基于多尺度特征融合的驾驶员注意力分散检测方法

Driver Distraction Detection Based on Multi-scale Feature Fusion Network

近年来,道路交通事故的发生逐年增加。驾驶员注意力不集中是造成交通事故的主要原因之一。该项工作利用多源数据来检测驾驶员是否注意力分散。由于每个数据源能为其余数据源提供一定的信息,即多源数据之间的关联性较强,因此对不同来源的数据进行同等处理或对多源特征进行简单的连接整合会导致特征耦合度高,不能保证挖掘任务的有效性。另外,注意力分散驾驶可能受到许多因素的影响,当已知类别的集合中不存在驾驶员注意力分散的类型时,常见的有监督方法可能会导致分类错误。对此,提出了一种基于多尺度特征融合的驾驶员注意力分散检测方法(Multi-Scale Feature Fusion Network,MSFFN)。首先,通过多个嵌入式子网络从多源数据中学习低维表示。然后,提出一种多尺度特征融合方法,从时空关联性的角度聚合这些特征表示,降低多源特征之间的耦合度。最后,设计基于卷积长短期记忆的编解码模型进行无监督检测。在训练阶段,模型仅对正常驾驶实例进行训练,确定正常数据的一类分类边界。在检测阶段,计算模型重构误差并将其作为每一个测试数据的评分,从而做出细粒度的检测决策。该方法在公开的驾驶员行为数据集上取得了很好的实验结果,优于现有方法。

The occurrence of road traffic accidents has increased year by year.Driver inattention during driving is one of the major causes of traffic accidents.In this paper,we utilize multi-source data to detect driver distraction.However,the correlations derived from multi-source data will generate feature of high-dimensional entanglement.Existing methods perform similar processing for data of different sources or simply stick to concatenate multi-source features,which are not easy to catch the key feature of high-dimensional entanglement.And distracted driving can be affected by many factors.Supervised methods might cause misclassification when the type of driver distraction does not exist in the set of the known categories.Therefore,we propose a multi-dcale feature fusion network approach to tackle these challenges.Basically,it first learns low-dimensional representations from multi-source data through multiple embedding subnetworks,and then proposes a multi-scale feature Fusion method to aggregate these representations from the perspective of spatial-temporal correlation,thereby reducing the entanglement of feature.Finally,we utilize a ConvLSTM encoder-decoder model to detect driver distraction.Experimental results on a public loaded drive dataset show that the proposed method outperforms the existing methods.