复杂交通监控场景下运动目标检测与跟踪方法研究

发布时间：2018-03-01 02:25

本文关键词： 码本混合高斯模型梯度统计直方图稀疏表示块正交匹配马尔可夫随机场　出处：《长安大学》2014年博士论文　论文类型：学位论文

【摘要】：交通监控场景中情况各异、环境复杂，运动目标的任意性和随机性，以及光照、遮挡、目标姿态等不确定性，使运动目标检测与跟踪过程中出现的问题具有不可预测性。本文在复杂交通监控场景中应用单目摄像机获取目标区域，建立借助目标特征实现目标检测与跟踪方案。围绕形变与尺度变化下目标的检测与跟踪方法，研究了视频中目标区域的检测和提取、基于目标特征的识别与分类，以及复杂场景中尺度变化的目标的跟踪问题，这些问题形成了基于复杂交通监控场景下运动目标检测与跟踪方法的技术研究。论文的主要内容如下： 1)提出一种基于量子聚类分析技术的像素块编码的交通背景提取方法。实际交通监控中的场景状况是不确定的，为了精确的提取前景需要用有效的方法对背景建立模型，了解视频序列中像素变化并考虑到像素之间的相互关系。本文在经典Codebook方法的基础上进行探索和研究，将视频图像划分为像素块，对像素块进行聚类学习和编码，在该编码的基础上用交替学习和更新的方法对编码进行实时更新。实验证明，本方法获取的前景干扰较少，目标区域图像较清晰；另外该方法计算简单，加之采用交替更新的方法，实时性好，提取的前景具有较好的鲁棒性。 2)提出一种基于MRF的自适应车辆阴影检测和消除方法。对车辆阴影通过将前景与背景像素的颜色和局部纹理特征进行对比。颜色特征是用HSI颜色空间特性，局部纹理则用SILTP编码的汉明距离对阴影像素检测，其中得阈值用极大似然估计的方法估计。在以上特征的基础上，用马尔可夫随机场对像素标记及其邻域的相关性进行表示，进而对阴影和非阴影像素进行分割。实验结果表明，和其他方法相比该方法有相似或者更优越的性能，能适应光照的变化环境。 3)提出基于Gabor特征图像上提取HOG特征的行人识别方法(在此简称GHOG方法)。针对场景中目标姿态、光线等不断变化的需求，该方法将视频图像和Gabor小波进行卷积，得到的40个小波图像进行尺度和方向上进行融合，形成一幅Gabor的融合图像；在此基础上进行HOG特征提取，根据提取的HOG特征用Real Adaboost级联分类的方法进行目标识别。实验证明，该方法能有效降低错误检测率；对目标在Gabor特征图像融合过程中采取了编码的方式，使计算量也有效降低。 4）提出一种对于目标图像分块稀疏表示和贝叶斯估计进行目标跟踪的方法。针对目标跟踪过程中遮挡问题，该方法根据基础样本库子空间的块对目标的外观进行稀疏线性组合表示，为了实时更新目标模板，采用了增量学习的方法来适应不断变化的目标。然后建立了基于重建图像和观察目标的近似误差的概率观察模型，这个观察模型用一个随机的仿射运动模型形成粒子滤波进行目标跟踪。本文提出的跟踪方法比IVT和L1T方法在处理遮挡、姿态变化较大、突然光线变化和尺度变化方面有较好的效果。
[Abstract]:In the traffic surveillance scene is different, the environment is complex, arbitrary and random moving target, and illumination, occlusion, object pose uncertainty, so that the moving target detection and tracking problems in the process of unpredictability. Get the target area using monocular camera based on complex traffic monitoring scene, establish with the help of the target characteristics to achieve target detection and tracking scheme. Based on the method of detecting and tracking target deformation and scale change, studies the detection and extraction of the target area in the video, recognition and classification based on the features of the target tracking problem, and target scale in complex scene changes, these problems form the technology of detection and tracking method moving target in complex traffic monitoring based on the scene.
The main contents of the paper are as follows:
1) extraction method is proposed for analysis of quantum clustering technology based on pixel block encoding traffic background. The actual traffic monitoring scene in the situation is uncertain, in order to accurately extract the foreground with the effective method for background modeling, pixels in video sequence changes and considering the relationship between pixels in the article. The exploration and Research Based on the classical Codebook method, the video image is divided into blocks of pixels, clustering learning and encoding of pixel blocks, updated the encoding method by alternative learning and updating based on the encoding. Experiments show that this method gets the prospect of less interference, the target region image is clear; the method is simple, and the method of renewal, good real-time performance, the extraction of foreground is robust.
2) proposed an adaptive vehicle shadow detection and elimination method based on MRF. The vehicle shadow through the foreground and background pixel color and local texture features were compared. The color feature is characteristic of HSI color space, local texture with SILTP encoding the Hamming distance of shadow pixel detection, the threshold method for maximum likelihood estimated. Based on these features, represented by a Markov random field on correlation between pixel labeling and its neighbor, and the shadow and non shadow pixel segmentation. The experimental results show that compared with other methods this method has similar or better performance, can adapt to the changes of illumination.
3) proposed pedestrian recognition method of HOG feature extraction based on Gabor image features (here referred to as GHOG). According to the attitude of the target in the scene, the light and the changing needs of the video image and Gabor wavelet convolution, 40 wavelet image obtained by the scale and direction of fusion, forming a fusion image Gabor; based on HOG feature extraction, HOG feature extraction method based on the Real Adaboost cascade classifier for target recognition. The experimental results show that this method can effectively reduce the error detection rate; the target in the Gabor feature in the process of image fusion by encoding, the calculation is reduced effectively.
4) proposed a target image for block sparse representation and Bias estimation for target tracking. The occlusion problem in the process of target tracking, the method based on sample subspace block to target the appearance of sparse linear combination, in order to update the target template, using incremental learning method to adapt to the the change of target. Then established observation model of reconstructed image and the observed object approximation error probability based on the observation model with a random affine motion model into particle filter for target tracking. The tracking method proposed in this paper than the IVT and L1T methods in dealing with occlusion, attitude changes greatly, has good effect of a sudden change of light and scale changes.

【学位授予单位】：长安大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：TP391.41;U495

【参考文献】