Abstract
Recently, there are many researches on active safety system of intelligent vehicle. To reduce the probability of collision caused by driver's inattention and mistakes, the active safety system gives warning or controls the vehicle toward avoiding collision. For the purpose, it is necessary to recognize and analyze circumstances around. In this paper, we will treat the problem about collision risk assessment. In general, it is difficult to calculate the collision risk before it happens. To consider the uncertainty of the situation, Monte Carlo simulation can be employed. However it takes long computation time and is not suitable for practice. In this paper, we apply neural networks to solve this problem. It efficiently computes the unseen data by training the results of Monte Carlo simulation. Furthermore, we propose the features affects the performance of the assessment. The proposed algorithm is verified by applications in various crash scenarios.
Introduction
Active safety technologies, such as Autonomous Emergency Braking (AEB), are essential for mitigating traffic accidents resulting from driver error. A core requirement of these systems is the ability to assess collision risk accurately and instantaneously. While Monte Carlo (MC) simulations provide a robust framework for modeling the probabilistic nature of driver behavior and vehicle trajectories, their high computational cost makes them impractical for real-time deployment. This study proposes a Neural Network (NN) based approach to approximate MC results, enabling high-fidelity risk assessment at millisecond speeds.
Methodology
The proposed framework bridges the gap between complex simulation and real-time execution through a two-step process. First, offline Monte Carlo simulations are conducted to generate a comprehensive dataset of collision risks across various scenarios. These simulations account for uncertainties in driver acceleration and steering by modeling them as Gaussian distributions.
Second, a Multi-layer Perceptron (MLP) with 200 hidden nodes is trained to predict the collision risk score. A key contribution of this work is the identification of specific "collision features" that enhance the network's predictive accuracy:
- Time-to-Collision (TTC): The projected time until impact assuming constant velocities.
- Overlap: The physical width of the expected contact area between vehicles.
- Angle: The relative heading angle between the vehicles at the moment of impact.
By incorporating these physics-informed features, the model can better capture the non-linear dynamics of vehicle interactions compared to using raw state data alone.
Experimental Setup
The system was validated using synthetic data generated in Tass PreScan, a professional-grade simulation environment for Advanced Driver Assistance Systems (ADAS). The evaluation utilized 48 distinct crash scenarios provided by Hyundai Mobis, covering rear-end collisions, intersections, lane changes, and head-on impacts.
Results and Discussion
The experimental results highlight a significant performance leap in computational efficiency. The Neural Network achieved a risk assessment time of 0.0067 seconds, a 99.4% reduction compared to the 1.0928 seconds required by the original Monte Carlo simulation. This 163x speedup is crucial for vehicles traveling at high speeds where every millisecond counts.
In terms of accuracy, the inclusion of the three proposed collision features reduced the Mean Absolute Error (MAE) from 0.2458 to 0.2182, an 11% improvement. The system successfully issued warnings at least 1 second before impact in 79.2% of all test cases. It was most effective in intersection (100% success) and rear-end (90.5%) scenarios.
However, the system faced challenges in head-on collision scenarios, where the success rate dropped to 37.5%. This is primarily attributed to the high relative closing speeds and the 30-meter range limitation of the simulated sensors, which often left insufficient time for a 1-second pre-impact warning.
Conclusion
This paper demonstrates that Neural Networks can effectively serve as high-speed surrogate models for computationally expensive safety simulations. By integrating domain-specific geometric features, the proposed system achieves the necessary balance of accuracy and real-time performance for active safety applications. Future research will focus on extending sensor range and refining the network architecture to better address high-speed frontal collision scenarios.
Summary updated