Multi-source temporal attention fusion network (MTAFN) for driving risk assessment based on naturalistic driving data

Real-time and short-term driving risk assessment is crucial for advancing Advanced Driver Assistance Systems (ADAS) by enabling proactive warning strategies and enhancing driving safety. However, the complexity and dynamic nature of real-world driving environments, coupled with diverse data sources, challenge traditional risk assessment methods in capturing the coupled effects of various factors. To address these limitations, this study proposes a novel framework for driving risk assessment based on multi-source data, integrating both unsupervised learning and deep learning techniques. The proposed Multi-source Temporal Attention Fusion Network (MTAFN) integrates three core components: a Feature Selection Network to dynamically identify critical input features, a Static Feature Encoder to fuse static and dynamic data through information flow propagation, and a Temporal Attention Fusion Network employing a modified multi-head attention mechanism to capture long-term temporal dependencies. The experimental results demonstrate that the proposed model outperforms other models on the performance, showing excellent transferability across different scenarios, along with further improvements in both performance and efficiency. Furthermore, the model exhibits comprehensive interpretability in multi-source feature fusion, temporal dependencies, and scenario transferring. Its robustness across different risk labelling strategies has also been validated. This study highlights MTAFN’s effectiveness in leveraging multi-source data for driving risk assessment and its potential to advance proactive warning strategies, offering a robust solution for enhancing safety of ADAS in complex environments