A Deep Multichannel Network Model for Driving Behavior Risk Classification

Driving behavior risk classification is a crucial issue in transportation systems because the prediction of vehicle hazard levels in advance can effectively reduce the occurrence of unnecessary traffic accidents. This paper proposes a novel Deep Multichannel Network Model (DMNM) for driving behavior risk classification. Based on real historical driving data, we present a driving behavior portrait framework with multidimensional factors and a dimensionless loss polymerization method. In this approach, first, we divide the related factors into three dimensions by their work methods, which are the inputs of the network. Second, the data of three original dimensions are extracted through fully connected layers to obtain embedding dimensions. Third, a new dimension indefensible factor is obtained by filter operation to eliminate the correlation between external factors and internal factors, which denotes the degree that the internal abnormal operational behavior can be explained by the external environment. Last, we regard the driving information as four channels of driving behaviors and extract the channel information through convolutional neural networks (CNNs). The numerical calculation and comparison results with real traffic data demonstrates the superior performance of our framework, and the accuracy of the proposed method in vehicle hazard classification is 95%.