【作者】 王丰年；

【导师】 戴国骏；

【作者基本信息】 杭州电子科技大学 ， 计算机应用技术， 2010， 硕士

【摘要】 视频运动目标分析是计算机视觉领域的一个核心问题,在军事、视频监控、等许多方面有着广泛的应用前景。本文主要针对视频运动目标分析应用于智能交通的场景,重点研究了基于注意机制的运动目标检测、单体运动目标跟踪、以及多运动目标之间跟踪焦点转移三个方面的问题,并加以实现和改进。本文按照基于注意力机制的显著区域的提取、显著区域内运动目标跟踪焦点的保持,以及显著区域内运动目标跟踪焦点转移的流程来撰写。首先在显著区域提取部分,提出基于视觉注意机制的目标检测算法。通过引入注意力机制概念,结合图像的动态特征和静态进行目标检测。对传统的基于动态特征的目标检测方式进行改进:结合混合高斯建模以及帧间差分法进行前景提取,并在图像的最佳尺度中进行图像填充,有效解决了检测目标的空洞现象,同时在背景更新中提出局部更新方法,避免局部背景突变而发生误检;基于静态特征的目标检测根据提取的显著区域保持原始物体的一致性特点,从面积和体积上对检测目标进行约束,成功解决因形态学作用而“膨胀“现象。其次在显著区域内跟踪焦点的保持方面,提出对于跟踪区域显著性的保持以及连续跟踪的算法。一旦确定跟踪显著区域,保持该区域的显著性大小,抑制其他显著区域的显著性。在此基础上,保持对该区域内的运动目标连续跟踪。并在跟踪过程中提出了基于粒子滤波的多相似度的条件密度跟踪算法。改进了粒子滤波的重采样过程,利用区域局部性原理,使权值大的粒子周围分布较多的粒子,有效地较少了粒子的贫化现象,然后对图像中的各个方向进行跟踪,将非刚体运动形式中的倾斜、旋转转化成各个方向上的线性运动,降低计算复杂度,并将该算法与目标常用的跟踪算法进行比较,结果显示,该算法有效地提高了跟踪精度。最后在跟踪焦点的切换机制方面,提出通过视觉注意机制的切换机制,进行跟踪焦点转移的算法。检测图像内所有显著区域的显著性大小,抑制中心区域,对其他显著区域的显著性进行排序,根据自底向上的任务驱动,使跟踪焦点总是集中在显著性最大的显著区域内的运动目标中。更多还原

【Abstract】 Video-based analysis of moving objects is a core issue in the field of computer vision and there are broad prospects in the military, video surveillance, and many other areas. In this paper, I select ITS as my research example, and focused on the problems of object detection, single moving object tracking and tracking focus tracking shift, and implement and improve it.This paper is about three topics which are objects detections, single object tracking and tracking focus shift.In the part of extraction in a prominent area, I proposed a algorithm of visual attention-based objection detection. This paper introduces the concept of attention and detects the moving objects which integrated the moving feature and visual feature in the image. For the former this paper improves the detection method as follows: extract the foreground by Gaussian Mixture Model, fill them in the best gauge of the image, and then update the local background which solves the problem of image holes effectively. For the latter, the main job is to control the areas of the detected objects.In the part of the retention of tracking focus, I proposed a algorithm of the definition of the tracking region and tracking focus. The tracking was defined when system is initialed, keep the value of the prominent area, inhibit other significant area and track the moving objects in this place continuously. In the tracking process, this paper proposes a new algorithm based on multi-likelihood condensation-conditional density propagation. This algorithm can reduce the dimension of non-rigid state vector object such as rotation and the degree of tilt by verifying likelihood of the parameters in each particle (the level and vertical of coordinates) each. The experimental data shows that this algorithm plays a good job.In the part of tracking focus shift, I proposed a algorithm of the tracking shift based on the visual attention shift mechanism. All the values of prominent area were computed and sorted by the value. Finally our tracking focus shift mechanism would always concentrate the tracking focus whose value is the biggest.更多还原