Myrto Konstantinidou

This paper presents FireVision, an innovative platform and model for real-time fire detection and monitoring. The platform utilizes automated drone flights to collect high-resolution imagery in both suburban and forested settings. Ensemble deep learning inference, based on Mask R-CNN weak learners, is employed to trigger alerts. Detection performance is further enhanced by integrating ResNet-50, ResNet-101, and ResNet-152 classifiers, which can be deployed in the cloud or on the drone’s edge co-processing units. Additionally, a fire criticality index is introduced, leveraging detection bounds and masks to assess the severity of fire events, alongside an automated drone path-planning algorithm for identifying critical fire incidents. Experiments were conducted using a supervised, mask-annotated dataset to evaluate model accuracy and inference speed across various cloud and edge computing configurations. Results indicate that ResNet-101 surpasses ResNet-50 by 5 to 12.5 percent in mAP@0.5 mask accuracy, with an 18 percent increase in inference time on the cloud and a 27 percent increase on the drone edge device GPU. In comparison, ResNet-152 achieves a 0.5 to 1.2 percent improvement in mAP@0.5 over ResNet-101, but its inference time is nine times slower in the cloud and 1.3 times slower on the GPU.

This paper introduces FireVision, a novel detection platform and model designed for real-time fire detection and monitoring. The platform employs automatic drone flights to capture high-resolution images across both suburban and forest environments. It uses ensemble deep learning inference from Mask R-CNN weak learners to trigger alerts. These inferences are enhanced by the accurate detection capabilities of ResNet-50, ResNet-101, and ResNet-152 classifiers, which can be deployed either in the cloud or on the drone's edge co-processing units. The authors also implemented an index, called the Fire Criticality Index, that uses detection bounds and masks to indicate the criticality of a fire event, as well as an automated drone path-planning algorithm for detecting a fire critical event. The authors conducted experiments with their proposed model using a mask-annotated dataset comprising of 12,000 images. They evaluated the model's accuracy and inference speed across various cloud and edge computing setups. The experimental results revealed that ResNet-101 outperformed ResNet-50 by 5--12.5\% in maP@0.5 mask accuracy and demonstrated an 18\% increase in inference time when executed on the cloud and a 27\% increase on the drone edge device GPU. For the ResNet-152 compared to ResNet-101, the map@0.5 increased by 0.5--1.2\%. However, the ResNet-152 inference time was 9x slower in the cloud and 1.3x slower on the GPU.

Wildfires are increasingly recognized as a climatological hazard, able to threaten industrial and critical infrastructure safety and operations and lead to Natech disasters. Future projections of exacerbated fire regimes increase the likelihood of Natech disasters, therefore increasing expected direct damage costs, clean-up costs, and long-term economic losses due to business interruption and environmental remediation. While large industrial complexes, such as oil, gas, and chemical facilities have sufficient resources for the implementation of effective prevention and mitigation plans, small-to-medium-sized industrial hubs are particularly vulnerable due to their scattered distribution and limited resources for investing in comprehensive fire prevention systems. This study targets the vulnerability of these communities by proposing the deployment of Wireless Sensor Networks (WSNs) as cost-effective Early Wildfire Detection Systems (EWDSs) to safeguard wildland and industrial domains. The proposed approach leverages wildland–industrial interface (WII) geospatial data, simulated wildfire dynamics data, and mathematical optimization to maximize detection efficiency at minimal cost. The WII delimits the boundary where the presence of wildland fires impacts industrial activity, thus representing a proxy for potential Natech disasters. The methodology is tested in Cocentaina, Spain, a municipality characterized by a highly flammable Mediterranean landscape and medium-scale industrial parks. Results reveal the complex trade-offs between detection characteristics and the degree of protection in the combined wildland and WII areas, enabling stakeholders to make informed decisions. This methodology is easily replicable for any municipality and industrial installation, or for generic wildland–human interface (WHI) scenarios, provided there is access to wildfire dynamics data and geospatial boundaries delimiting the areas to protect.

In viticulture, downy mildew is one of the most common diseases that, if not adequately treated, can diminish production yield. However, the uncontrolled use of pesticides to alleviate its occurrence can pose significant risks for farmers, consumers, and the environment. This paper presents a new framework for the early detection and estimation of the mildew’s appearance in viticulture fields. The framework utilizes a protocol for the real-time acquisition of drones’ high-resolution RGB images and a cloud-docker-based video or image inference process using object detection CNN models. The authors implemented their framework proposition using open-source tools and experimented with their proposed implementation on the debina grape variety in Zitsa, Greece, during downy mildew outbursts. The authors present evaluation results of deep learning Faster R-CNN object detection models trained on their downy mildew annotated dataset, using the different object classifiers of VGG16, ViTDet, MobileNetV3, EfficientNet, SqueezeNet, and ResNet. The authors compare Faster R-CNN and YOLO object detectors in terms of accuracy and speed. From their experimentation, the embedded device model ViTDet showed the worst accuracy results compared to the fast inferences of YOLOv8, while MobileNetV3 significantly outperformed YOLOv8 in terms of both accuracy and speed. Regarding cloud inferences, large ResNet models performed well in terms of accuracy, while YOLOv5 faster inferences presented significant object classification losses.

Studies reveal that drivers’ behavior is the most significant factor in road accidents worldwide. Regarding tunnels, which are the most critical element of road infrastructure, despite the significant efforts that have been conducted toward the enhancement of drivers’ education all these years, studies illustrate that there are still serious deficiencies need to be tackled. To address this issue, this research endeavor develops a virtual reality tool based on the serious game idea in order to inform and educate potential users about the specific rules and behavioral patterns that should govern their safe driving when passing through tunnels. To do so, the appropriate behavioral patterns are determined using applicable norms and guidelines while the specific educational requirements are identified. Following that, the novel tool for training users is developed. The tool consists of a virtual reality gaming environment based on the notion of serious games that simulates driving through a tunnel from a first-person perspective. Various scenarios are developed within this environment based on the knowledge gaps identified in the literature, with the aim of assessing users’ knowledge as well as educating them when required. The developed tool was tried by more than 50 drivers, professional and non-professional during tool’s launch activities. In particular, drivers who had recently obtained a driver’s license confirmed that such a tool would be especially useful in the context of their training. The ultimate goal of this study is to provide an efficient tool in order to support both practitioners and authorities to significantly improve the safety level of road tunnels by emphasizing on the driving behavior, since this is considered the most crucial component of each tunnel system.

Automated deep learning and data mining algorithms can provide accurate detection, frequency patterns, and predictions of dangerous goods passing through motorways and tunnels. This paper presents a post-processing image detection application and a three-stage deep learning detection algorithm that identifies and records dangerous goods’ passage through motorways and tunnels. This tool receives low-resolution input from toll camera images and offers timely information on vehicles carrying dangerous goods. According to the authors’ experimentation, the mean accuracy achieved by stage 2 of the proposed algorithm in identifying the ADR plates is close to 96% and 92% of both stages 1 and 2 of the algorithm. In addition, for the successful optical character recognition of the ADR numbers, the algorithm’s stage 3 mean accuracy is between 90 and 97%, and overall successful detection and Optical Character Recognition accuracy are close to 94%. Regarding execution time, the proposed algorithm can achieve real-time detection capabilities by processing one image in less than 2.69 s.