Abstract
In intelligent connected environments, there is a growing demand for driver identification. Intelligent connected vehicles, relying on identity recognition, can provide more personalized driving experiences based on user preferences and habits. However, existing identity recognition algorithms often do not optimize input feature parameters, leading to poor performance in multi-user identification scenarios. This study first constructs features encompassing 15-dimensional driving state data categories. These features are used as inputs to compare the performance of different parameter types. Next, a multilayer perceptron model is employed for driver identification, with traditional machine learning models used for reference. Bayesian optimization is then applied to fine-tune hyperparameters utilizing the expected improvement and noise-adding strategy. The macro F1 score serves as the evaluation metric for the recognition models. To validate performance in multi-user identification, experiments are conducted using a natural driving dataset containing 44 drivers. Results indicate that under the selected 8-dimensional CAN signal conditions, MLP consistently achieves an above 96% macro F1 score with increasing driver numbers. Notably, under 22 driver conditions, MLP reaches a 97% recognition rate, surpassing most existing studies. When inputting different feature combinations from 44 drivers, the model performs best at 97.25% recognition with operational and posture features combined. After eliminating features with multicollinearity, recognition improves to 97.53%, a 2.23% enhancement. By adjusting sample sizes and optimized parameters, MLP achieves 92.97% recognition with 2.75% of the sample dataset, demonstrating effective multi-user identification with limited samples. Experimental results affirm the robustness of MLP in identifying a large number of drivers’ identities.
Keywords
Get full access to this article
View all access options for this article.