Deep Vision Net: An AI-Based System for Dynamic Traffic Scene Reconstruction and Safety Prediction with Explainable AI
Year: 2023
Type: article
Abstract: The research paper explores innovative, dynamic traffic scene reconstruction methodologies and multi-modal data fusion in safety-critical applications. Leveraging deep learning techniques, we propose an AI-based system capable of processing traffic images and LiDAR data to predict safety measures for connected vehicles. Our system utilizes the popular ResNet50 model and LSTM layers to create a DeepVisionNet model, enabling efficient multi-modal data fusion. To ensure comprehensive model training and address the limitations, we employ synthetic data generation techniques, which facilitate the analysis of various traffic scenarios. Through extensive experiments carried out by executing validation using real-world traffic data and connected vehicle simulations, we evaluate the performance and effectiveness of our AI-based system. Our results demonstrate superior accuracy, reliability, and interpretability compared to existing approaches in the literature. By providing interpretable safety recommendations by adopting Explainable-AI (XAI) approach to drivers and traffic management authorities, our system contributes significantly to road safety and traffic optimization. The AI-based system proves to be an invaluable asset for dynamic traffic scene reconstruction and multi-modal data fusion. It offers the potential to revolutionize the field of traffic analysis and safety prediction, providing a safer and more efficient transportation ecosystem.
Language: English
Subfield: Automotive Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7218181623
