【作者】 付朝霞；

【导师】 王黎明；

【作者基本信息】 中北大学 ， 信号与信息处理， 2013， 博士

【摘要】 基于视频流的目标检测、跟踪与识别是计算机视觉和模式识别领域的热点问题,在智能视频监控、高级人机交互、移动机器人定位与导航、虚拟现实等领域具有广泛的应用前景。经过几十年学者们的不懈努力,上述技术已经有了较多的研究成果。由于视觉应用系统中环境的复杂性以及目标本身的多样性,给目标检测、跟踪和识别技术带来了极大的困难。实践表明一般意义上的目标检测、跟踪与识别技术还远未成熟,距离实用化尚存在一定差距,还需要开发出更为实用鲁棒的算法。本论文从理论和实际应用的角度出发,对以视频为输入的运动目标识别的相关关键技术进行研究,研究内容主要涉及运动目标的检测、运动目标的跟踪、运动特征的表征和识别方法等。本文研究了背景建模方法,提出了一种基于像素统计分类的视频流目标检测算法,借助把图像的像素值看成是前景高斯分布和背景高斯分布的组合,进行背景估计和自适应背景更新；以统计当前帧前景像素的点数来判定光照突变,并结合帧间差分法来检测运动目标。仿真实验表明,该算法可以实时准确地检测出前景运动目标,具有更强的适应性。通过复杂背景下的人脸检测实验表明,该算法在基于肤色信息的人脸检测中也具有一定的实际应用价值。本文还提出了一种基于链码标定的圆检测算法,利用数学形态学方法有效地去噪填充和提取二值图像的边缘,再利用链码方法确定圆度参数。实验表明,该算法简单有效,计算精度小于1个像素,具有较好的实用效果。针对多目标跟踪问题,本文提出了融合角点特征的多目标跟踪算法。利用改进的Harris算子提取运动目标的均匀稳定的特征点,通过特征匹配和匹配优化,完成视频运动多目标的跟踪。跟踪实验表明,该算法能够完成视角变化、旋转、仿射变换、光照变化等多种情况下的稳定匹配,可以实现小部分遮挡状态下目标的稳定跟踪。本文研究了经典的Mean shift跟踪算法,由于该算法对于快速运动的目标跟踪是无效的,而且还存在误差累积的问题,因此本文提出了基于质心加权的Kalman滤波的跟踪算法。利用背景差锁定动态目标跟踪区域,在目标跟踪开始时利用Kalman滤波来预测目标的位置,然后采用质心加权算法优化修正跟踪目标的位置,并以修正后的状态预测值进行观测更新,进而实现对跟踪目标较为精确的定位。经过仿真实验分析,该算法在有效检测到运动物体的同时能够快速准确地跟踪运动物体,具有较好的实时性与较强的鲁棒性。针对复杂多变光照下的人脸识别问题,本文提出了基于LBP算子与EMD的人脸识别算法,首先对图像进行一系列简单有效的预处理以提高算法的鲁棒性,然后提取图像的局部LBP特征,获得图像的LBP直方图。采用EMD方法对LBP直方图进行计算,完成对图像相似性的度量。在GTAV标准人脸库上实验结果表明,该算法显著提高了识别率。人体行为识别与理解属于更高一层的视觉任务。本文在探讨了各种人体行为识别算法的基础上,提出了一种基于时空兴趣点的人体行为识别算法,采用3D Harris角点提取不同行为的时空特征,然后采用K-means聚类和LLE结合的方法对提取的运动特征进行降维和分类,训练识别过程则采用平均Hausdorff距离的几何特征方法完成相似性配准。KTH数据库上的实验表明该算法是有效可行的,基于流的轨迹识别方法进一步提高了识别的准确率。更多还原

【Abstract】 Object detection, tracking and recognition based on video stream are hot issues in the field of computer vision, pattern recognition, intelligent video surveillance, senior human-computer interaction, mobile robot localization and navigation, virtual reality, which have broad application prospects.In the decades’unremitting efforts of scholars, these technologies have already more research achievements. Due to the complexity of the environment and the diversity of the goal itself in the visual system, the technologies of object detection, tracking and recognition have brought great difficulties. The practical experience shows that the technologies of object detection, tracking and recognition are far from mature in the general sense and there are still certain gaps away from practical application. So they also need to develop more practical and robust algorithms. In both the theoretical and the practical perspective, this paper studies on some correlative key issues of moving object recognition with the input video. The issues mainly focus on moving object detection, moving object tracking, motion feature representation and recognition.In the paper, the methods of background modeling are studied and an algorithm of object detection of video stream is proposed based on the pixels’ statistics classification. The pixel values of the image are seen as the combination of the foreground Gaussian distribution and the background Gaussian distribution, and the background estimation and the adaptive background update will be put up. The statistical number of the foreground pixels of the current frame determines whether the light has a larger change, and the algorithm needs to combine with the frame-difference method to detect moving object. Simulation results show that the algorithm can quickly and accurately detect the foreground object with greater adaptability. The experiment of face detection under complex background shows that the algorithm has a certain practical worth in the face detection based on skin color information. Circle detection algorithm is proposed in this paper. It firstly uses mathematical morphology method to denoising, filling and contour extraction for the binary image, and then calculates circularity index using chain code method. Experimental results show that the circle detection algorithm is simple and effective and the accuracy is less than one pixel.To solve multi-target tracking problems, this paper proposes a multi-target tracking algorithm based on a combination of corner feature. It extracts stable and symmetrical feature points of moving object using the improved Harris operator, and completes the tracking of video moving multi-target by feature matching and matching optimization. Tracking experiments show that the algorithm can complete stable matching under the changes of angle view, rotation, affine transformation, illumination and other circumstances, and can achieve stable tracking under a small partial shelter state. The classic tracking algorithm of Mean shift is not valid to fast moving object, and has also the problem of error accumulation. So this paper proposes an algorithm based on centroid weighted Kalman filter for object tracking. The algorithm firstly uses background subtraction method to lock dynamic target tracking area, and then uses the Kalman filter to predict the target’s position at the beginning of the target tracking, and then optimizes the predictive state value adopting centroid weighted method, finally updates the observation data according to the corrected state value. Simulation results show that the algorithm can detect effectively moving objects and at the same time it can quickly and accurately track moving objects with good robustness.To solve face recognition problems in a complex and changing light, the paper proposes an algorithm of face recognition based on LBP operator and EMD. Firstly, it uses a series of simple and efficient image preprocessing for improving the robustness of the algorithm, and then the LBP histogram of the image is obtained by extracting the local LBP feature. The use of EMD can complete measuring the similarity of the images by calculating the LBP histogram. The experimental results in the GTAV standard face database show that the algorithm improves significantly the recognition rate. Recognition and understanding of human behavior are the higher level of visual tasks. On the basis of various algorithms of human behavior recognition, the paper proposes an algorithm of human behavior recognition based on space-time interest point. It firstly uses3D Harris corner to extract the spatial-temporal features of different behavior, and then classifies these motion features and reduces their dimensions using K-means clustering combining with LLE method in the data space. In the process of training recognition, the geometric characteristics method of the mean Hausdorff distance completes similarity registration between image sequences. The experiments on the KTH database show that the algorithm is effective and feasible, and the stream-based trajectory method improves further the recognition accuracy.更多还原