【作者】 谭论正；

【导师】 夏利民；

【作者基本信息】 中南大学 ， 交通运输工程， 2014， 博士

【摘要】 随着汽车保有量的急剧增加,交通安全问题日益突显。驾驶行为建模是自主驾驶和车辆安全辅助驾驶的关键技术之一,对减少交通事故、提高交通安全具有重大的理论意义和应用价值,本文围绕基于视觉的驾驶行为建模的关键技术进行了深入研究。论文的主要工作和创新点如下：(1)提出了一种基于序列图像的交通场景三维重建方法。采用EM算法进行静态点集和基础矩阵估计,E步利用极线约束和基础矩阵检测交通场景中的静态特征点,M步根据静态点采用8点法估计基础矩阵；利用奇异值分解和绝对二次曲线约束估计投影矩阵；提出了基于PCA的交通运动场景点的重建方法,使用PCA确定轨迹基元,将运动场景点的轨迹看作是轨迹基元的线性组合,利用最小二乘法求得场景点的三维轨迹,实现交通运动场景点的重建。该方法省去了摄影重建的过程,大大减少了计算复杂度。实验结果表明了该方法的有效性。(2)提出了基于协方差描述子的交通场景理解方法。为了克服基于单一特征的交通场景分割与识别的不足,采用交通场景的运动结构特征、纹理和颜色特征,并利用协方差描述子进行多特征融合,以消除特征冗余以及不同特征数值悬殊对图像分割的影响；使用多类LogitBoost分类器进行交通场景分割与识别。实验结果表明该方法有效地提高了交通场景分割与识别的效果。(3)将视觉注意机制引入交通险情的检测,提出了基于视觉注意力模型的交通险情的检测方法。为了提高检测速度,采用半球形稀疏采样法减少了计算量；使用贝叶斯概率模型和高斯核函数对交通视频进行非参数显著性度量,分析视频显著性,采用多尺度显著图计算方法以提高检测精度。实验结果表明该方法能有效检测交通险情。(4)提出了基于贝叶斯模型的驾驶行为建模方法,以贝叶斯概率模型作为驾驶模型,根据交通场景和车辆自身的速度、位置信息,预测正常情况下的驾驶行为；利用稀疏贝叶斯学习方法对模型参数进行估计。该模型实现了直行、变道、加速、减速等多种驾驶行为的预测。实验结果证明本文方法有较好的预测性能。(5)提出了基于模糊规则的险情驾驶行为建模方法。根据驾驶人的经验,建立了7种情况险情驾驶行为模糊规则集,结合高斯函数和Sigmoid函数给出了混合型隶属函数,并利用C均值聚类法和梯度下降法对模糊规则模型参数进行估计。实验结果证明该方法能很好的描述人的驾驶行为,可用于各种险情的驾驶行为决策。更多还原

【Abstract】 Transportation safety problems have become more and more obvious with rapid growth of vehicle population. Driving behavior modeling is one of the key technologies of autonomous driving and safety vehicle assistant driving, which has great theoretical significance and applied value in reducing accidents and improving traffic safety. In this dissertation, the exploratory research work has been done regarding driving behavior modeling based on vision, which include3D traffic scene reconstruction, traffic scene understanding, dangerous traffic events detecting, and driving behavior modeling.The main contributions of this dissertation are summarized as follows:(1) The3D reconstruction method of traffic scene based on image sequence is proposed. The EM algorithm is adopted to estimate static feature points and the fundamental matrix. In step E, static feature points of traffic scene are detected by using epipolar constraint and fundamental matrix. In step M,8-points algorithm is applied to compute the fundamental matrix based on the static feature points.This method also estimates Projection Matrix by using singular value decomposition and the restriction of absolute conic. The reconstruction method of traffic scene’s moving points based on PCA is proposed. This method determines the trajectory basis by using PCA, and the trajectory of scene’s moving points can be seen as the linear combination of trajectory basis. The reconstruction of traffic scene is achieved by calculating the3-D trajectory using least squares method. This method omits the process of projective reconstruction and reduces the computational complexity distinctly. Experimental results prove the effectiveness of this method.(2) An approach for traffic scene understanding is proposed based on covariance descriptor. In order to overcome the drawback of segmenting and recognizing the traffic scene based on single feature, this method adopts the movement structure features, texture and color features in traffic sense, and uses covariance descriptors to integrate multi-feature for eliminating feature redundancy and effects on image segmentation causing by the numerical disparity of different features. The multiclass LogitBoost classifier is used for image segmentation to improve the accuracy of segmentation. Experimental results show that this method can effectively improve the effect of traffic scene segmentation and recognition.(3)The method is proposed to detect dangerous traffic events based on visual attention models. The Hemispherical sparse sampling method has been adopted to improve the detecting speed; The Bayesian probability models and Gaussian kernel function has been used to do nonparametric saliency measure of video in order to analyze visual attention; The calculation method base on multi-scale saliency map has been used to improve the detection accuracy. The experimental results show that this method can effectively detect traffic danger.(4) The method of driving behavior modeling based on Bayesian model is proposed. This method can predict the appropriate driving behavior based on the traffic scene and the car’s own speed, location information. Besides, the model parameters are estimated by adopting the sparse Bayesian learning method. This model achieves the prediction of several driving behaviors such as going straight, changing lanes, accelerating, decelerating and so on. The experimental results show that our method has a good performance on driving behavior predicting.(5) The method to model driving behaviors is put forward based on fuzzy rules under a dangerous traffic environment. The fuzzy rules of seven driving behavior is established according to the driver’s experience. It gives a mixed membership function by combining Gaussian function with Sigmoid function and an estimation of the model parameters of fuzzy rules by adopting a C-means clustering method and gradient descent. The experimental results show that this method can be very good for driver’s behavior description and can be used in all kinds of driving behavior decisions when it is dangerous.更多还原