【作者】 王国良；

【导师】 梁德群；

【作者基本信息】 大连海事大学 ， 通信与信息系统， 2008， 博士

【摘要】 图像处理技术在智能交通系统中应用的研究,是智能交通系统的重要前沿研究领域,具有十分重要的理论意义和应用价值。图像处理技术在智能交通系统中的应用领域非常广阔,大体上可归纳为基于视觉的智能车辆导航、基于视觉的交通监控和基于视觉的交通管理三大应用领域。本文主要对后两大领域中的若干问题进行了研究,主要包括车辆分割、车辆跟踪、运动车辆的阴影检测和车牌识别等。研究成果如下:提出了一种用于分割运动目标的非参数多模态背景模型。该模型采用分箱核密度估计算法从训练图像序列中得到背景的密度函数。分箱核密度估计算法利用基于网格数据重心的分箱规则,很好地提取了训练图像序列的关键信息,避免了采用全样本数据点的重复计算,大大提高了运动目标分割算法的实时性。通过与全样本算法的对比,表明了该背景模型在运动目标分割中的有效性。固定核带宽的Mean Shift跟踪算法在跟踪尺度逐渐增大的目标时,会同时导致尺度定位偏差和空间定位偏差;跟踪尺度逐渐缩小的目标时,虽然空间定位偏差很小,但是会导致尺度定位偏差。在跟踪快速运动目标时,由于目标区域在相邻两帧间出现不重叠的情况,迭代往往收敛于背景中与目标特征比较相似的区域。因此,本文提出了基于特征匹配的快速运动目标自适应Mean Shift跟踪算法,根据目标的缩放因子对核带宽进行自适应选取,实现尺度定位,目标缩放因子可以通过特征匹配的方法求取;同时,通过特征匹配又可以对迭代的初始中心位置进行确定,从而实现了对快速、尺度可变运动目标的有效跟踪。实验结果验证了该算法的有效性。为了在复杂背景下鲁棒跟踪视频序列中的多自由度运动目标,本文基于粒子滤波理论提出了一种多自由度运动目标的稳健跟踪算法。该算法在核函数下颜色直方图的基础上,对目标的中心位置和表征目标形状的协方差矩阵进行更新,从而自适应地调整了核函数带宽的大小,修正了跟踪窗口的尺寸,实现了对多自由度运动目标的跟踪。在不同场景和不同目标的跟踪实验中,提出的算法能够稳健、可靠地跟踪多自由度运动目标,对目标尺度和角度变化具有良好的适应性。在户外的视频监控系统中,运动目标的阴影降低了系统对目标识别与跟踪的能力。传统的基于像素的阴影检测算法易受噪声的影响。为了提高阴影检测算法的准确性,提出了一种基于区域与光照不变性的运动阴影检测算法。该算法从阴影的物理特性出发,考虑了区域内像素的总体特征。将运动区域采用EM聚类算法进行分块,对其中的小块向邻近的大块进行合并。对其中的每一块,根据阴影区域和相对应的背景区域之间的光照不变性进行阴影检测。实验结果表明,本文算法能够很好地抑制噪声,准确地检测出阴影,明显比基于像素的算法有效。提出了一种基于字符特征匹配的车牌定位与倾斜校正方法。该方法考虑到我国车牌首位字符为汉字的显著特征,利用标准车牌汉字库,采用特征匹配对车牌中的汉字进行定位。由于汉字在我国车牌中的位置严格固定,因此,对汉字的成功定位,也就实现了对整个车牌的定位与倾斜校正。对不同背景、不同光照条件下的车牌进行大量实验,结果表明本文方法能够准确地进行车牌定位与倾斜校正,具有良好的鲁棒性。以上成果构成了本文的主要内容。本论文的研究内容涉及到智能交通中图像处理技术的主要方面,其研究成果是对智能交通中图像处理技术的发展。本文提出的新思想和新方法对智能交通中图像处理技术应用的研究具有指导性意义和应用价值。更多还原

【Abstract】 Today,image processing technology plays an ever-important role in ITS(Intelligent Transportation System).Applying image processing technology to ITS is a challenging field which has great theoretical significance and practical value.In ITS,image processing technology is applied to a variety of areas such as vision-based intelligent vehicle guidance,vision-based traffic surveillance and vision-based traffic management.In this thesis,we focus on the latter two fields which mainly include moving vehicles segmentation,moving vehicles tracking,shadow detection of moving vehicles and license plate recognition.The main contributions of the thesis are as follows:A novel nonparametric multimodal background model is proposed to segment moving objects.The binned kernel density estimators are exploited to estimate the probability density function of background intensity in training sequence.Based on the gravity center of the data points,the binned kernel density estimators describe the key information of the original whole sample set and avoid the repetition computation in the evaluation phase.Compared with algorithm based on the whole samples,the proposed approach is proved to be efficient in traffic surveillance systems.The Mean-Shift algorithm of the fixed kernel-bandwidth is applied to track vehicle changing big in size with the huge error of size and space localization,and changing small in size with the huge error of size localization.At the same time,the Mean-Shift algorithm,strictly depends on the assumption that object regions overlap between the consecutive frames,is applied to track fast motion objects without converging to real place of objects.Therefore,a object tracking algorithm is proposed,this algorithm gets the target’s scale using automatic selection of kernel-bandwidth based on feature matching.At the same time,find the starting position of Mean-Shift iterative through the Feature Matching.Experimental results show that the proposed algorithm can track successfully fast moving objects of changing in size.In order to robustly track the muti-degrees of freedom moving objects in video sequences in the presence of cluttered backgrounds,a tracking algorithm of muti-degrees of freedom moving object is proposed based on the Particle Filter Principle.This algorithm gets the target’s scale using automatic selection of kernel-bandwidth based on updating the position of the object and the covariance matrix that describes the shape of the object.Test results tracking various objects in different scenarios show that the proposed algorithm can track muti-degrees of freedom moving objects,and can adapt to change of the object’s scale and angle.Cast shadows from moving objects reduce the general ability of robust classification and tracking of these objects,in outdoor surveillance applications.Classic pixel-based object shadow detection algorithm limits the performance,owing to noise.A algorithm for segmentation of cast shadows is proposed with improved accuracy,combining region with illumination invariant.This algorithm takes into account the features of all the pixels in a region.Using EM Cluster,the moving region is segmented into blocks with the smaller blocks combined with the neighbor bigger blocks.Shadow detection is performed in every block based on the illumination invariant between the shadow region and the corresponding background region.Experimental results show that the proposed algorithm is the most robust to noise,can detect accurately shadows and is more efficient than the algorithm based on pixel.A novel approach for license plate location and tilt correction is proposed based on the feature matching of character.Considering that the primacy character of the PRC license plates is Chinese characters,this approach gets the Chinese character’s position using standard Chinese characters of the PRC license plates based on feature matching. The Chinese character’s position in the PRC license plates holds fixed,thus,the proposed algorithm achieves license plate location and tilt correction using Chinese characters location.To demonstrate the effectiveness of the proposed algorithm,it conducts extensive experiments over a large number of real-world vehicle license plates. It reports that the proposed approach has high accuracy and robustness.The above contributions constitute the main parts of this thesis.The content of this thesis includes the main aspects of image processing technology in ITS.The contributions of this thesis are further development of image processing technology in ITS.The proposed new idea and methods have great guiding significance to the application research of image processing technology in ITS and have greater practical value.更多还原