基于注意力机制改进的疲劳驾驶检测方法

Improved fatigue driving detection method based on attention mechanism

由于疲劳驾驶采集过程中存在因识别角度不佳,部分区域遮挡等原因,在不同时间段丢失了不同特征的时间信息,导致算法的泛用性较差。此外,驾驶疲劳的检测需要在保证综合准确率的同时,需要具有更低的漏检率。针对以上问题,提出了一种基于注意力机制和长短期记忆(LSTM)神经网络的疲劳驾驶检测模型。通过对不同特征定位点计算多维特征向量,并对特征向量的时间序列进行学习,同时引入注意力机制,赋予各维度隐含状态不同的概率权重,加强重要信息对疲劳状态判定的影响和降低丢失特征信息的历史数据对参数的影响。根据实验可得,该方法在更普遍的检测环境下有着92.19%的准确率和1.9%的漏检率,同时在丢失部分特征的环境下漏检率仅有3.07%。

Due to the poor recognition angle and partial area occlusion in the process of fatigue driving acquisition,the time information with different characteristics is lost in different time periods,resulting in poor universality of the algorithm.In addition,the detection of driving fatigue needs to not only ensure the comprehensive accuracy,but also have a lower missed detection rate.To solve the above problems,a fatigue driving detection model based on attention mechanism and long short-term memory(LSTM)neural network is proposed.By calculating multi-dimensional feature vectors for different feature localization points and learning the time series of feature vectors,the attention mechanism is introduced to give different probability weights to the hidden states of each dimension,so as to strengthen the influence of important information on the determination of fatigue state and reduce the influence of historical data losing feature information on parameters.According to the experiment,this method has accuracy of 92.19%and missed detection rate of 1.9%in more general detection environment,and the missed detection rate is only 3.07%in the environment where some features are lost.