A Model Fusion-based Distracted Driving Behavior Detection Method

Safe driving has always been a hot issue of traffic research,and has great research significance for public safety and economic application.In recent years,with the continuous upgrading of chip manufacturing technology,the high-performance computing chip facilitates the research and engineering application of complex and high-efficiency algorithms,enables the complex algorithms to run on more compact chip platforms,and makes the driving assistance systems more intelligent.Because driver’s distracted driving behavior is the most principal factor leading to traffic accidents,standardizing driving behavior is the most important task of the assistant systems in order to cope with the complicated road environment,reduce the accident occurrence probability and protect the life safety of passengers.This dissertation aims to research the key techniques of the assistant systems to restrain the driving behavior,and mainly focuses on the distracted driving behavior.The principal contributions and innovative works are summarized as follows:1.Driver’s behavior characteristic is a kind of limb synthesis feature.Convolved Neural Networks(CNNs)can extract high efficient features,and abstractly generalize the contour posture,and thus,express the information of the whole image with less data.Because CNNs are limited by the connection of local features,it is difficult to correlate long-distance features from continuous supervised learning.The vision transformer(Vi T)networks realize the feature connection of remote regions of an image,which inherit the attention mechanism of the transformer networks,but the inherent defect of incomplete feature extraction is exposed.Based on the above problems and the characteristics of distracted driving behavior detection,the innovative works are as follows: 1)a fusion model,namely residual vision transformer(RVi T)network,synthesizes the correlation between the excellent feature extraction efficiency of Residual Networks(Res Net)and the long-distance feature of Vi T; 2)proposing an serialized image filtering processing(i.e.,local average pooling of sequence blocks)to solve the problem of over-fitting and accuracy degradation caused by too many model parameters.Compared with convolution networks and visual transformer networks,RVi T improves the accuracy and stability.2.For the imbalance of the number of categories in training data,an isolated center loss(ICL)is proposed to enlarge the discrimination of deeply learned features under the conventional Softmax loss(SL).Based on the principle that the inter-class dispersion should be as large as possible and the intra-class diversity should be as small as possible,ICL is composed of three parts: 1)the first part adopts fixed weights with equal-angle distribution to minimize the sum of cosine values of all angles between classes and ensure the maximum distance of different classes in angle domain; 2)the second part is the idea of center clustering,which minimizes the Euclidean distance between each sample and the center of its class,and urges the same kind of samples to gather together as much as possible; 3)the third part maximizes the Euclidean distance between different classes so that the samples of different classes are separated in Euclidean space as much as possible.Compared with the conventional SL,ICL not only improves the accuracy but also is more stable(the variation of multiple experimental results under the same experimental settings is less).ICL runs only slightly slower than the SL loss method and is still faster than some other state-of-the-art methods.3.Based on the innovative research contents of this dissertation,we design an engineering simulation system.This system firstly estimate image noise level,and applies a denoising step to reduce the influence of the image noise on recognition accuracy if the noise level is greater than a given threshold.Then,it employs our proposed model to recognize driving behavior.This system accords with the research purpose of distracted driving detection algorithm,and solves the problems existing in the model.