COVID, various criteria have been introduced and evaluated in COVID-19 spread risk [54,55]. Transmission ways of coronavirus and prioritizing their importance are still under debate and more studies are underway to understand transmission ways better [56]. Although the risk calculation could be very complicated based on time, room volume, air circulation, etc. [57,58], our research’s scope includes a general risk assessment function which is a simple weighted average. The COVID-19 viral load risk function was modelled using a weighted average of risky criteria (as mentioned above) over each IndoorGML node (e.g., a meeting room). For the initial prototype calculation of risk, Equation (1) was used:(1) { RiskCOVID19 = W1C1+W2C2+W3C3+W4C4W1=W2=W3=W4=1, to simplify the model
For simplicity sake, we assigned each of the above map layers with a similar weight and computed the risk factor over each IndoorGML node using an aggregation function.
However, this risk function can be easily manipulated and configured by the users on the client side. So, we evaluated a set of different weights and evaluated them in Section 5.7. In this new risk model, the wights are as follows:
Risk from Cleaning : W1=0.1: since cleaning is a measure of potentially unknown risks of COVID-19 transmission, such as airborne particles passing through the ventilation.
Risk from Contact Tracing: W2=0.4: it is one of the strongest risks for the place
Risk from People Density: W3=0.3: people using the space or engaging in risky behaviors in the space are stronger indicators and higher weight.
Risky behaviours: W4=0.2: We assumed the number of people being the strongest due to the prevalence of airborne virus being multiplicative on the number of people in the room, whereas risky behaviors are less frequent, and therefore harder to weight (W3=0.2).
Moreover, the above Equation (1) does not take into account the duration of time spent occupying the space, the actions taken, and the decay of the virus particles over time. The algorithm restarts the people count and cough numbers after the space is cleaned. A slightly more complicated set of equations (Equation (2)) expands on the simple risk calculation by taking into account the amount of time that people remain in a particular space, and the decay of the virus particles over time (assuming the worst-case scenario of 72 h for all of the particles deposited to become inactive). (2) {C3= number_of_people × person_dwell_time(hours)×72− time_since_departure(hour)72 C4=AverageLastRiskyBehavior×72−TimeLastRiskyBehavior(hour)72
For future research, we will consider different risk profiles (i.e., optimistic and pessimistic) for various user groups (e.g., pessimistic risk profile for COVID-19 vulnerable people).

4.2. Interoperable IoCT Using STA
In order to integrate multiple COVID-19 sensor systems, the OGC STA was used to support the interoperability between the sensing layer, cloud data management, and cleaning risk assessment application. Figure 7 shows an example of the OGC STA data model being used in the cleaning scenario for a specific Thing—an IndoorGML cell.
Every IndoorGML cell has a Location in space and time. This geospatial encoding was performed by GeoJSON (Geographical JavaScript Object Notation) [59]. Every sensor was referenced by the IndoorGML cell in which the sensor was installed. Each Thing can have multiple Datastreams, which are collections of Observation entities grouped together using the same Observed Property. For the cleaning use case, a different Datastream for each sensor’s phenomenon was used. Each Datastream contained a Sensor and an ObservedProperty. This refers to the instruments that can observe a phenomenon. For this paper, eight different Datastreams were defined, including, proximity, density, and coughs. An ObservedProperty specifies the phenomenon and also contains the unit of measurement. A Datastream can have several Observations, and they dictate the value for the phenomena encoded by the OGC Observations and Measurements (OM). For our example, this can refer to the values taken from a sensor measurement. FeatureOfInterest identifies the characteristics of the Thing. The Thing entity is an IndoorGML cell and the FeatureOfInterest entity describes the characteristics of this cell. For example, Figure 7 shows that “Duration” spent by a smartphone user in a room recorded within the proximity of a BLE beacon is considered a Datastream entity that kept the “Time” duration in seconds as an ObservedProperty. This Datastream entity used “Smartphones” as the sensor entity to keep Observations which are the duration of time that users spend in each cell in seconds.
Table 2 lists all Datastream entities that were used together with their sensing profile, whereby each property indicates the type of format that was encoded. For this research, all of the observations were sent to the Amazon IoT Core using smartphones and the Jetson Xavier NX development kit [49]. The next step was to map observations to an instance of the OGC STA endpoint using the Amazon Lambda functions. Interested readers can see and test the JSON payloads that were used to send all eight types of observations in Supplementary Materials.