SIA OpenIR  > 光电信息技术研究室
Alternative TitleMoving Objects Invasion Detection and Tracking
Thesis Advisor惠斌
Keyword目标检测 全局运动补偿 显著性检测 目标跟踪
Call NumberTP391.4/S98/2015
Degree Discipline模式识别与智能系统
Degree Name硕士
Degree Grantor中国科学院沈阳自动化研究所
Place of Conferral沈阳
Abstract运动目标检测与跟踪作为计算机视觉领域的关键技术,已广泛应用于人机交互、交通安全、视频监控、公共安全管理和军事装备等多个领域。本文在对运动目标检测与跟踪算法进行研究的基础上,提出了一些改进方法,并通过实验验证了改进方法的有效性和稳定性。本文主要工作如下:运动目标检测方面,在深入研究现有运动目标检测方法的基础上,提出基于全局运动补偿和图像显著性相结合的目标检测算法,根据图像分割及图像特征点提取结果对场景进行自动判别,然后在不同场景中采用不同的方法检测感兴趣目标。本文采用的基于全局运动补偿的方法是提取相邻三帧图像的SIFT特征点进行匹配;根据Mean Shift算法进行聚类选取背景特征点,估计背景运动参数进行运动补偿;将当前帧图像与运动补偿后得到两幅背景图像分别进行差分、然后相乘,阈值化后得到运动目标区域。基于图像显著性的目标检测方法首先对图像进行预处理,提高目标与背景对比度,利用图像频谱残差得到显著目标候选区域,然后在目标候选区域内进行显著性检测,最后对得到的显著性图进行二值分割提取显著目标。单帧检测结果容易出现虚警、漏检等问题,为提高目标检测结果的准确性,对多帧检测结果进行聚类分析,得到准确的目标模板,对目标跟踪过程进行初始化。运动目标跟踪方面,研究了几种不同的目标跟踪算法,特别是近几年提出的几个较有代表性且具有较好跟踪性能的目标跟踪算法。在研究了多种目标跟踪算法的基础上,对TLD目标跟踪算法进行了深入研究。为实现多目标跟踪,对TLD目标跟踪算法进行改进。通过使每个目标拥有一个正样本集合,多个目标共用一个负样本集合,实现检测模块分类器的训练和分类。本课题依托课题组某型无人机光电吊舱预研项目,为运动目标自动检测与跟踪提供了基础,具有重要的理论和应用价值。
Other AbstractObject detection and tracking, as the key technology in computer vision, is widely used in human-computer interaction, traffic safety, video surveillance, public safety management, military equipment, etc. Improvement methods are proposed in this paper based on research of object detection and tracking algorithms, and are proved to be valid and static by experiments. The main jobs of this paper are as follows.In the object detection part, based on lucubration of object detection methods proposed up to now, we present an object detection method of combining global motion compensation and salience detection, and judging the scenes according to the result of image segment and feature extraction, then detecting interesting objects in different scenes by different means. The global motion compensation method applied in this paper is to extract and match the SIFT feature points in continuous three frames, and select the points locating in the background by Mean Shift clustering method,evaluate the global motion parameters to compensate the global motion. Then difference images is calculated with the present frame and the two background images acquired by motion compensation, and multiplied with each other. Finally, the moving object regions are extracted by threshold segment. The object detection method based on image salience detection states as follows: The original image is pre-processed to improve the image contrast between objects and background. Then, we obtain the candidate salient object regions and compute the image salience map in the candidate regions. Finally, salient object is extracted by threshold segment of the salience map. As detection errors, such as, false alarm, leak detection, etc. is prone to occur in the object detection result of a single frame, clustering analysis of multiple frames detection results is applied to improve the accuracy of object detection. And the exact object model is acquired to initialize the object tracking process.We study different target tracking algorithms in depth, especially several respective methods proposed in recent years which achieved satisfying tracking performance. We study the TLD method on the basis of research of different target tracking methods. And we modify the TLD method to adapt multi-target tracking by one single target holding a positive sample set and all targets share a common negative sample set.The research of this topic relies on a certain kind of UAV electro-optical pod pre-research project, and provide the basis for automatic detecting and tracking moving targets, is estimable in both theory and application.
Contribution Rank1
Document Type学位论文
Recommended Citation
GB/T 7714
孙照蕾. 运动目标侵入侦测与跟踪[D]. 沈阳. 中国科学院沈阳自动化研究所,2015.
Files in This Item:
File Name/Size DocType Version Access License
运动目标侵入侦测与跟踪.pdf(3535KB) 开放获取CC BYApplication Full Text
Related Services
Recommend this item
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[孙照蕾]'s Articles
Baidu academic
Similar articles in Baidu academic
[孙照蕾]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[孙照蕾]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.