To safely protect workplaces and the workforce during and after the COVID-19 pandemic, a scalable integrated sensing solution is required in order to offer real-time situational awareness and early warnings for decision-makers. However, an information-based solution for industry reopening is ineffective when the necessary operational information is locked up in disparate real-time data silos. There is a lot of ongoing effort to combat the COVID-19 pandemic using different combinations of low-cost, location-based contact tracing, and sensing technologies. These ad hoc Internet of Things (IoT) solutions for COVID-19 were developed using different data models and protocols without an interoperable way to interconnect these heterogeneous systems and exchange data on people and place interactions. This research aims to design and develop an interoperable Internet of COVID-19 Things (IoCT) architecture that is able to exchange, aggregate, and reuse disparate IoT sensor data sources in order for informed decisions to be made after understanding the real-time risks in workplaces based on person-to-place interactions. The IoCT architecture is based on the Sensor Web paradigm that connects various Things, Sensors, and Datastreams with an indoor geospatial data model. This paper presents a study of what, to the best of our knowledge, is the first real-world integrated implementation of the Open Geospatial Consortium (OGC) Sensor Web Enablement (SWE) and IndoorGML standards to calculate the risk of COVID-19 online using a workplace reopening case study. The proposed IoCT offers a new open standard-based information model, architecture, methodologies, and software tools that enable the interoperability of disparate COVID-19 monitoring systems with finer spatial-temporal granularity. A workplace cleaning use case was developed in order to demonstrate the capabilities of this proposed IoCT architecture. The implemented IoCT architecture included proximity-based contact tracing, people density sensors, a COVID-19 risky behavior monitoring system, and the contextual building geospatial data.

Emerging deep learning (DL) approaches with edge computing have enabled the automation of rich information extraction, such as complex events from camera feeds. Due to the low speed and accuracy of object detection, some objects are missed and not detected. As objects constitute simple events, missing objects result in missing simple events, thus the number of detected complex events. As the main objective of this paper, an integrated cloud and edge computing architecture was designed and developed to reduce missing simple events. To achieve this goal, we deployed multiple smart cameras (i.e., cameras which connect to the Internet and are integrated with computerised systems such as the DL unit) in order to detect complex events from multiple views. Having more simple events from multiple cameras can reduce missing simple events and increase the number of detected complex events. To evaluate the accuracy of complex event detection, the F-score of risk behaviour regarding COVID-19 spread events in video streams was used. The experimental results demonstrate that this architecture delivered 1.73 times higher accuracy in event detection than that delivered by an edge-based architecture that uses one camera. The average event detection latency for the integrated cloud and edge architecture was 1.85 times higher than that of only one camera. However, this finding was insignificant with regard to the current case study. Moreover, the accuracy of the architecture for complex event matching with more spatial and temporal relationships showed significant improvement in comparison to the edge computing scenario. Finally, complex event detection accuracy considerably depended on object detection accuracy. Regression-based models, such as you only look once (YOLO), were able to provide better accuracy than region-based models.

Complex Event Detection (CED) in video streams involves numerous challenges such as object detection, tracking, spatio–temporal relationship identification, and event matching, which are often complicated by environmental variations, occlusions, and tracking losses. This systematic review presents an analysis of CED methods for video streams described in publications from 2012 to 2024, focusing on their effectiveness in addressing key challenges and identifying trends, research gaps, and future directions. A total of 92 studies were categorized into four main groups: training-based methods, object detection and spatio–temporal matching, multi-source solutions, and others. Each method’s strengths, limitations, and applicability are discussed, providing an in-depth evaluation of their capabilities to support real-time video analysis and live camera feed applications. This review highlights the increasing demand for advanced CED techniques in sectors like security, safety, and surveillance and outlines the key opportunities for future research in this evolving field.

Complex event detection (CED) adds value to camera stream data in various applications such as workplace safety, task monitoring, security, and health. Recent CED frameworks have addressed the issues of limited spatiotemporal labels and costly training by decomposing the CED into low-level features, as well as spatial and temporal relationship extraction. However, these frameworks suffer from high resource costs, low scalability, and an increased number of false positives and false negatives. This paper proposes GICEDCAM, which distributes CED across edge, stateless, and stateful layers to improve scalability and reduce computation cost. Additionally, we introduce a Spatial Event Corrector component that leverages geospatial data analysis to minimize false negatives and false positives in spatial event detection. We evaluate GICEDCAM on 16 camera streams covering four complex events. Relative to a strong open-source baseline configured for our setting, GICEDCAM reduces end-to-end latency by 36% and total computational cost by 45%, with the advantage widening as objects per frame increase. Among corrector variants, Bayesian Network (BN) yields the lowest latency, Long Short-Term Memory (LSTM) achieves the highest accuracy, and trajectory analysis offers the best accuracy–latency trade-off for this architecture.