|
Citation
|
Niranjan B, Thomas S, Michael D, Niclas T. 27th International Technical Conference on the Enhanced Safety of Vehicles (ESV); April 3-6, 2023; Abstract #: 23-0215, pp. 22p. Washington, DC USA: US National Highway Traffic Safety Administration, 2023 open access. |
|
Abstract
|
27th International Technical Conference on the Enhanced Safety of Vehicles (ESV): Enhanced and Equitable Vehicle Safety for All: Toward the Next 50 Years
https://www-esv.nhtsa.dot.gov/Proceedings/27/27ESV-000215.pdf
Urban traffic is characterized by limited space, varying traffic flows and multiple types of road users. Despite increasing automation and design efforts, the joint use of traffic areas poses a particular risk for vulnerable road users (VRUs). In order to make traffic as safe as possible, the severity of injuries to VRUs in unavoidable collisions must be reduced. In future applications, predicting situation-specific injury risks for VRUs in real-time using machine learning (ML) could support decision making in determining risk minimization strategies. The predictive capability of any ML model is determined by the quality of the used training data. While there are no real-world training data available for injury prediction, simulation data, which is frequently employed in passive safety engineering, can be used as synthetic data. Since deliberate training data generation consumes substantial resources, particular attention is focused on the iterative generation of optimized simulation data sets. This study presents and discusses an adaptive simulation data generation pipeline to generate simulation data sets that reflect the overall system's behavior with the overall goal of efficiency and sustainability.
The novel pipeline involving nine steps is divided into two phases, "Data Generation" and "Data Exploitation". The "Data Generation" phase predominately focusses on the adaptive strategies to generate a generalist training data set. Along with the fundamental techniques for adaptively adding new points, metrics for assessing the information content of the present data set and for tracking the iterative sampling progress are also discussed in this study. Additionally, experiments to understand the effects of batch size is conducted and the potential use of information content metrics for process termination and dynamic, adaptive batch size adjustment is discussed.
The pipeline is initially tested using a generic example and is then applied to a simulation setup modeling a human head crashing onto a vehicle windshield. The observations from applying the pipeline to the simulation setup are compared with the observations from applying it to the generic function to evaluate the novel pipeline. It is shown, that the pipeline is generally applicable to such real-world problems and that the anticipated dynamic behavior of the data generation process is confirmed in the generic and real application example. This lays fundamental groundwork which needs to be extended along multiple routes in future work.
Not only in recent years, urban traffic systems have shown a significant trend towards multimodality [1]. Next to motivating an extensive amount of research activities on the design of multimodal urban mobility systems, this poses significant challenges towards traffic safety to all stakeholders involved [2] [3]. Particularly vulnerable road users (VRUs), e.g. pedestrians or cyclists, exhibit an overproportionate share, an increased injury severity and relatively high death-rates in the accident data [4] [5] [6]. In order to respond to these VRU-specific needs in traffic safety, a consortium of industrial and academic partners has teamed up in the research project ATTENTION ("artificial intelligence for real-time injury prediction"*) to develop a framework, as well as constituting methods and tools to dynamically predict injury risks for VRUs in accident scenarios [7]. Given the requirement of (near) real-time prediction, conventional engineering methods to predict the behavior or performance of structures under dynamic loading conditions – namely the finite element method (FEM) – are not feasible. More precisely, while the comprehensive simulation of a crash scenario can take up to 30 h on an advanced compute cluster, the collision of a vehicle is avoidable until ca. 1.5 seconds before impact [8]. Hence, the potentials and applicability of artificial intelligence (AI) or – more accurately – machine learning (ML) to predict the system's responses in such scenarios are studied in this project ... |