EMPIRICAL BENCHMARKING OF A LIGHTWEIGHT MULTI-MODAL FACIAL LANDMARK FRAMEWORK FOR REAL-TIME DRIVER VIGILANCE ON RESOURCE-CONSTRAINED EDGE DEVICES

Abstract

Drowsiness, fatigue, and lack of attention of the driver are among the key factors in road traffic accidents [1], [2]. Despite the high efficiency that end-to-end deep learning methods can provide, the computational needs can limit their deployment on resource-constrained edge devices [4], [6], [11]. In this paper, we propose a lightweight solution to monitor multi-modal aspects of driver vigilance, namely, drowsiness and distraction on resource-constrained edge hardware, called DRIVEX. DRIVEX utilizes MediaPipe Face Mesh to get 468 dense facial landmarks [12], and the following metrics are employed to analyze specific types of driver vigilance: Eye Aspect Ratio (EAR) for eye-closure detection [8], Mouth Aspect Ratio (MAR) for yawning assessment [13], and horizontal facial geometry for distraction estimation. The above features are combined through a lightweight heuristic decision-making engine without the need of a specialized convolutional classification phase at runtime. The whole pipeline was implemented and evaluated on Raspberry Pi 3 B+, CPU-only. Over 520 frames analyzed sequentially, DRIVEX provides 33.42 FPS throughput and 29.21 ms latency on average per frame.

Keywords:

Driver Vigilance, Edge AI, Eye Aspect Ratio, Facial Landmark Mesh, Computer Vision, Real-Time Systems, Benchmarking