交通场景中运动目标检测和车牌复原算法的研究
文献类型：学位论文
| 作者 | 刘玉兰 |
| 学位类别 | 工学博士 |
| 答辩日期 | 2009-05-24 |
| 授予单位 | 中国科学院研究生院 |
| 授予地点 | 中国科学院自动化研究所 |
| 导师 | 彭思龙 |
| 关键词 | 运动检测 图像配准 图像复原 退化模型 活动轮廓模型 motion detection image registry image recovery degradation model active contour model |
| 其他题名 | Motion Detection and License Plate Recovery in Real Traffic Videos |
| 学位专业 | 模式识别与智能系统 |
| 中文摘要 | 车牌图像复原是智能交通系统中的重要研究内容，可辨认的车牌图像能为交通事故的追查提供重要的线索和证据。本文对交通视频中序列图像复原的相关技术进行了研究，包括运动目标检测、图像配准和序列图像复原，在分析现有文献的基础上对以上算法进行了深入的研究，并根据实际问题提出了改进。 首先，针对现有运动目标检测算法容易受到噪声影响，并且容易在目标内部产生空洞的问题，提出了基于自适应阈值的图像块重心差分法对目标进行初始分割。重心位置代表了图像块中灰度的总体分布，因此受噪声的影响很小；自适应阈值则是由两个因素决定：当前图像块的重心位置在时间域上的差分值，及当前图像块周围目标图像块的个数。和固定阈值相比，自适应阈值能有效解决目标内部的空洞问题。 其次，针对活动轮廓模型的控制点容易产生聚集或受到噪声的“牵引”而停留在错误的位置，提出了用改进的活动轮廓模型与Mean Shift相结合的算法对块差分法的结果做进一步分割。该算法以改进模型的负能量值作为各个像素点的权重，计算控制点的Mean Shift向量，并进行Mean Shift迭代，直至整个轮廓收敛。实验证明该方法能有效抑制噪声对控制点的“牵引”，并保持曲线的连续性和光滑性，防止控制点聚集。 再次，在序列图像的粗配准过程中，针对现有加权光流匹配法所用的权值函数为非自适应的核函数，无法适应图像中不同特征的区域，提出了基于图像梯度的多尺度方法来估计自适应的加权核函数。实验表明自适应加权核函数光流法估计出的运动矢量具有较小的平均角度误差和平均幅度误差。 最后，针对现有图像复原算法需要事先估计模糊核，因此无法考虑时时变化的运动模糊和大气扰动等因素的影响，提出了一种线性图像退化模型，通过用线性算子来表示模型中的各种退化，如亚像素平移、形变及模糊等因素，把对降晰参数的估计转化为求线性方程组，从而可以和高质量图像的估计可以交替进行。实验证明，用最大后验概率的方法对基于上述退化模型的序列图像复原能得到较好的结果。 |
| 英文摘要 | License plate image recovery is an important research area in Intelligent Transportation Systems. A clear, identifiable license plate can supply important clue and proof to investigation of traffic accident. This paper studies main techniques of sequence image recovery in real traffic videos, which includes motion detection, image matching and sequence image recovery. First, with the two problems met in the existing motion detection algorithms: influence of noise and holes inside objects, this paper proposes a new block barycenter difference method based on self-adaptive threshold to find the rough object area. Barycenter presents the light distribution of an image block, and is rarely influenced by noise. The self-adaptive threshold is determined by two factors: the barycenter difference in time domain and the number of object block around current image block. Unlike the fixed threshold, the self-adaptive threshold can solve the problem of holes inside objects. Second, with the problem that control points of an active contour model tend to aggregate or tend to be trapped by noise and stop in the wrong location, this paper proposes a new algorithm based on improved active contour model and Mean Shift to make further separation of the result of image block difference. In the new algorithm, negative energy of the improved model is used as weight of each pixel. Then Mean Shift vector of each control point is computed and Mean Shift iteration is taken until the contour converges. Experiments show that the new algorithm can suppress the traction of noise and prevent aggregation of control points, keeping the contour smooth. Thirdly, optical flow method is frequently used for sequence image registry in pixel level. But the weight functions used in these algorithms are not self-adaptive and can not adapt to the different feature areas in an image. With this problem, the paper proposes a multi-scale method based on gradient to estimate self-adaptive kernel functions. Experimental results show that optical flow method with a self-adaptive kernel function can get smaller average angle error and smaller average magnitude error, compared with the unimproved method. Finally, in the existed image recovery algorithms, blur kernels are always estimated beforehand, then the influence of motion blur and atmospheric disturbance which changes with time can not be considered. With this problem, the paper proposes a linear degradation model. In this model, image degradation... |
| 语种 | 中文 |
| 其他标识符 | 200518014628070 |
| 源URL | [http://ir.ia.ac.cn/handle/173211/6153]  |
| 专题 | 毕业生_博士学位论文 |
| 推荐引用方式 GB/T 7714 | 刘玉兰. 交通场景中运动目标检测和车牌复原算法的研究[D]. 中国科学院自动化研究所. 中国科学院研究生院. 2009. |
入库方式： OAI收割
来源：自动化研究所
浏览0
下载0
收藏0
其他版本
除非特别说明，本系统中所有内容都受版权保护，并保留所有权利。