A Data-Driven Framework for Braking Distance Prediction in Automatic Emergency Braking Systems Using CarSim Simulation and Machine Learning

Accurate braking distance prediction is essential for optimizing active safety deployment in Automatic Emergency Braking (AEB) systems. In rear-end emergency scenarios, stopping distance is governed by strongly coupled and nonlinear interactions among vehicle kinematics, tire–road adhesion, aerodynamic resistance, and dynamic load transfer, making conventional linear or simplified analytical models inadequate. This study proposes an integrated data-driven framework that combines high-fidelity multi-body vehicle dynamics simulation using CarSim, Principal Component Analysis (PCA), and machine learning regression algorithms for braking distance prediction. A dataset of 700 emergency braking cases was generated under three standardized Car-to-Car Rear (CCR) scenarios, including stationary, moving, and braking targets, with the road friction coefficient continuously varied from 0.5 to 1.0. Five physical input variables, namely ego-vehicle velocity, target velocity, initial distance, road friction coefficient, and effective braking force, were mapped to the corresponding break distance. PCA was applied to decorrelate the feature space and reduce redundancy, preserving more than 99.7% of the total variance within a four-dimensional latent subspace. Four regression models, including Linear Regression, Random Forest, Support Vector Regression, and an Artificial Neural Network, were evaluated using 10-fold cross-validation and statistical significance testing. The results show that nonlinear models substantially outperform the linear baseline, confirming the nonlinear nature of emergency braking dynamics. In the PCA-reduced feature space, the Artificial Neural Network achieved strong predictive performance, with a mean MAE of 0.929 m and an R² score of 0.993, closely followed by XGBoost (MAE = 0.921 m, R² = 0.991). However, pairwise Wilcoxon signed-rank tests reveal that the performance differences among all evaluated non-linear models (RF, SVR, XGBoost, LightGBM, and ANN) are not statistically significant (p>0.05). Consequently, non-linear machine learning models as a class exhibit equivalent high-precision predictive capabilities ("MAE"<1.0" m" , R^2≥0.989), offering flexible algorithm selection for embedded safety deployments. An ablation study reveals an engineering trade-off: the original feature space provides higher prediction accuracy, whereas PCA reduces computational runtime by 3.16% to 25%. These findings demonstrate that the proposed framework provides accurate braking-distance prediction while reducing computational cost. However, its real-time implementation on embedded AEB hardware remains to be experimentally validated