对象密集场景的选择式特征背景减除方法研究及应用
文献类型：学位论文
| 作者 | 胡志鹏 |
| 答辩日期 | 2011-05-29 |
| 文献子类 | 硕士 |
| 授予单位 | 中国科学院研究生院 |
| 授予地点 | 北京 |
| 导师 | 黄铁军 |
| 关键词 | 背景减除 背景建模 前景对象检测 特征背景 智能监控 视频分析 视频编码 密集场景 |
| 学位专业 | 其它专业 |
| 英文摘要 | With the “safe city” plan，video surveillance systems are gradually covering all crucial places and the surveillance data generated would increase hundreds of times. The traditional video analysis based on eye monitoring can not satisfy the need brought by such rapid increasing data. Therefore, intelligent video surveillance becomes more and more necessary. One of the most important functions in intelligent surveillance system is separating target objects from background, which is usually utilized through background subtraction. Although background subtraction technology has been researched more than a decade, it remains confronted with many severe challenges, one of which is the foreground detection in crowded scenes. In such scenes, most foreground objects keep still, which leads to inefficiency of the traditional background subtraction algorithms, such as GMM, KDE, etc, because they can only detect moving foreground objects. Eigenbackground method can detect motionless objects, where several eigenbackgrounds are used to reconstruct the background. However, when the scene becomes crowded, some foregrounds may be absorbed into the eigenbackgrounds, leading to severe miss detections and false alarms. In order to solve this problem, selective eigenbackground methods are proposed in this thesis. The main contributions can be summarized as follows: Firstly, this thesis presents a block-level eigenbackground algorithm based on best eigenbackground selection, where the original video frame is divided into blocks and each block is processed independently. Through this blocking strategy, the foreground proportion in the training samples and the spatio-temporal complexity of the algorithm are significantly reduced. Moreover, in order to introduce the foreground information into the background as less as possible, the algorithm selects the best eigenbackground for each block to reconstruct its background, rather than using all the eigenbackgrounds in the traditional eigenbackground method. Experimental results show the performance of the proposed algorithm exceeds those of some states-of-arts on crowded scene dataset. Secondly, to improve the eigenbackground method furthermore, this thesis proposes a pixel-level selective eigenbackground algorithm based on virtual frame. On one hand, clean pixels are selected from the video to construct “virtual frames”, which contain no foreground objects. With virtual frames as training and update samples, the stability and the purity of the eigenbackground model are raised significantly. On the other hand, in the detection stage, the best eigenbackground is selected for each pixel to reconstruct its background value, thus each pixel can get the best background reconstruction result. Experimental results show that the background subtraction performance is further raised with the proposed algorithm. Thirdly, background subtraction technology is applied in video analysis and video encoding, where two intelligent surveillance systems based on background subtraction are developed and evaluated. The first one is the event detection system “eSur”, which includes the subsystem “eSur_Person” oriented to person surveillance and “eSur_Vehicle” oriented to traffic surveillance. Particularly, “eSur_Person” is used in the TRECVID-SED evaluation, where we got outstanding results. The second system, “SmartCam”, is a video encoding and decoding system based on background modeling, which includes the software implementation and partial hardware implementation. In this system, background modeling is applied in video encoding. Experimental results show that the encoding bit rate is greatly reduced under the same video quality. In summary, this thesis studies the background subtraction in crowded scenes and makes a preliminary exploration of its application in intelligent surveillance system. Tentative experimental results demonstrate the feasibility and effectiveness of the work in this thesis. |
| 学科主题 | 计算机应用 |
| 语种 | 中文 |
| 公开日期 | 2011-07-01 |
| 源URL | [http://ictir.ict.ac.cn/handle/311040/1174]  |
| 专题 | 中国科学院计算技术研究所学位论文_2011硕士 |
| 推荐引用方式 GB/T 7714 | 胡志鹏. 对象密集场景的选择式特征背景减除方法研究及应用[D]. 北京. 中国科学院研究生院. 2011. |
入库方式： OAI收割
来源：计算技术研究所
浏览0
下载0
收藏0
其他版本
除非特别说明，本系统中所有内容都受版权保护，并保留所有权利。