| 摘 要: 针对面部检测精度不足、车载终端算力限制等问题,提出一种改进YOLOv8的检测算法。首先将MobileNetV3-Small作为主干网络,实现模型轻量化;引入SE(Squeeze-and-Excitation)注意力机制与跨尺度融合模块(Cross-Scale Feature Fusion Module,CCFM),增强对小目标的检测能力;结合BoT-SORT算法对面部特征进行实时ID追踪;再融合Dlib关键点,根据单位时间内打哈欠数量、点头数、闭眼时间以及PERCLOS值综合判定疲劳。实验结果表明:相比原YOLOv8算法,改进算法体积降低32.3%,平均精度均值(mAP50)和查全率分别提升至97.7%和96.1%。验证了该算法对面部特征检测的鲁棒性和可靠性。 |
| 关键词: YOLOv8 疲劳驾驶检测 轻量化 BoT-SORT跟踪算法 |
|
中图分类号:
文献标识码: A
|
| 基金项目: 基于有限元分析的交叉韧带生物力学特性与膝骨关节炎致病机理的研究(黔科合基础-ZK[2023]一般052) |
|
| Lightweight Fatigue Driving Detection Based on Improved YOLOv8 |
|
GUO Aoxiang, CHEN Yajiang
|
(School of Sciences, Zhejiang Sc-i Tech University, Hangzhou 310018, China)
371080335@qq.com; yjchen@zstu.edu.cn
|
| Abstract: To address issues such as insufficient facial detection accuracy and computational limitations of vehicle-mounted terminals, an improved YOLOv8 detection algorithm is proposed. First, MobileNetV3-Small is adopted as the backbone network to achieve model lightweighting. The Squeeze-and-Excitation ( SE) attention mechanism and Cross-Scale Feature Fusion Module (CCFM) are introduced to enhance smal-l target detection capability. The BoT-SORT algorithm is then integrated to enable rea-l time ID tracking of facial features. Finally, Dlib keypoints are fused to comprehensively determine fatigue states based on metrics including the number of yawns, head nods, eye closure duration, and PERCLOS values within a unit timeframe. Experimental results show that, compared to the original YOLOv8, the improved algorithm reduces model size by 32.3% while increasing mean Average Precision (mAP50)and recall to 97.7% and 96.1% , respectively. This validates the robustness and reliability of the algorithm for facial feature detection. |
| Keywords: YOLOv8 fatigue driving detection lightweight BoT-SORT tracking algorithm |
