In view of the fact that the current fatigue driving detection methods based on computer vision can rarely be applied to the driver's side and oblique profile status, and the driver's video angles collected in the complex real environment may be different, a fatigue driving detection method based on the driver's side states was proposed to solve this problem. The method was applicable to the driver's front, side and oblique side states from various angles. The feature states of eyes, mouth and head were directly extracted and distinguished under each angle state, and the related fatigue indexes were calculated. After that, the multiple indexes calculated by multi-features were integrated by weighted calculation method, and the fatigue discrimination of drivers under front, side and oblique side states was realized. The feature extraction method of eyes, mouth and head was improved based on YOLOv8, and then the state identification of eyes and mouth was realized by MobileNetV3, the three-dimensional head pose was calculated by coordinate system transformation, and the optimal weight allocation of multiple indicators was calculated by analytic hierarchy process. To realize multi-index comprehensive analysis of driver fatigue state under front, side and oblique side conditions. The experimental results showed that the accuracy of the method was 94.8% on average, which can be applied to the driver fatigue detection under various side conditions.
当前疲劳驾驶检测多基于驾驶员正面状态,通过分析眼睛、嘴巴和头部姿态等特征来判断疲劳程度。例如,梁元辉等[3]提出了一种多特征融合的眼睛状态检测算法,通过融合眼睛纵横比 (eye aspect ratio,EAR)、虹膜特征和眼角张开角度,实现疲劳驾驶检测。朱艳等[4]结合人脸跟踪算法提取面部特征点,通过曲线拟合和归一化指标计算眼睛和嘴巴的开合度,并利用卷积神经网络 (convolutional neural network,CNN)建立疲劳状态识别模型。郑伟成等[5]使用改进的多任务级联卷积神经网络进行人脸检测,并提取眨眼频率、嘴部张开程度和点头频率等疲劳特征。Zhu等[6]通过轻量化人脸检测模型和轻量化人脸关键点检测模型提取面部特征并评估疲劳状态。Savas等[7]使用多任务CNN结合Dlib算法计算眼睑闭合时间百分比(percentage of eyelid closure over the pupil over time,PERCLOS)和打哈欠频率,从而判定疲劳状态。现有研究主要针对正面状态,但这些方法存在角度受限、易受遮挡、对光照变化敏感以及实时性不足等问题。侧面检测能够更好地适应驾驶员头部的转动,减少遮挡影响,对光照变化的鲁棒性更强,因此,本文中提出一种基于侧面特征的疲劳驾驶检测方法。依据图1中展示的眼部、嘴部和头部特征来判别驾驶员的疲劳状态[8]。为了验证方法的有效性,构建了一个包含0°、30°、60°、90°不同角度的正常和疲劳驾驶状态的数据集,并使用该数据集进行了实验验证。
检测时,首先输入驾驶员视频,再将连续视频分为连续帧图像,在帧图像中分别使用YOLOv8-ROI网络提取眼部感兴趣区域(region of interest,ROI)、嘴部ROI并使用YOLOv8-Keypoints网络提取面部重要关键点;在眼部ROI、嘴部ROI提取完成后,通过MobileNetV3网络判别眼部状态、嘴部状态;在面部关键点提取完成后,通过坐标系转换的方法计算判别头部姿态,并根据3种特征判别结果计算疲劳指标,最后根据层次分析法计算出的最优权重做多指标融合,综合多指标分析判别驾驶员疲劳状态。具体方法如图2所示。
眼部、嘴部在所有角度下特征提取实验综合结果如表2所示。根据实验结果可知,本文中提出的YOLOv8-ROI模型能够有效用于提取驾驶员侧脸的眼部和嘴部特征提取,模型在准确性和实时性方面都具有很好的性能。0°、30°、60°、90°下模型综合精确率达到了95.3%,YOLOv8-ROI模型的每秒帧数(frames per second,FPS)达到197帧/s,仅次于Dlib方法的261帧/s,表现出良好的实时性,适用于驾驶员疲劳驾驶检测。
在眼部和嘴部特征状态判别阶段,使用了开源的野外闭眼数据集(closed eyes in the wild,CEW)和打哈欠检测数据集(yawning detection dataset,YawDD)以及自建的侧面多角度数据集。CEW提供了眼部开闭样本,YawDD提供了嘴部开闭样本,自建数据集则提供了多角度的眼部和嘴部样本。所有样本被分为训练集(9 080张)、验证集(1 442张)和测试集(1 135张),这些数据集包含不同光照和角度下的图像,有助于准确判别眼部和嘴部的状态。
AMODIOA, ERMIDOROM, MAGGID, et al. Automatic detection of driver impairment based on pupillary light reflex[J]. IEEE Transactions on Intelligent Transportation Systems, 2019, 20(8): 3038-3048.
