热成像技术能够探测不可见的长波红外辐射并以图像的形式显示，在科学研究、安防刑侦及****中有着举足轻重的地位。如果可以用全景图的方式显示所观测场景的大视场热成像则能够极大地扩大观测者的视野、提升场景感知能力。然而，由于热辐射成像模糊、信噪比低，图像特征提取往往存在着较大误差，进而导致特征点匹配不稳定，图像拼接失败。针对这一问题，改进了匹配过程，提出了一种基于鲁棒特征匹配的热成像全景图像生成算法。在增加特征匹配鲁棒性方面的改进主要包括2 方面：第一，利用PCA（主成分分析）对SIFT 算子进行降维以降低算子相关性，提高特征向量的鉴别能力；第二，利用快速搜索密度峰聚类算法预先筛选匹配点集以剔除错误匹配点，提高特征点的匹配准确度。实验结果表明，本文提出的算法可有效且稳定地生成热成像全景图，具有实用价值。

Thermal imaging equipment can detect invisible long-wave infrared radiation and provide visible display. It has a pivotal position in the areas of scientific research, security and criminal investigation and national defense. Instead of observing single image, it is highly possible to improve the perception of observers via observing a thermal panorama obtained by image stitching technique. However, there are usually large errors in thermal image feature extraction due to the blur details and low signal-noise ratio of thermal imaging. Therefore, the feature points matching process is not stable enough for the image stitching. Aiming at this problem, our paper improves the matching process and provides a robust feature matching based stitching method for thermal panorama. The improvements comprise two aspects: first, using PCA to reduce the dimensions of SIFT features in order to reduce correlations between features, improving the discriminative ability of feature vectors; second, using density peak clustering algorithm to eliminate the unstable matching points in order to improve the matching accuracy. Experimental results show that the proposed algorithm can efficiently and stably generate thermal panoramas with high practical values.