Internet of Things (IoT) systems are becoming ubiquitous in various cyber–physical infrastructures, including buildings, vehicular traffic, goods transport and delivery, manufacturing, health care, urban farming, etc. Often multiple such IoT subsystems are deployed in the same physical area and designed, deployed, maintained, and perhaps even operated by different vendors or organizations (or “parties”). The collective operational behavior of multiple IoT subsystems can be characterized via (1) a set of operational rules and required safety properties and (2) a collection of IoT-based services or applications that interact with one another and share concurrent access to the devices. In both cases, this collective behavior often leads to situations where their operation may conflict, and the conflict resolution becomes complex due to lack of visibility into or understanding of the cross-subsystem interactions and inability to do cross-subsystem actuations. This article addresses the fundamental problem of detecting and resolving safety property violations. We detail the inherent complexities of the problem, survey the work already performed, and layout the future challenges. We also highlight the significance of detecting/resolving conflicts proactively, i.e., dynamically but with a look-ahead into the future based on the context.

The emerging edge computing applications often use high definition cameras as edge devices to capture video streams that need to be analyzed in real-time for situational understanding and answering queries. However, such devices suffer from limited energy (and hence limited computing power) and limited bandwidth available to stream the data to the edge controllers that provide much higher computing capacities. In this paper, we address these issues in the context of vehicular traffic monitoring and develop a scheme that has two components: YLLO and BATS. YLLO is a lightweight object recognition algorithm that runs on the edge device itself and substantially reduces the frame rate sent to the edge controller without removing the important information. BATS adapts the transmissions to the available bandwidth by taking advantage of further redundancy in the video stream in both single and multi-camera scenarios. We show that these mechanisms together can maintain object identification accuracy of above 95% while transmitting just 5-10\% of all the frames recorded by the cameras.

This paper introduces TU-DAT, a novel, freely downloadable computer vision dataset for analyzing traffic accidents using roadside cameras. TU-DAT addresses the lack of public datasets for training and evaluating models focused on automatic detection and prediction of road anomalies. It comprises approximately 280 real-world and simulated videos, collected from traffic CCTV footage, news reports, and high-fidelity simulations generated using BeamNG.drive. This hybrid composition captures aggressive driving behaviors—such as tailgating, weaving, and speeding—under diverse environmental conditions. It includes spatiotemporal annotations and structured metadata such as vehicle trajectories, collision types, and road conditions. These features enable robust model training for anomaly detection, spatial reasoning, and vision–language model (VLM) enhancement. TU-DAT has already been utilized in experiments demonstrating improved performance of hybrid deep learning- and logic-based reasoning frameworks, validating its practical utility for real-time traffic monitoring, autonomous vehicle safety, and driver behavior analysis. The dataset serves as a valuable resource for researchers, engineers, and policymakers aiming to develop intelligent transportation systems that proactively reduce road accidents.

In this paper, we demonstrate that Deep Learning (DL) and spatiotemporal reasoning can effectively identify driving behavior based on the videos captured by roadside cameras. The use of roadside infrastructure for such determination is twofold: (1) a global view of the vehicles and their interactions, and (2) no involvement or awareness of the vehicles or their drivers, so the determination is inexpensive, easy to deploy, and entirely non-intrusive. Furthermore, our method uses deep learning only for object detection and tracking and builds a flexible and explainable reasoning model to identify the driving behavior. The essential advantage of this approach is that we use deep learning only for tasks that can be accomplished efficiently and with high accuracy (i.e., object detection and tracking), which can be done in real-time. Although there are deep learning models for detecting complex activities (e.g., aggressive driving), they are much harder to train, require higher accuracy, and inferencing time may not satisfy real-time constraints. By using a setup with program-controlled robocars, we demonstrate that we can achieve accuracies of 98–99% for driving behavior characterization, and the mechanism can provide detection of 650 ms on a very dated desktop. The characterization can provide feedback to the driver (or the automated car) for improved traffic safety and roadway throughput.

In this paper, we present C-FAR, a framework for reasoning about anomalies in road-based intelligent transportation systems (ITS) based on video monitoring by the roadside camera infrastructure. The anomalies could span broad temporal and spatial ranges, including fine-grain (e.g., unsafe interactions among moving vehicles in real-time), medium-grain (e.g., aggressive/unsafe driving styles of individual vehicles over extended periods/distances), and coarse-grain (e.g., ensemble properties of the traffic over even longer time horizons). Unlike traditional approaches that utilize deep learning to recognize individual activities, C-FAR does so only for primitive movements and activities and then builds a comprehensive event logic framework. It also provides an optimal resolution of the detected/predicted anomalies by identifying the minimal changes in the controllable parameters of the system. We implemented a prototype system and tested it on three distinct real-world traffic data sets. We demonstrate that the proposed scheme can predict anomalies with over 84% recall level at 95% confidence level approximately 4.05 seconds before the incident.

Vision-Language Models (VLMs) offer the ability to generate high-level, interpretable descriptions of complex activities from images and videos, making them valuable for situational awareness (SA) applications. In such settings, the focus is on identifying infrequent but significant events with high reliability and accuracy, while also extracting fine-grained details and assessing recognition quality. In this paper, we propose an approach that integrates VLMs with traditional computer vision methods through explicit logic reasoning to enhance SA in three key ways: (a) extracting fine-grained event details, (b) employing an intelligent fine-tuning (FT) strategy that achieves substantially higher accuracy than uninformed selection, and (c) generating justifications for VLM outputs during inference. We demonstrate that our intelligent FT mechanism improves the accuracy and provides a valuable means, during inferencing, to either confirm the validity of the VLM output or indicate why it may be questionable.