基于Kohonen-SVM模型的驾驶行为险态动态辨识

On the dynamic identification of the driving risk liability status based on the Kohonen-SVM method

对驾驶行为的危险状态进行动态辨识并提前预警是防止交通事故发生的重要手段。提出一种基于Kohonen神经网络和支持向量机（SVM）的驾驶行为险态动态辨识方法。基于国内外相关研究,选取油门、方向盘转角、刹车、离合、X轴速度、Y轴速度、X轴加速度、Y轴加速度、发动机转速作为驾驶行为状态指标。应用Kohonen神经网络对9个指标组成的向量进行非监督聚类。用聚类结果组成的时间序列表示驾驶员行为指标的动态变化特征并以此作为输入,通过训练SVM实现驾驶行为险态辨识,解决了高维指标数据监督聚类困难和险态识别的静态性问题。最后,采用驾驶模拟器进行试验设计,对方法的有效性进行验证。以8个危险场景作为诱发驾驶行为险态出现的刺激,10个被试共产生8 400组识别序列,选取600组标识为险态的时间序列进行验证。结果表明：该模型的驾驶行为险态识别正确率为82.22%。不同被试的正确率差异控制在6%以下,表明此模型具备一定的泛化能力。

The present paper introduces a driving risk status identifying method based on the Kohonen neural network system of the vector-supporting machine（ short for VSM） in hoping to reduce and prevent the traffic accidents by giving an early warning. For the said research purpose,we have first of all made an investigation of the effects of the 9 indicators on the speed of the X axis and Y axis,that is,the throttle angle,the rotation angle of the steering wheel,the working state of the brake and the clutch by using the engine speed as the vehicle＇s performance given in the up-to-date technical documents. And,then,it would be possible to use the Kohonen neural network for non-supervised clustering group of the nine indicators. At this,it is possible for us to take the time series of the results of the clustering factors as the training set for the vector-supporting machine to drive the risk status identification. In so doing,the method can help to solve the difficulty in supervising the high-dimensional data clustering by joining the Kohonen neural network and the vector-supporting machine. What is more,since there is no need for the generated time series to accumulate a certain amount of priori information in the time window,they can naturally demonstrate the dynamic characteristics of the driver＇s behavior. And,last of all,we have also conducted a driving simulator-based experiment to verify the effectiveness of the aforementioned method,with 8 hazard scenarios being quoted as stimuli for the driving risk behaviors. In the test we have laid out,we have managed to collect the original testing data gained with the 10 participants in the 8 400 sets of testing and identifying sequences via the SVM identification model for the Kohonen neural network system. Furthermore,we have also adopted 600 sets of recognition sequences labeled as the driving risk status-in-situ as the testing set to examine the accuracy rate and the generalization ability of the model. As a result,the best penalty coefficient and the radial basis coefficient of SVM have been found to be 256 and 0. 108 82,respectively,the accuracy rate of which has been proven as high as 82. 22%. Therefore,the gap of the accuracy rate for the different participants has been found merely below 6%,hence,indicating that the model we have adopted has had really significant generalizing power.