| 决策树弱分类支持的卫星视频运动检测 |
| |
| 引用本文： | 吴佳奇,蒋永华,沈欣,李贝贝,潘申林.决策树弱分类支持的卫星视频运动检测[J].武汉大学学报(信息科学版),2019,44(8):1182-1190. |
| |
| 作者姓名： | 吴佳奇 蒋永华 沈欣 李贝贝 潘申林 |
| |
| 作者单位： | 1.辽宁工程技术大学测绘与地理科学学院, 辽宁 阜新, 123000 |
| |
| 基金项目： | 国家重点研发计划2016YFB0500801国家自然科学基金91538106国家自然科学基金41501503国家自然科学基金41601490国家自然科学基金41501383珠海市引进创新团队项目ZH0111-0405-160001-P-WC广东省“珠江人才计划”本土创新团队项目2017BT01G115 |
| |
| 摘 要： | 针对卫星视频的运动检测存在的局部伪运动虚警问题，提出一种决策树支持的局部有差别更新背景减除法。首先，对ViBe背景模型加以改进，在原模型基础上新增更新因子参数，建立初始的背景模型；其次，利用改进的模型进行运动检测，并对检测结果进行连通域标记，得到分割目标集；然后，根据目标特征属性建立决策树模型，并对分割目标集进行弱分类处理；最后，根据分类结果对背景模型样本和对应的更新因子进行局部有差别更新，进而实现“伪运动”虚警消除。利用SkySat和吉林一号视频数据进行实验，结果表明所提方法的效果和性能良好，检测率优于0.909，针对经典ViBe方法的误检测消除率优于90%。
|
| 关 键 词： | 决策树 卫星视频 运动检测 "伪运动" 有差别更新 |
| 收稿时间： | 2018-09-02 |
Satellite Video Motion Detection Supported by Decision Tree Weak Classification |
| |
| Institution: | 1.School of Geomatics, Liaoning Technical University, Fuxin 123000, China2.School of Remote Sensing and Information Engineering, Wuhan University, Wuhan 430079, China3.Zhuhai Orbita Aerospace Science & Technology Co Ltd, Zhuhai 519080, China4.State Key Laboratory of Information Engineering in Surveying, Mapping and Remote Sensing, Wuhan University, Wuhan 430079, China5.Chang Guang Satellite Technology Co Ltd, Changchun 130000, China |
| |
| Abstract: | For the satellite video motion detection, two issues are harmful to the subsequent application. One is the local parallax"pseudo-motion", and the other is"Ghost pseudo-motion", which may lead to serious error detection. For these two issues, a decision tree-supported local differentiated updating background subtraction method is proposed. Firstly, the initial background model is established. One other parameter, the updating factor, is added based on the ViBe model. Each pixel in the model has a unique dynamic updating factor. In the next place, we perform the pixel-wise motion detection based on the modified model. The detection results are segmented into targets by using the connected component analysis-labeling method. Afterwards, the decision tree model is established based on the target characteristics. And the targets can be weekly classified into different categories. In the end, the background model samples and the factor parameters are differentiated updated according to the categories results. Using SkySat and Jilin No. 1 video data to do experiments. The experimental results show that our proposed method has favourable effect and performance, with the recall rate of results superior to 0.909. Comparing with the classical ViBe method, the error detection removal rate is superior to 90%. |
| |
| Keywords: | |
| 本文献已被 CNKI 等数据库收录！ |
| 点击此处可从《武汉大学学报(信息科学版)》浏览原始摘要信息 |
| 点击此处可从《武汉大学学报(信息科学版)》下载免费的PDF全文 |
|
