| Aggressive driving behavior is a major contributor to traffic accidents, increased energy consumption, and accelerated vehicle lifespan. As a result, automatic aggressive driving detection has become an important component of intelligent transportation systems. While numerous studies report high classification accuracy on public driving behavior datasets, many rely on accuracy evaluation and fixed preprocessing pipelines, limiting their applicability to real-world scenarios. This paper presents a systematic experimental analysis of aggressive driving detection based on inertial sensor data, with particular emphasis on the role of temporal segmentation and class imbalance mitigation. The study investigates how preprocessing decisions influence detection performance across different deep learning models. Experiments are conducted using Long Short-Term Memory (LSTM) networks and a hybrid architecture combining Convolutional Neural Networks (CNNs) with LSTM layers, evaluated on two complementary datasets: a public driving behavior dataset and a real-world dataset collected from an electric vehicle. The results show that temporal window size and class imbalance mitigation strategies, including class weighting and minority class oversampling, significantly affect detection performance. Longer temporal windows generally improve aggressive driving detection by capturing complete driving maneuvers, while appropriate imbalance mitigation substantially improves the recognition of rare aggressive events. The experiments further show that temporal segmentation can significantly alter the effective class distribution of the generated samples. Additionally, the findings confirm that accuracy alone may overestimate model performance under strong class imbalance. These results highlight the importance of carefully designing preprocessing and evaluation strategies for real-world aggressive driving detection. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.