Artificial intelligence for real-time injury prediction

Fraunhofer EMI with focus on VRU-car collision simulations

The goal of the ATTENTION project is to develop a method for real-time injury prediction of vulnerable road users (VRUs), such as pedestrians or cyclists. For this purpose, data-driven methods are used to determine a situation-specific injury risk from vehicle-based video data and virtual tests with the help of digital human models. Prospectively, injury prediction enables both safe and efficient traffic through automated vehicle risk mitigation strategies.

© Fraunhofer EMI
Example of a collision simulation between a muscle-driven cyclist and a vehicle model.

Multimodal traffic and human individuality as major challenges for autonomous driving

In many German inner cities, urban mobility is characterized by the multiple use of limited and narrow traffic areas and a large number of different road users, which in their entirety constitute multimodal traffic. A key future issue for cities and mobility service providers is the efficient and, at the same time, safe use of shared public spaces. Participating road users differ substantially in terms of speed, maneuverability and vulnerability. The large-scale establishment of automated traffic promises to reduce traffic accidents and fatal road injuries through their interaction. Nevertheless, complex inner-city scenarios, undirected traffic flows and human individuality pose major challenges to increasing automation.


Classification in existing safety concepts

Collisions between vehicles and VRUs still lead to a large number of accident fatalities. The number of traffic accidents involving cyclists has not decreased in the EU since 2010 and has even increased in Germany. Even in future multimodal traffic, collisions between vehicles and VRUs cannot be ruled out due to the previously mentioned aspects, such as human individuality.

With regard to the road safety of VRUs, the primary goal must be to avoid an accident as best as possible. Sensor technology and communication systems implemented in the vehicle are already being used to identify potential collisions at an early stage and, for example, to adjust the safety distance via assistance systems and avoid collisions with active braking and steering systems. But what if the accident is unavoidable?

If the “Vision Zero” (no more traffic fatalities by 2050) is to be addressed realistically, it must be accepted that not all accidents are preventable and that current and future traffic is therefore accompanied by an uncertainty factor, for example, due to human individuality. Consequently, in addition to accident avoidance strategies, accident mitigation measures and, thus, injury reduction must also be considered, which raises the question: In the event of an unavoidable collision: Which technical measures can be taken to reduce the situation-specific injury risk of VRUs? The project ATTENTION addresses this gap in a proof-of-concept study.


© Adobe Stock
Visualization of VRU and object detection in highly automated vehicles.

Concept and aim of the project

With the aim of a situation-specific prediction of injury risks, various databases are being set up in the project, and artificial intelligence (AI) methods are being used and linked together.

At the beginning of the project, vehicle-based video data of actual vehicle-VRU collisions are analyzed. In this framework, biomechanical and AI-based motion prediction methods were applied to build a VRU position and motion database. The subsequent stage involves extracting the most frequently occurring pre-impact positions of pedestrians and cyclists and positioning the corresponding human body model (HBM) utilized in the project in the finite-element (FE) environment accordingly. These positions include, for example, a mean cyclist posture and various pedestrian defensive postures, which were observed accordingly in the analyzed video data. In addition, a muscle model with position-specific muscle activities is integrated into the human model to best approximate the real collision behavior of VRUs. 

© Fraunhofer EMI
Technical classification of the ATTENTION method in existing autonomous driving (AD) system concepts.

The positioned HBM equipped with muscles is then subsequently integrated with an FE bicycle and vehicle model in a virtual collision scenario in order to numerically compute a vehicle-VRU collision in the FE code after defining boundary conditions (for example, initial velocity and impact angle). The HBM is instrumented beforehand to ensure extraction of injury-relevant information such as strains or accelerations for specific body regions from the simulation. This makes it possible to calculate situation-specific injury risks via the FE simulation for a wide variety of collision scenarios. With the goal of data-driven injury prediction, an injury database is built over a large number of vehicle-VRU collision simulations. Single, representative collision simulations are compared with real data from Bosch accident research and checked for plausibility, thereby increasing confidence in the simulation-based injury prediction. In addition, the real accident data are used to define the limits of the parameter space (for example, permissible combinations of relative speeds and collision angles) for the collision simulations.

The synthetic data sets generated by FE simulations are used in the next step to train different AI models. This shall make it possible to predict a plausible injury risk even for non-simulated parameter combinations, for which a time-intensive FE simulation would otherwise be necessary. The ambitious goal of the project is to predict this situation-specific injury risk in real time.

In the final step of the project, the AI-based prediction of injury risks will be demonstratively tested in a virtual driving environment. In detail, the motion prediction and AI-based injury prediction methods of this project will be integrated as an active safety tool into a virtual vehicle in the driving environment. The corresponding VRU model is detected by a virtual camera and its further movement until the time of collision is predicted, taking into account the distance between the vehicle and the VRU, the impact angle and the relative speeds. Based on this information, the braking and steering actions are then adjusted to best reduce the VRU injury risk in the remaining time period. Through prior training with the synthetic data from the collision simulation, the AI derives the parameter combination that can still be achieved to reduce the predicted VRU injury risk as best as possible.


Fraunhofer EMI with focus on VRU-car collision simulations

As part of the ATTENTION project, Fraunhofer EMI will contribute to the development of the collision and injury database and is represented by the groups Human Body Dynamics and Digital Engineering. Different competences of the two groups will be combined. On the one hand, the individual FE models must be prepared appropriately for data generation and extraction of biomechanical injury information. Further, single simulations are compared with real accident data for single collision scenarios to increase trust in the generated simulation data. To create a large variety of vehicle-to-VRU collision simulations, which are also observed in real accident data, the vehicle and VRU FE-models must be arbitrarily combinable to allow for different parameter combinations of e.g. the impact angle and the VRU-position in front of the bumper. On the other hand, in order to cover the large parameter space of possible collision scenarios as efficiently as possible, the development and application of methods for adaptive data generation is beneficial. Thus, despite the long computation times of FE collision calculations, a sufficient information basis can be generated to train corresponding AI models with the goal of injury prediction for this specific application case of frontal impact. Based on the expertise in biomechanics and human body models as well as on profound knowledge in data-driven methods and efficient generation of simulation-based training data, EMI builds up a collision and injury database within the ATTENTION project.

The ATTENTION project will end in June 2024 and is funded by the German Federal Ministry for Economic Affairs and Climate Action within the framework of the research program “New Vehicle and System Technologies”.