Visual measurement network for intelligent cockpit driving state combining lightweight temporal convolution and cross-granularity state mapping
TL;DR
This study proposes LTC-CGMN, a lightweight driver state evaluation framework that integrates cross granularity state mapping and temporal convolution, and shows that LTC-CCMN provides a good balance between recognition accuracy, continuous risk estimation, and edge deployment efficiency for intelligent cockpit DMS applications.
Abstract
The actual deployment of driver monitoring system (DMS) is still limited by the limited edge computing resources and the sudden risk transition caused by discrete state classification. To this end, this study proposes LTC-CGMN, a lightweight driver state evaluation framework that integrates cross granularity state mapping and temporal convolution. Firstly, the CGSM module aligns fine-grained head/hand cues with coarse-grained torso features through cross attention, and then the LTC module uses depthwise separable causal dilation convolution to capture temporal dependencies with low computational cost. In addition, the knowledge extraction scheme guided by the AHP of the National Highway Traffic Safety Administration generates a continuous Driver Risk Index (DRI), reducing the step discontinuity in risk estimation. Finally, the experimental results on the 3MDAD multimodal driving dataset showed that LTC-CGMN achieved a Top-1 accuracy of 96.37%, a DRI prediction RMSE of 0.058, and a inference rate of 33.6 FPS on NVIDIA Jetson Orin. Research has shown that LTC-CCMN provides a good balance between recognition accuracy, continuous risk estimation, and edge deployment efficiency for intelligent cockpit DMS applications.