Robust Multimodal Driver Identification Using Dash-Cam and In-Vehicle Sensor Data

Abstract

With the increasing number of vehicles and rising concerns about traffic safety and vehicle security, accurate driver identification has become essential not only for forensic analysis and accident investigation but also for secure access control in connected vehicles. However, in real-world scenarios, the availability and reliability of in-vehicle data may vary due to sensor malfunctions or data corruption, leading to degraded performance in single-modality systems. To address this challenge, this paper introduces the Dataset for Driver Identification via Multimodal Approaches (DDIMA), which integrates CAN-bus data, IMU sensors, and dash-cam video footage to enable robust driver identification even under incomplete modality conditions. Comparative experiments demonstrate that the proposed multimodal approach significantly outperforms single-modality methods, especially under stringent evaluation settings such as course-based validation, where adaptability to new environments is critical. Furthermore, our identification method adopts a dual-stream deep learning architecture with a transformer-based cross-attention module to effectively fuse visual and motion information, along with an ensemble of CAN-bus and IMU data to reinforce security and reliability. The proposed DDIMA dataset and methodology provide a strong foundation for developing next-generation driver authentication and identification systems, enabling flexible use of available modalities and ensuring trusted performance across diverse road environments and potential threat scenarios.