Scholarlink

Driver behaviour monitoring is crucial for intelligent transportation systems, yet affordable and privacy-preserving solutions without cloud dependence remain limited. This paper presents DriveSense, a low-cost embedded driver behaviour detection system that combines an ESP32-CAM in-cabin camera with on-device inference on a companion mobile device. All frames are processed locally, and only compact event metadata is logged, eliminating continuous video streaming and cloud upload. A face-gated region-of-interest pipeline reduces computation and focuses inference on the driver’s face and upper body, enabling real-time operation under embedded constraints. We comparatively evaluate YOLOv8, YOLOv11n, and RF-DETR Nano on 849 images with 884 annotated instances across six behaviours: attentive driving, drowsiness, distraction, phone use, yawning, and alcohol impairment. YOLOv8 offers the best accuracy–latency trade-off (mAP 50 = 0.987 at 2.5 ms per frame), YOLOv11n achieves the highest recall (0.974) with the smallest footprint (2.5 M parameters, 10 MB), while RF-DETR Nano attains the highest precision (0.988) but, at 30 M parameters, exceeds the practical embedded budget. The DriveSense prototype achieves mAP 50 ≥ 0.986, demonstrating accurate, real-time, privacy-preserving driver monitoring on ultra-low-cost and cloud-free hardware.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

DriveSense: Low-Cost, Face-Gated, Privacy-Preserving Driver Behaviour Detection for Embedded Systems

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Chai Ming Jing,
  • Nor Azizah Saadon,
  • Muhamad Najib Zamri

摘要

Driver behaviour monitoring is crucial for intelligent transportation systems, yet affordable and privacy-preserving solutions without cloud dependence remain limited. This paper presents DriveSense, a low-cost embedded driver behaviour detection system that combines an ESP32-CAM in-cabin camera with on-device inference on a companion mobile device. All frames are processed locally, and only compact event metadata is logged, eliminating continuous video streaming and cloud upload. A face-gated region-of-interest pipeline reduces computation and focuses inference on the driver’s face and upper body, enabling real-time operation under embedded constraints. We comparatively evaluate YOLOv8, YOLOv11n, and RF-DETR Nano on 849 images with 884 annotated instances across six behaviours: attentive driving, drowsiness, distraction, phone use, yawning, and alcohol impairment. YOLOv8 offers the best accuracy–latency trade-off (mAP 50 = 0.987 at 2.5 ms per frame), YOLOv11n achieves the highest recall (0.974) with the smallest footprint (2.5 M parameters, 10 MB), while RF-DETR Nano attains the highest precision (0.988) but, at 30 M parameters, exceeds the practical embedded budget. The DriveSense prototype achieves mAP 50 ≥ 0.986, demonstrating accurate, real-time, privacy-preserving driver monitoring on ultra-low-cost and cloud-free hardware.