基于长方形点过程的遥感图像汽车提取  

Vehicle extraction from remotely sensed images based on rectangle marked point processes

作　　者：余煇 柴登峰[1] YU Hui;CHAI Deng-feng(Institute of Space Information and Technique,Zhejiang University,Hangzhou 310058,China)

机构地区：[1]浙江大学空间信息技术研究所,浙江杭州310058

出　　处：《浙江大学学报（工学版）》2019年第9期1741-1748,共8页Journal of Zhejiang University：Engineering Science

基　　金：国家重点研发计划资助项目(2017YFD0300402-3);国家自然科学基金资助项目(41571335).

摘　　要：针对遥感图像中汽车的识别和提取问题,分析大规模数量汽车的分布和排列,提出基于长方形点过程的汽车分布概率模型.采用先验模型刻画汽车在空间中的分布规律,采用数据项表达模型与图像的联系.采用模拟退火策略结合可逆跳马尔科夫链蒙特卡洛采样方法,对模型进行优化求解,实现图像中汽车的自动识别和提取.实验结果表明,所提方法能很好地用于道路和停车场等规则场景,提取精确度达到99%,回收率达到90%;由于先验模型的约束可以很好地解决汽车相互重叠和方向摇摆的问题,汽车提取的效果比传统算法更好.A rectangle marked point processes was proposed for vehicle identification and extraction from remotely sensed images,and to analyze the distribution and arrangement of a large number of vehicles.A prior model was developed to represent spatial distribution of vehicles,and a data term was applied to establish the relationship between the model and image.The optimal configuration was searched by simulated annealing coupled with Reversible Jump Markov Chain Monte Carlo,so that vehicles could be identified and extracted automatically.The experimental results show that the proposed method works well for the scenes of road and parking lot.The extraction precision is 99%,and the recall is about 90%.The constrain of prior model can solve the problem of vehicle overlapping and uncertain direction well,and the extraction results are better than that of the traditional algorithm.

关 键 词：长方形点过程 汽车提取 模板匹配 模拟退火 可逆跳马尔科夫链蒙特卡洛采样 

分 类 号：TP751.1[自动化与计算机技术—检测技术与自动化装置]