E-BTS: A low-power Event-driven Blink Tracking System with hardware-software co-optimized design for real-time driver drowsiness detection

초록

Existing driver drowsiness detection methods based on RGB cameras and neural networks face critical limitations in variable lighting conditions and impose high computational burdens, rendering them unsuitable for real-time, low-power embedded applications. This paper presents an Event-driven Blink Tracking System (E-BTS) that leverages event-based images, which capture changes in light intensity rather than full-frame data, significantly reducing computational requirements. A key innovation of this work is the separation of blink tracking into two tasks: eye blink detection and eye region detection. Dedicated algorithms and hardware modules were developed for each task and integrated into the specialized hardware accelerator, BlinkTrack. By integrating this accelerator into the embedded processor, the developed E-BTS processor achieves real-time performance for blink tracking at 200 FPS. The proposed system was implemented and validated on the FPGA-prototyped E-BTS processor, demonstrating robust performance across diverse lighting conditions and scenarios. Experimental results highlight substantial performance improvements, including a 1275× speedup in blink detection and a 310–496× speedup in eye region detection, while the hardware contributes only 27.2% to the overall processor power consumption. Additionally, quantitative evaluation under varying lighting and noise conditions showed an average F1-score of approximately 86%, which is notably high considering the real-time operation on a low-power embedded platform without relying on neural networks. These results establish the E-BTS as a practical, efficient, and robust solution that effectively overcomes the limitations of conventional methods for real-time drowsiness detection in embedded systems.

키워드