The ACFR Five Roundabout Trajectory Dataset is used in this study to compare personalized federated learning to traditional centralized machine learning for predicting a vehicle’s path at roundabouts. The training of a model in traditional centralized machine learning, which aggregates data from all sources into a single central server, may not effectively handle diverse data distributions. In contrast, personalized federated learning distributes the dataset across five clients, each representing a distinct roundabout, and trains models locally. Federated averaging then updates the global model, and meta-learning fine-tunes it. SHAP (Shapley Additive Explanations) performs feature selection, and user clustering methods to organize the data. Both approaches utilize Long Short-Term Memory (LSTM) models. The results show that personalized federated learning does a much better job than the centralized model at handling different types of data. It achieves test accuracy rates between 96.74% and 99.58% and also does a better job of recalling information. This highlights the federated model’s effectiveness in classifying vehicle path categories, as well as its robustness and adaptability in real-world scenarios.