一种基于阴影检测的建筑物变化检测方法

A Method for Shadow Detection and Change Detection of Man-made Objects

提出了一种基于背景模型的针对建筑物的阴影检测及变化检测方法。传统的基于背景模型的目标检测算法认为影像局部区域的自然背景符合高斯正态分布,而含有人工目标的区域则不符合这种分布,从而将目标区与自然地物区区分开来。然而,这种背景模型不适用于中等比例尺的航空影像。本文通过对背景模型的改进,把自然地物和人工地物都视为背景,而把阴影视为检测目标,可以很好地实现建筑物的阴影检测,然后采用阴影补偿法来检测建筑物的变化。试验表明了本方法的有效性。

We present a method for detecting shadows and changes of man-made objects based on background model. Original object detection method based on background model assumes that image pixels of nature background can be modeled as a Gaussian distribution, however regions containing man-made objects do not match with this model. So objects can be detected from the nature background. Our tests proved this background model is not applicable to medium scale aerial photos because in that scale man-made objects take up more percentages in the image and cannot be reqarded as a tail in Gaussian distribution. This paper improves the background model, and treats both nature and man-made objects as background, and shadows as detection objects. Using this model, we firstly detect object＇ s shadows from anomaly of Gaussian distribution and use some filter operators to eliminate those shadows not from man-made objects. Our tests proved that the detection effect of man-made object＇ s shadows is quite impressive. Then the original images were compensated using the shadow images. After compensating, we got images without shadows. At last change detection was made base on those compensating images. For convenience of the experiment, images were partitioned into several areas of interest（AOI）, and coarse matching of counterpart AOIs of old and new images was made. And then we detected man-made objects＇ shadows in AOIs, and those containing man-made objects are remained and those shadows detected will be considered as man-made objects information and be discarded. Then we compensated the AOIs with shadows detection images, and used difference change detection method to compare these AOIs and got the final change detection results. Our tests used aerial images taken at the year 2000 as old images, whose scale is 1： 8000, and aerial images taken at the year 2002 as new images, whose scale is 1： 15000. Experiment results prove our method is effective.