基于神经网络的车辆交通协调性评价模型

Evaluation Model of Harmony with Traffic Based on Neural Network

摘要:

为了研究自动驾驶汽车交通协调性的主客观映射评价模型，以高速公路匝道汇入为研究场景，首先基于自然驾驶数据的交互样本数据，以自车平均行驶速度、并线时刻侧向速度、并线时刻车头TTC、并线时刻两车相对侧向速度和对手车减速程度等客观指标数据和交通协调性主观评价结果作为模型输入和输出，构建映射评价模型；然后设计2×2交叉对比实验，并分析数据预处理方法及神经网络类型对评价模型效果的影响。研究结果显示，基于线性函数归一化处理的BP神经网络模型和Dropout神经网络模型总精度分别为95.71%和80.00%，基于阶梯函数归一化处理的BP神经网络模型和Dropout神经网络模型总精度分别为94.60%和73.25%。由此可见，模型评价效果较好，所建立的客观表征指标集能够较好地表达专家对车辆交通协调性表现的评价。在建模方法方面，基于BP神经网络的映射评价模型的表现优于Dropout神经网络模型，能够根据客观数据更准确地得到符合专家评判标准的评价结果。在样本数据预处理方法方面，线性函数归一化处理方法能够在消除数据间数量级差异的同时保留客观表征指标集数据在不同样本之间的差异，能提高映射评价模型的评价精度。因此，在小样本数据量下，基于线性函数归一化和BP神经网络的映射评价模型构建方案的评价效果更优。

Abstract:

In order to study the subjective and objective mapping evaluation model of harmony with traffic for autonomous vehicles， this paper takes the highway on-ramp as a research scenario. First， based on the interaction sample data of natural driving data， a mapping evaluation model is constructed with objective index data and subjective evaluation results as model inputs and outputs， and objective indexes include average driving speed， side velocity， TTC， relative side velocity， deceleration degree of critical traffic participant and so on. Then， the influence of data preprocessing method and neural network type on the evaluation effect is analyzed using the 2×2 cross-comparison test. The results show that the precision of the BP neural network model and the Dropout neural network model based on Min-Max scaling is 95.71% and 80.00% respectively， and the precision of the BP neural network model and the Dropout neural network model based on staircase-function normalization is 94.60% and 73.25% respectively. It can be seen that the evaluation effect of the model is good， and the objective index set can express the evaluation of experts on the harmony with traffic. In terms of modeling method， the performance of the mapping evaluation model based on BP neural network is better than that of the Dropout neural network model， which can get more accurate evaluation results. In terms of data preprocessing method， the Min-Max scaling method can eliminate the magnitude difference between the data and keep the difference of objective index data between different samples， which can improve the evaluation accuracy of the mapping evaluation model. Therefore， when the sample size is small， the evaluation effect of the mapping evaluation model based on Min-Max scaling and the BP neural network is better.