视频监控中的预处理、目标检测和跟踪方法研究

发布时间：2018-06-08 21:47

本文选题：视频监控 + 图像预处理　；参考：《南京邮电大学》2014年硕士论文

【摘要】：自物联网的概念被提出以来，人们希望自己的生活往智能化方向发展。视频监控属于智慧安防的一个重要方面，一直是物联网、计算机视觉的研究焦点，有极其重要的研究意义。本文从理论和实际应用出发，重点研究了视频监控中的图像预处理、目标检测和目标跟踪等方面，解决了视频监控中视频质量不高、目标检测率低和目标跟踪处理时间长及不准确等问题。研究内容包括以下几个方面：一、对视频传播中被噪声影响而造成视频图像质量降低的情况，研究了图像预处理来改善视频质量。为了去除图像中普遍存在的两种噪声：高斯噪声和椒盐噪声，本文对能够去除高斯噪声并平滑图像的双边滤波器进行了研究和改进，提出了一种能够同时去除高斯噪声和椒盐噪声的加权双边滤波器，弥补了双边滤波器在去除高频噪声方面的不足。实验结果证明，该算法能够在平滑图像、保留边缘的同时去除椒盐噪声。二、对目标检测算法进行了研究。目标检测分为运动目标检测和静态图像中的目标检测。在视频中进行检测，既可以检测到运动目标，又可以将视频帧作为静态图片来进行检测，因此本文将两者相结合，以提高检测率。在运动目标检测上，本文利用背景差分法，建立背景模型，以检测运动目标。在静态图像的目标检测中，本文利用梯度方向直方图特征加支持向量机分类器的方法进行目标检测。本文提出了一种将两者有效结合、同时进行检测的方法，实验证明，该方法能提高检测率。此外，在目标特征提取的过程中，本文发现，特征提取的好坏对目标检测的结果有着很重大的影响，而特征提取又是人为的、同时需要大量经验技巧的，因此本文希望能利用一种自动提取目标特征的方法来提取较好的特征，，而当前在机器学习领域快速发展的深度学习方法恰好能够满足本文的需求。因此，本文研究了基于深度学习的目标检测方法，虽然该方法目前还不是很成熟，但是实验表明，该方法还是有不错的性能。三、在视频目标跟踪方面，本文对传统的算法进行了总结。然后对实用性很好的跟踪学习检测（Tracking-Learning-Detection，TLD）算法进行了详细的研究，该算法将跟踪和检测算法相结合，解决了长时间跟踪过程中被跟踪目标发生的形变、部分遮挡导致跟踪失败的问题。针对该算法检测过程中会对整张图片进行扫描，本文利用视频帧间的联系，联合Camshift跟踪方法提出了一种改进的检测模块，减小了扫描区域，提高的检测的速度和准确率。
[Abstract]:Since the concept of the Internet of things was put forward, people want their lives to develop intelligently. Video surveillance is an important aspect of intelligent security and has been the focus of the Internet of things and computer vision. Based on the theory and practical application, this paper focuses on the image preprocessing, target detection and target tracking in video surveillance, and solves the problem that video quality is not high in video surveillance. The target detection rate is low and the target tracking processing time is long and inaccurate. The research includes the following aspects: firstly, the image preprocessing is studied to improve the video quality, which is caused by noise in video transmission. In order to remove two kinds of noise, Gao Si noise and salt and pepper noise, the two-sided filter which can remove Gao Si noise and smooth the image is studied and improved in this paper. A weighted two-sided filter which can simultaneously remove Gao Si noise and salt and pepper noise is proposed, which makes up for the deficiency of the two-sided filter in removing high-frequency noise. The experimental results show that the algorithm can smooth the image and preserve the edges while removing the salt and pepper noise. Secondly, the target detection algorithm is studied. Target detection is divided into moving target detection and still image detection. In video detection, moving targets can be detected, and video frames can be detected as static pictures, so this paper combines the two to improve the detection rate. In this paper, the background model is established by using background difference method to detect moving targets. In the object detection of static image, the method of gradient direction histogram feature plus support vector machine classifier is used to detect the target. In this paper, a method of combining the two effectively and detecting simultaneously is proposed. The experimental results show that this method can improve the detection rate. In addition, in the process of target feature extraction, we find that, The quality of feature extraction has a great influence on the result of target detection, and the feature extraction is artificial and requires a lot of experience and skill. Therefore, this paper hopes to use an automatic target feature extraction method to extract better features, while the rapid development of depth learning methods in the field of machine learning can meet the needs of this paper. Therefore, this paper studies the target detection method based on depth learning. Although the method is not very mature, but the experiment shows that the method has good performance. Third, in video target tracking, This paper summarizes the traditional algorithm. Then, the tracking learning detection algorithm Tracking-Learning-DetectionTLDalgorithm is studied in detail. The algorithm combines tracking and detection algorithm to solve the deformation of the tracked target in the long time tracking process. Partial occlusion causes tracking failure. In view of the whole picture will be scanned in the detection process of the algorithm, this paper proposes an improved detection module by using the link between the video frames and the Camshift tracking method, which reduces the scanning area and improves the speed and accuracy of detection.
【学位授予单位】：南京邮电大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP391.41;TN948.6

【相似文献】