【作者】 何强；

【导师】 杨波；

【作者基本信息】 重庆大学 ， 模式识别与智能系统， 2007， 硕士

【摘要】 图像序列中的运动目标检测与跟踪是数字图象处理和模式识别以及计算机视觉领域研究的主要内容之一。在机器人导航、安全监控系统、医学图象分析和虚拟现实等领域中都有广泛应用。对于运动目标跟踪系统,其核心主要包括两个部分:一是运动目标的检测与提取;二是运动目标的跟踪。运动目标的提取是目标跟踪任务最为关键的部分。如果提取的目标不正确或不完整,将会极大的影响目标跟踪部分的任务,甚至不能进行有效跟踪。在运动目标检测方面,在主要包括连续帧间差分法、背景差分法、和光流法,并对各种方法进行了比较和实验分析,指出其优缺点和适用范围。图像分割是目标提取中的关键一环,图像分割质量的好坏将直接影响到目标的提取。本文提出了二维OTSU-GA自适应阈值分割算法,它综合了二维OTSU阈值分割算法和遗传算法的优点,经实验证明:该算法既保证了图像分割质量,又保证了分割效率。在目标跟踪研究方面,本文深入讨论了均值偏移算法背后深厚的非参数估计理论,详细地进行了算法的推导,讨论其收敛性。采用了以均值偏移算法为基础的CamShift(Continuous Adaptive Meanshift)跟踪算法。CamShift跟踪算法是一种基于目标颜色特征的跟踪方法,该方法由于其无参数,运算速度快而被广泛应用。在简单的背景环境中Camshift算法能够取得较好的跟踪效果,但在复杂的背景中,由于该算法对物体不做任何预测,当有大面积背景颜色与目标颜色接近时,将无法跟踪到目标。于是本文提出CamShift和卡尔曼滤波器相结合的算法,卡尔曼滤波器预测目标在当前帧的可能位置,CamShift算法在该位置邻域内搜索,算法对复杂背景下的运动目标有很好的跟踪效果,而且对遮挡问题也有很好的处理。经实验证明,此算法有较强的鲁棒性和实用性。更多还原

【Abstract】 Detection and tracking of moving object in image sequence is an important area in the domain of digital image processing, pattern recognition and computer vision. Its applications include that robot navigation, security surveillance, medical image analysis and virtual reality.For a visual tracking system, the core includes two parts: one part is moving object detection and extraction, and the other is moving object tracking. Moving object extraction is the most important part. With incorrectly extracted or half-baked object, it may seriously affect the tracking operation, even can not track the object. For moving object detection, three methods which included“differential between adjacent frames”,“background subtraction”and“optical flow”, are discussed. Through comparison and analysis, the merits, weaknesses and application field of each method are concluded. Image segmentation is an important step of object extraction. The object extraction is directly affected by the quality of image segmentation. In this paper, a“Two-dimensional OTSU-GA Adaptive Thresholding Segmentation Algorithm”is proposed. This Algorithm make Genetic Algorithm combine with Two-dimensional OTSU threshold Algorithm. It can make calculation faster by GA’s capacity of searching the best answer. This algorithm improves the efficient of Image Segmentation.In this paper, nonparametric density estimation which is base of mean shift iteration is formulized thoroughly. The derivation of mean shift procedure as well as the strict proof of its convergence is given. On the object tracking research, the CamShift algorithm which means Continuous Adaptive Mean-Shift is mainly discussed. While Camshift is a tracking method based on the object’s color character, and it is widely used for the parameter-less and fast calculation. CamShift algorithm works well in the simple background, but not in the complicated background. So the CamShift algorithm combined with Kalman filter is proposed in my paper. Kalman filter is used to forecast possible position of target, then Camshift search the real position near the possible position. The algorithm has good effect to moving target in the complicated background, and can deal well with simple occlution. The results proved that the algorithm is robust and practical.更多还原