考虑视觉特性的分流区换道风险评估

发布时间：2018-08-06 19:52

【摘要】：高速公路分流区是高速公路安全瓶颈,为有效评价分流区安全服务水平,本文借助D-Lab人因数据采集与分析系统对分流区眼动数据进行采集、分析,利用交通冲突技术提取车辆轨迹,构建考虑视觉特性的分流区换道风险评估模型。具体研究内容如下:(1)提出获取视觉特性的试验方案。选取共25位驾驶员开展平行式有可选车道的单车道类型分流区实车试验,获取眼动数据,并利用参照物将注视点的像素坐标转换为唯一固定的二维坐标。(2)实现兴趣区域下眼动数据分析。从基本眼动、目标注视、视线转移三特征分析试验数据,利用二元变量分析将注视点分布划分为两类,选取资深驾驶员注视点分布利用近邻传播聚类算法,通过调控阻尼系数λ与偏向参数p确立聚类数目,依照聚类结果将兴趣区划分为7部分,其中前窗采用放射线划分,以此分析不同驾驶行为间视觉差异,共选取11类差异性显著数据。(3)构建换道决策模型。确定视觉特性指标体系,利用主成分分析法降维,并提出基于支持向量机的非线性驾驶行为分类,以视觉特性参数获取换道概率,判别驾驶行为类型。同时对比四种核函数,结果表明,高斯径向基函数核函数准确率91.67%,灵敏度90.21%,适用于小样本量、低维度情况。(4)提出基于预测轨迹的冲突严重程度判别方法。采用视频检测技术,对固定背景下运动目标进行轨迹提取,利用神经网络实时预测轨迹,并引入量化指标J以碰撞概率分析分流区两车冲突严重程度。同时探讨融合换道决策的碰撞概率算法,构建融合视觉特性的分流区风险评估模型。结果显示,指标J考虑避险行为对潜在冲突点出现时间的影响,准确性更高;融合视觉特性的碰撞概率模型,准确度85.71%、灵敏度92.75%,更接近于实际。
[Abstract]:In order to evaluate the security service level of the diversion area effectively, this paper collects and analyzes the eye movement data of the diversion area by means of the D-Lab Human cause data acquisition and Analysis system. Traffic conflict technology is used to extract vehicle trajectory, and a risk assessment model of diverging area with visual characteristics is constructed. The main contents are as follows: (1) the experimental scheme of obtaining visual characteristics is proposed. A total of 25 drivers were selected to carry out a parallel single-lane diverging area test with optional lanes to obtain eye movement data. The pixel coordinate of the fixation point is transformed into a unique two-dimensional coordinate by using the reference object. (2) the eye movement data analysis under the region of interest is realized. From the basic eye movement, target gaze, line of sight transfer three characteristic analysis test data, using the binary variable analysis to divide the fixation point distribution into two categories, select the senior driver fixed point distribution and use the nearest neighbor propagation clustering algorithm. The number of clusters is established by adjusting damping coefficient 位 and bias parameter p, according to the clustering results, the region of interest is divided into 7 parts, in which the front window is divided by radiation, so as to analyze the visual differences between different driving behaviors. A total of 11 types of significant difference data were selected. (3) the decision model of changing channels was constructed. The visual characteristic index system is determined, the dimension is reduced by principal component analysis, and the nonlinear driving behavior classification based on support vector machine is proposed. The change probability is obtained by the visual characteristic parameter, and the driving behavior type is distinguished. At the same time, the results show that the accuracy of Gao Si radial basis function kernel function is 91.67, the sensitivity is 90.21, and it is suitable for small sample size and low dimension. (4) A method for judging the severity of conflict based on prediction trajectory is proposed. Based on the video detection technique, the trajectory of moving target in fixed background is extracted, and the trajectory is predicted by neural network in real time. The impact probability is introduced to analyze the severity of collision between two vehicles in the split area. At the same time, the collision probability algorithm of fusion path changing decision is discussed, and the risk assessment model of shunt region is constructed based on fusion vision characteristics. The results show that the impact of risk avoidance behavior on the time of occurrence of potential conflict points is more accurate, and the collision probability model with visual characteristics is more accurate, with a sensitivity of 92.755.75, which is closer to reality.
【学位授予单位】：南京理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：U491

【参考文献】