Experimental study demonstrates robust driver identification using multimodal sensor and video fusion, indicating improved vehicle security and forensic reliability under incomplete data.
Key Points
To develop a robust multimodal driver identification framework capable of accurately verifying drivers even during sensor dropouts or data corruption.
Introduced the Dataset for Driver Identification via Multimodal Approaches (DDIMA), combining CAN-bus signals, inertial measurement unit (IMU) readings, and dash-cam video.
Designed a dual-stream deep learning architecture featuring a transformer-based cross-attention module to fuse visual representations with an ensemble of CAN-bus and motion data.
Evaluated system robustness under challenging conditions, including missing modalities and course-based validation across different environments.
Multimodal data fusion significantly outperformed single-modality baselines across all tested configurations.
The transformer-based cross-attention architecture retained high identification accuracy during course-based validation across unseen driving environments.
System performance remained resilient even when individual sensor inputs were unavailable or corrupted.