基于深度学习的船舶驾驶员疲劳检测算法

针对日益凸显的船舶值班人员疲劳驾驶问题，为有效预警值班驾驶员的疲劳状态，保障船舶航行安全，研究了基于深度学习的疲劳检测算法。考虑到船舶驾驶台空间大、背景复杂等特点，使用深度可分离卷积改进RetinaFace人脸检测模型，优化模型的检测速度；基于Channel Split和Channel Shuffle思想，结合批量归一化、全局平均池化等技术搭建改进的ShuffleNetV2网络，自动提取图像特征，识别眼睛、嘴巴的开闭状态；根据PERCLOS准则融合眼睛、嘴巴2个特征参数综合判定驾驶员是否疲劳。实验结果表明:改进后RetinaFace模型的检测速度由9.33帧/s提升至22.60帧/s，人脸检测精度和速度均优于多任务卷积神经网络（MTCNN）；改进的ShuffleNetV2网络识别眼睛、嘴巴状态的准确率高达99.50%以上；算法在模拟驾驶台环境中识别疲劳状态的精确率达到95.70%，召回率达到96.73%，均高于目前常见的Haar-like+Adaboost以及MTCNN+CNN疲劳检测算法。算法检测每帧图片仅需0.083 s，基本满足实时检测的要求。 

A Detection Algorithm for the Fatigue of Ship Officers Based on Deep Learning Technique

Aiming at preventing Officers on Watch (OOW) from fatigue driving, a fatigue detection and alert algorithm based on deep learning technique is developed. Considering the large space and complex background of the ship bridge, the RetinaFace model is improved by using Depthwise Separable Convolution to optimize the detection speed. An upgraded ShuffleNetV2 network is then developed by adopting the concepts of Channel Split, Channel Shuffle, and other techniques such as batch normalization and global average pooling. The proposed algorithm can extract image features and automatically identify the opening and closing of the eyes and mouth of the OOW. According to the PERCLOS criteria, the two features of the eyes and mouth are integrated to determine whether the OOW is fatigued. Experimental results show that the detection speed of the improved RetinaFace model improves from 9.33 to 22.60 frames/s. The detection accuracy and speed for the face detection are superior to the multi-task convolutional neural network. The upgraded ShuffleNetV2 network achieves over 99.50% accuracy in recognizing the states of eyes and mouth. The algorithm has an accuracy of 95.70% and a recall rate of 96.73% in identifying the fatigue state in a simulated ship bridge scenario, which are higher than Haar-like+Adaboost and MTCNN+CNN fatigue detection algorithmsused in practice. It only takes 0.083 s for the algorithm to complete the process, which indicates that the algorithm is capable of carrying out real-time detection.