Oleksandr Byzkrovnyi

This study aims to determine the prerequisites for the occurrence of road traffic accidents, analyze the most dangerous maneuvers of motor vehicles that can lead to hazardous situations, and develop the most effective method for promptly informing the driver about potential danger. The goal of this study is to develop an information system that ensures timely notification of drivers about possible road traffic accidents in designated hazardous areas. The tasks include: investigating existing computer vision models for classification and object tracking tasks and determining the most suitable ones for deployment on a single-board computer Nvidia Jetson, while examining their performance and technical limitations; developing an optimized solution for the prompt notification of drivers about danger; creating an algorithm for detecting potential vehicle collisions that integrates computer vision methods and mathematical modeling; developing a comprehensive danger warning system based on the obtained results and testing its functionality. The following methods were applied in this study: a process-based approach to investigate the mechanisms of road traffic accident occurrence, statistical analysis of hazardous areas and maneuvers, and performance analysis of computer vision models for real-time object detection and tracking and driver notification. Additionally, road situations were simulated and modeled using the BeamNG.tech environment. The results include the development of a methodology based on computer vision and mathematical models for identifying hazardous situations on the road and the creation of an approach for prompt notification of road users using cloud technologies, IoT devices, and the GeoHash algorithm. An information system that allows drivers to receive warnings about potential hazards along their route has been proposed. Conclusions: this study confirms the successful development of a software system for forecasting and notifying drivers about the risk of road traffic accidents. The conducted studies have demonstrated the effectiveness of the proposed algorithm for detecting hazardous situations and technological solutions for road infrastructure integration. Experiments conducted using BeamNG.tech have confirmed the functionality of the developed system, which can be applied to minimize the risk of road traffic accidents in designated hazardous areas.

This research focuses on developing an information technology solution capable of delivering real-time notifications about potential car accidents to Internet of Things (IoT) devices installed in vehicles. Since vehicles are mobile and frequently change location, there is a need to have stable and timely notifications. To address this, the proposed solution leverages cloud technologies, which offer scalability, low latency, and high availability for real-time data transmission. The main objective is to design and implement architecture that enables effective real-time messaging between cloud services and moving IoT devices. The system is tested across various simulated driving scenarios to evaluate its performance and reliability. A key aspect of the methodology involves measuring the time interval between when a message is sent and when it is received by the IoT device, followed by assessing whether the driver has adequate time to respond appropriately. The experimental setup is based on test cases created using Python, the BeamNG.tech simulation environment, and AWS IoT Core as the cloud service provider. The results demonstrate that the proposed technology can reliably handle real-time messaging in dynamic conditions. As a conclusion, the research confirms the potential of the developed solution for real-life applications in connected vehicle safety systems.

The subject matter: this study focuses on the preconditions for the occurrence of road traffic accidents at intersections and areas with limited visibility; the use of computer vision models for classifying the preconditions of road traffic accidents and evaluating the effectiveness of their use in real-time operation. The goal of the study is to compare computer vision models for the task of classifying the preconditions of traffic accidents for real-time operation. The study involves comparing models using the Jetson TX2 platform and determining the effectiveness of this approach for generating real-time warning signals for drivers. Tasks: explore computer vision models such as Detectron2 and YOLOv7 for the task of classifying traffic accident preconditions in terms of model performance, ease of dataset creation, model training, and deployment. Compare YOLOv8 and DetectNet_v2 on a single-board computer Jetson TX2 in terms of processing speed. The methods used include training and using machine learning models, as well as simulating hazardous situations using software such as BeamNG.tech and CARLA. A comparative analysis of the application results of the models was conducted using performance evaluation metrics. The main results of the study include identifying the most effective model for classification tasks on the Jetson TX2 single-board computer – DetectNet_v2; a positive evaluation of the effectiveness of this approach for real-time driver warning, although certain limitations were noted regarding the size of the training dataset and the complexity of its preparation. Conclusions. The following computer vision algorithms were examined: Detectron2, YOLOv7, and DetectNet_v2. It was found that YOLOv7 outperforms Detectron2 in detecting the preconditions of traffic accidents in images. However, DetectNet_v2 was found to be better suited for deployment on the Jetson TX2 single-board computer compared to YOLOv7. Additionally, based on experimental findings, it was concluded that the application of this approach to predicting accident preconditions is problematic due to difficulties in creating a training dataset – specifically, the variability of precondition scenarios.

The research subject is road crash accident nature and approaches for its preventions or predictions in the real-time using computer vision algorithms and usage of edge devices. The goal of this research is to create a model for prediction of potential vehicles collision, which works for real-time. The methodology used in the research is a combination of computer vision model TrafficCamNet_1.3 output with the math approaches to determine the possible vehicles collision. The exact math methods include calculation of cars’ movement projections and usage it for checks whether vehicles collision may occur or not. The experiments setup is based on the scenarios designed using BeamNG.tech, usage of Nvidia Jetson Orin Nano as a platform for running real-time classification and determination of possible road crash accidence. The main results of this research are outlining the exact time spent for having car stopped before crash, exact cars’ characteristics and case setup and the percentage of happened road crash accidents to determine the model robustness and ability to real life introducing. As a conclusion, this research reveals the facts, that model works for the cases, when cars did not exceed allowed speed limit on the particular road. With the allowed speed, driver will be able to be notified in time and will have enough time to stop the car, otherwise amount of time to react on the threat is being significantly reduced. As a model improvement, the usage of models’ ensemble with different training dataset sizes can be considered for early car classification on the image. The results of this research can be used for building the intelligent software system for the preventions of road traffic accidence on the defined as a dangerous road parts.

Nowadays the robotics is relevant development industry. Robots are becoming more sophisticated, and this requires more sophisticated technologies. One of them is robot vision. This is needed for robots which communicate with the environment using vision instead of a batch of sensors. These data are utilized to analyze the situation at hand and develop a real-time action plan for the given scenario. This article explores the most suitable algorithm for detecting potential road accidents, specifically focusing on the scenario of turning left across one or more oncoming lanes. The selection of the optimal algorithm is based on a comparative analysis of evaluation and testing results, including metrics such as maximum frames per second for video processing during detection using robot’s hardware. The study categorises potential accidents into two classes: danger and not-danger. The Yolov7 and Detectron2 algorithms are compared, and the article aims to create simple models with the potential for future refinement. Also, this article provides conclusions and recommendations regarding the practical implementation of the proposed models and algorithm.