ZHUJ W, MAY E, XIAJ, et al. A lightweight fatigue driving detection method based on facial features[J]. Signal, Image and Video Processing, 2024, 18(1): 335-343.
[7]
SAVAŞB K, BECERIKLIY. Real time driver fatigue detection system based on multi-task ConNN[J]. IEEE Access, 2020, 8: 12491-12498.
KAURN, RANIS, KAURS. Real-time video surveillance based human fall detection system using hybrid haar cascade classifier[J]. Multimedia Tools and Applications, 2024, 83(28): 71599-71617.
[10]
PARVINS, ROZARIOL J, ISLAMM E. Vision-based on-road nighttime vehicle detection and tracking using taillight and headlight features[J]. Journal of Computer and Communications, 2021, 9(3): 29-53.
[11]
JIANGB, LIN, CUIX, et al. Research on facial expression recognition algorithm based on improved MobileNetV3[J]. EURASIP Journal on Image and Video Processing, 2024, 2024: 22.
[12]
CHENGG, CHAOP, YANGJ, et al. SGST-YOLOv8: an improved lightweight YOLOv8 for real-time target detection for campus surveillance[J]. Applied Sciences, 2024, 14(12): 5341.
[13]
WANGT, RENS, ZHANGH.Nighttime wildlife object detection based on YOLOv8-night[J]. Electronics Letters, 2024, 60(15): e13305.
[14]
LIY S, WANGJ R, ZHANGK H, et al. Lightweight object detection networks for UAV aerial images based on YOLO[J]. Chinese Journal of Electronics, 2024, 33(4): 997-1009.
[15]
BOCHKOVSKIYA, WANGC Y, LIAOH Y M. YOLOv4: optimal speed and accuracy of object detection[DB/OL]. (2020-04-23)[2025-01-15].
[16]
GEZ, LIUS, WANGF, et al. YOLOX: exceeding YOLO series in 2021[DB/OL]. (2021-07-18) [2025-01-25].
[17]
LIC, LIL, JIANGH, et al. YOLOv6: a single-stage object detection framework for industrial applications[DB/OL]. (2022-09-07)[2025-01-15].
[18]
WANGC Y, BOCHKOVSKIYA, LIAOH Y M. YOLOv7: trainable bag-of-freebies sets new state-of-the-art for real-time object detectors[C]//2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Vancouver, BC, Canada: IEEE, 2023: 7464-7475.
[19]
YANGL, ZHANGR Y, LIL, et al. SimAM: a simple, parameter-free attention module for convolutional neural networks[J]. PMLR, 2021, 139: 11863-11874.
[20]
FERRARIOV F, SFORZAC, SERRAOG, et al. Active range of motion of the head and cervical spine: a three-dimensional investigation in healthy young adults[J]. Journal of Orthopaedic Research, 2002, 20(1): 122-129.
[21]
StatesUnited, Federal Motor Carrier Safety Administration, DivisionTechnology. PERCLOS: a valid psychophysiological measure of alertness as assessed by psychomotor vigilance: FHWA-MCRT-98-006[R]. Washington, DC: Federal Motor Carrier Safety Administration, 1998.
[22]
MAY H, YUANG W, YUEK, et al. CJS-YOLOv5n: a high-performance detection model for cigarette appearance defects[J]. Mathematical Biosciences and Engineering, 2023, 20(10): 17886-17904.
[23]
WANGC Y, BOCHKOVSKIYA, LIAOH Y M. Scaled-YOLOv4: scaling cross stage partial network[C]//2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Nashville, TN, USA :IEEE, 2021: 13024-13033.
[24]
LIK N, GONGY B, RENZ L. A fatigue driving detection algorithm based on facial multi-feature fusion[J]. IEEE Access, 2020, 8: 101244-101259.
[25]
RENS Q, HEK M, GIRSHICKR, et al. Faster R-CNN: towards real-time object detection with region proposal networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 39(6): 1137-1149.
[26]
LIH, WANGY B, NAN Y. Motion fatigue state detection based on neural networks[J]. Computational Intelligence and Neuroscience, 2022, 2022: 9602631.