【作者】 陈鑫；

【导师】 章云；

【作者基本信息】 广东工业大学 ， 控制理论与控制工程， 2014， 硕士

【摘要】 随着数字图像处理、人工智能和模式识别等技术的发展,基于视频的运动目标检测与跟踪技术逐渐应用于智能交通、安全监控和机器视觉等领域,发展前景乐观,近年来研究不断升温,基于深度学习的图像识别技术得到了巨大提升。本文以智能交通系统中城市交通信息的获取为应用背景,以单目摄像机固定场景下的视频图像输出为研究对象,对现有的运动目标检测和跟踪算法展开研究,并对算法给出了一些改进方法,以期提高算法的准确性和实时性。论文分为三大部分,主要内容如下：1.背景建模和运动目标检测。背景建模是运动目标检测的关键,为了能适应户外交通场景下环境光照变化和运动目标阴影的影响,提出了一种基于改进码本模型的运动目标检测方法,利用码本模型对周期性环境光照变化的不敏感性,显著改善了光照小幅变化下的检测效果。针对阴影问题,构建了阴影码本模型,一定程度上抑制了运动目标阴影对检测效果的干扰。2.图像的滤波等前期处理。检测前我们对视频序列图像进行平滑滤波处理,消除噪声干扰。检测后输出的运动目标二值图像存在很多的孤立噪声点及面积较小的连通区域,对这些干扰采用形态学中膨胀、腐蚀处理,消除孤立噪声点的同时填充运动目标内部的空洞。另外通过设定连通区域面积阈值,去除不感兴趣的小目标,比如行人、非机动车等。3.运动目标的预测和跟踪。采用基于卡尔曼预测和特征匹配的方法,首先对处理后的运动目标二值图像,提取目标的质心、外接矩形框等特征,初始化卡尔曼预测模型,预测其在下一帧中将出现的区域。然后在与下一帧中该区域出现的目标进行特征距离的计算,设定匹配阈值,以实现目标间的跟踪。针对跟踪过程出现的遮挡情况,分背景遮挡和运动目标遮挡两种情形进行处理,能较好处理实际中出现的目标遮挡问题。本文针对运动目标检测和跟踪中出现的阴影、遮挡等问题,从算法的实时性和准确性出发,改进了检测算法并提出新的跟踪方法,并对道路交通视频的车辆进行了检测和跟踪测试。更多还原

【Abstract】 With the development of digital image processing, artificial intelligence, pattern recognition and other disciplines, video-based moving target detection and tracking technology is increasingly used in intelligent transportation, security surveillance, machine vision and other fields. The prospect of the research is good, and recent research continues to heat up, the depth learning instead of the traditional method of manual features plus classification has made good progress in the areas of detection and identification of the image. In this paper, the application background is how to get urban traffic information from intelligent transportation system, and the research object is the output video images of monocular camera with fixed scene, we study the existing moving target detection and tracking algorithms and give some improved methods in order to improve the accuracy and real-time performance of the algorithm. The paper is divided into three parts, and the main contents are as follows:1. Background modeling and moving target detection. Background modeling is the key point of moving target detection, in order to adapt to the effects of ambient light changes and shadows of moving target in outdoor traffic scenes, an improved codebook-based methods of moving target detection is presented. With the insensitivity of the changes of periodic environment illumination, we improved the detection result under the case that the light changes slightly. To solve the shadow problem, we construct the shadow codebook model to suppress the interference of moving target shadow.2. Image preprocessing. Before detecting, the video sequences should be smooth-filter processed in order to eliminate noise interference. And after detecting, the binary image of moving objects exist a lot of isolated noises and small connected regions, then we using morphological expansion and corrosion methods to eliminate isolated noises and fill the hollow inside the moving objects. Also, by setting the threshold of the connected areas, we can remove uninterested objects such as pedestrian, bicycle and so on. 3. Prediction and tracking of moving targets. First, we extract the target centroid and external rectangle, and initialize the kalman prediction model to predict the region in the next frame by using the method based on kalman prediction and feature matching. Then, calculating the feature distance between the target and the objects in the predicted region of the next frame, and setting match threshold in order to realize target tracking. To deal with the occlusion in the tracking process, we handle it as background occlusion or occlusion between two targets.To handle the issues of shadows, occlusion and other problems in the process of moving target detection and tracking, an improved method considering the timeliness and accuracy has been proposed in this article, and the method has been tested with traffic video.更多还原