On-Device Edge-Gated Attention and FFT for Smartphone-Based Aggressive Driving Detection

초록

Aggressive driving behaviors such as hard acceleration, braking, and sharp turning are major contributors to traffic accidents worldwide. While traditional detection systems rely on dedicated vehicle hardware, many smartphone-based approaches suffer from limited generalization and dependence on cloud-based computations. To address these challenges, this study proposes a practical and fully on-device deep learning framework for real-time aggressive driving detection using built-in smartphone inertial measurement unit (IMU) sensors. The proposed framework integrates a public benchmark dataset, real-world driving data collected at the Seoul National University (SNU) Autonomous Vehicle Test Track in Siheung, and simulated IMU data generated using the CARLA simulator. This hybrid dataset broadens the evaluated driving conditions by covering normal driving, aggressive maneuvers, and ISO 3888-2-based obstacle-avoidance scenarios, while also revealing remaining domain-shift limitations under strict source-holdout evaluation. The detection model is based on a Long Short-Term Memory (LSTM) backbone augmented with a lightweight Edge-Gated Temporal Attention (EGTA) module and a parallel Fast Fourier Transform (FFT) branch. EGTA is used as an engineering refinement to emphasize abrupt motion transitions, whereas the FFT branch provides compact frequency-domain cues related to vibration and window-level spectral structure. The experimental results show that the full LSTM + EGTA + FFT model improves Macro-F1 by about 7% over the standard LSTM baseline while achieving the best overall classification performance among the compared baselines and ablation variants. This paper presents a practical and privacy-aware on-device solution suitable for intelligent transportation systems, real-time driver feedback, and Usage-Based Insurance (UBI) applications. The privacy benefit arises from local smartphone inference and reduced external data transmission, not from formal privacy mechanisms such as differential privacy, encryption, secure aggregation, or federated learning.

키워드