中文摘要：

可见光和红外图像具有互补特性,融合可产生更好的召回率,但现有方法融合后总会导致精度下降。这项研究提出了一种在特征级进行融合检测行人目标的方法：1提取前景目标特征的极大稳定极值区域Maximally Stable Extremal Regions（MSERs）,计算红外图像MSERs稠密度和相似度特性,并根据此特性分类MSERs。2搜索匹配可见光图像中的相似MSERs区域,定位前景目标。3融合提取红外与可见光图像中的相似匹配MSERs区域,完成运动目标轮廓提取。该方法融合可见光图像信息,能有效滤除背景物,辅助定位在红外图像中检测的前景目标,并补充仅利用红外图像提取前景目标的缺失部分。已使用公共数据库对该方法进行测试,并与早期融合方法进行比较,能获得更好的召回率,同时融合后准确率不会下降。不需要对背景建模,因此比以往算法计算上更高效,更简单,单帧检测的效果也能达到实时处理要求。

英文摘要：

Visual and infrared cameras have complementary characteristics, and the detection performance can be improved by fusing visible and infrared images. The fusion of visible and infrared images can improve recall rate, which means more foreground pixels are correctly detected. But the existing fusion methods always lead accuracy dropped. A method extracting information about the two complementary fused domains to detect pedestrians is proposed： （1）Extract foreground features firstly： extract Maximally Stable Extremal Regions （MSERs） from visible and infrared images, calculate dense degree and similarity characteristics of MSERs in the infrared image, and then classify MSERs based on these features. （2） Locate foreground targets： search all categories of MSERs in the infrared images to match visible images, and extract similar matched MSERs areas in which include pedestrian contour. （3）Merge and output pedestrian contours extracted from similar match MSERs areas. The fusion of infrared and visible information can effectively filter out the background object, locate the foreground objects and replenish the missing part of the foreground object extracted only in infrared images. The method proposed has been tested in a common database and compared with the early fusion methods, which can improve the recall rate, does not drop the precision rate after the fusion, and no need for background modeling. It also can compute more efficiently and easier than previous calculations.