基于深度学习的疲劳状态识别算法  

Fatigue Recognition Algorithm Based on Deep Learning

作　　者：周慧[1] 周良[1] 丁秋林[1] 

机构地区：[1]南京航空航天大学计算机科学与技术学院,南京210016

出　　处：《计算机科学》2015年第3期191-194,200,共5页Computer Science

基　　金：江苏省产学研联合创新资金项目(SBY201320423)资助

摘　　要：目前国内外的疲劳状态识别算法大多数是基于单一的、人为制定的疲劳状态特征实现的,且大部分算法结构复杂,效率比较低下,对驾驶员的个人行为习惯的适应性不强。为此,提出一种基于深度学习的疲劳状态识别算法,它引入深信度网络(DBN)来模拟输入图像数据分布,完成对疲劳特征的分层自动抽取,进而基于时间窗实现视频流图像的疲劳状态识别;同时,算法自适应调整网络学习率以减少网络预训练时间,采用反馈机制实现网络自进化,从而提高对用户个性化疲劳特征的适应性。实验结果表明,算法可以使网络获取到良好的疲劳特征,且误判率会随使用时间的增加而逐渐降低。Current domestic and overseas fatigue recognition algorithms are implemented using fatigue features which are mostly singular and man-made.Most of those algorithms have complex structure,low efficiency and weak adaptability for drivers’ individual behavior habit.To this end,the paper put forward a fatigue recognition algorithm based on deep learning.It introduces deep belief network（DBN）to simulate the data distribution of input images,extracts fatigue features automatically layer by layer,and then recognizes state of fatigue from video images based on time window.The algorithm adjusts the learning rate of the net adaptively to reduce pre-training time,uses feedback mechanism to let the net evolve by itself and as a consequence improves its adaptability for user personalized fatigue features.The experimental result shows that our algorithm acquires good fatigue features,and its misjudgment rate reduces gradually along with incremental time.

关 键 词：疲劳状态识别 深度学习 深信度网络 疲劳特征 反馈机制 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]