基于卷积神经网络的不良驾驶行为辨识

Identification of Bad Driving Behavior Based on Convolutional Neural Network

为量化描述驾驶员驾驶行为的动态变化过程与不良程度,研究了不良驾驶行为谱的构建与分析方法,用于不良驾驶行为的实时辨识。首先,基于拉格朗日插值法对轨迹数据清洗处理后提取特征指标参数构建驾驶行为谱,采用风险度量法对急转向、急加速、急减速、超速4种不良驾驶行为进行量化表达。其次,使用大样本统计分布的IQR(interquartile range)与客观赋权的CRITIC(criteria importance though intercrieria correlation)方法确定不良驾驶行为特征指标参数阈值与权重,结合隶属度函数构造模糊综合评价模型对不良驾驶行为谱特征值进行确定以标定不良行驶车辆。最后,将不良驾驶行为谱特征值作为输入,基于人工智能卷积神经网络(convolutional neural networks,CNN)算法对不良驾驶行为进行辨识,并与SVM(support vector machine)、RF(random forest)、BP(back propagation)等传统机器学习算法在辨识误差上进行比较。结果表明:CNN算法对不良驾驶行为辨识的理论误差值MAE(mean absolute error)为0.059、RMSE(root mean squared error)为0.084、R 2高达0.911。可见,不良驾驶行为谱作为一种客观量化不良驾驶行为的方法与CNN算法相结合,能依据车辆运行轨迹对不良驾驶行为进行自动辨识,具有客观性、可靠性与适应性。

In order to quantitatively depict the dynamic fluctuations and gravity of driver behavior,a method for constructing and analyzing a bad driving behavior spectrum was studied for real-time identification of bad driving behavior.Firstly,based on the Lagrange interpolation method,trajectory data was cleaned and feature indicators were extracted to construct a driving behavior spectrum.The risk measurement method was used to quantify four types of bad driving behaviors,including abrupt turning,rapid acceleration,sudden deceleration,and over speeding.Secondly,the interquartile Range of a large sample statistical distribution and the CRITIC(criteria importance though intercrieria correlation)method with objective weighting were used to determine the threshold values and weights of the feature indicators for bad driving behavior.A fuzzy comprehensive evaluation model was constructed using a membership degree function to determine the characteristic values of bad driving behavior spectrum and calibrate bad driving vehicles.Then,based on the bad driving behavior spectrum characteristic values as input,an artificial intelligence CNN(convolutional neural network)algorithm was used to identify bad driving behavior,and the identification error was compared with traditional machine learning algorithms such as SVM(support vector machine),RF(random forest),and BP(back propagation).The results showed that the theoretical error value mean absolute error(MAF)of the(CNN)algorithm for identifying bad driving behavior was 0.059,root mean squared error(RMSE)was 0.084,and R 2 was as high as 0.911.Therefore,combining the bad driving behavior spectrum as an objective quantitative method for bad driving behavior with the Convolutional Neural Networks algorithm can automatically identify bad driving behavior based on vehicle operating trajectory,and has objectivity,reliability,and adaptability.