【作者】 王曾敏；

【导师】 杨兆选；

【作者基本信息】 天津大学 ， 信号与信息处理， 2009， 博士

【摘要】 随着现代经济的高速发展,智能交通系统的研究倍受关注。论文工作研究的视频车辆检测系统是智能交通系统的重要组成部分,它通过对道路现场视频图像序列的分析与处理,实现道路交通信息的自动检测和车辆特征的自动识别。论文工作结合科研项目的需求,较深入地研究了基于小波变换的图像去噪、配准和融合等多种算法;设计了道路现场视频车辆检测系统,该系统含有基于DSP的视频处理卡和工控机,其中的视频处理卡用于从视频图像中提取各种交通信息,工控机辅助用户查看和管理交通信息;针对小波图像融合运算量大和视频检测系统实时性要求之间的矛盾,论文探索性地把小波图像融合算法应用于视频车辆检测系统的设计,提高了检测系统的检测速度和正确率,基本上满足了视频信息检测的实时性要求。论文做了以下具有创新性的工作:(1)提出了小波模极大值阈值去噪算法和改进的角点特征的图像配准算法。去噪算法利用信号和噪声的小波模极大值随尺度传播的特性不同,采用特定的阈值区分小波模极大值点是否由噪声产生,达到去噪的目的;改进的图像配准算法利用梯度模值从小波分解后高频分量中提取角点,用特征点灰度信息匹配法剔除误匹配的角点对后,确立仿射变换方程。实验表明,论文提出的去噪算法对噪声的依赖性小,在去噪的同时较好的保留了图像的细节信息;改进的图像配准算法比传统算法的配准精度更高,且具有较低的运算复杂度。(2)提出了基于局部梯度平均模值的图像融合算法,利用小波变换后两幅图像低频系数的局部梯度平均模值确定低频融合系数;改进了传统的高频系数局部方差融合算法,利用局部方差定义高频系数的区域匹配度,通过区域匹配度确定高频融合系数。实验表明,论文提出的融合算法有效克服了传统融合算法对图像细节表现力不足和容易造成模糊的缺点,融合图像具有良好的视觉效果。(3)提出了区域化图像融合设计方案,即在设计视频车辆检测系统时,针对小波图像融合算法计算量大、难以实时处理的难点,只对虚拟检测线覆盖的图像区域进行融合,而不处理与车辆检测无关的区域。与对整幅图像进行的融合相比,区域化图像融合只关注感兴趣区域的图像信息,具有运算量小等特点,适用于对实时性要求高的系统。实验表明,将区域化图像融合应用于视频交通信息检测系统,较好地解决了小波图像融合算法计算量大难以进行实时视频处理的难题。更多还原

【Abstract】 With the development of world economics, intelligent transportation system has become more and more significant. As an important part of intelligent transportation system, vehicle detection system can detect and analyze traffic information automaticly through video processing.This dissertation researches the wavelet domain image denoising, registration and fusion algorithm; it designs a vehicle detection system which contains DSP based video processing cards and industrial personal computers, video processing cards are used to extract traffic information from video and image, industrial personal computers help users view and manage traffic information. Focusing on the conflict between computational complexity of wavelet domain image fusion algorithm and real-time capability of vehicle detection system, this dissertation has exploringly applied the wavelet domain image fusion algorithm to vehicle detection system, which improves the detection accuracy and meets the time-constraints of system requirements. The innovations involved in the research are as follow:(1) A wavelet modulus maximum threshold denoising method and an improved image registration algorithm based on corner feature are proposed. Based on the characteristics of random noise on different wavelet transform scales and the relationship between the noise Lipschitz and its wavelet modulus maximum, an image denoising algorithm is presented where the wavelet coefficients of noise signal are filtered by changeable threshold so as to reduce the noise; The improved image registration algorithm extracts the angular point by the gradient modulus, then it eliminates the mismatching points and gets all six raw parameters of the affine transformation equation. Some computer denoising and registration simulation results are given. The results show that the denoising method is effective both in removing the noise and in reserving the detail of image. It also shows that the improved image registration algorithm has high registration precision and low computational complexity.(2) This dissertation presents an image fusion algorithm based on local mean gradient modulus. After wavelet decomposition, the fused wavelet coefficients are conbined through the proportion of local mean gradient modulus of coefficients in each low frequency subimage. The improved local deviation fusion rule is used to fuse high frequency subimages, it defines the match measure by the local deviation and finally decides the fusion method for high frequence subimage. Experimental results show that the presented fusion method achieves comparatively high visual effects and performances than the traditional image fusion algorithm.(3) A design proposal of partition image fusion is brought forward in this dissertation. Focusing on the computational complexity of wavelet domain image fusion algorithm and its low real-time capability, this dissertation proposes a new method that fuses the given partition of images which is covered by the virtual detection line and ignores the the rest partition of this image. Comparing with fusion of the whole image, the partition image fusion has low computational complexity. Experimental results show that partition image fusion method applied in the vehicle detection system not only improves the image quality, but also meets the time-constraints of system requirements.更多还原