红外图像信噪比和对比度较低、缺乏颜色纹理信息、目标周围有光晕效应、边缘模糊，这些缺点对红外图像中人体目标检测提出了挑战。本文对复杂环境下红外图像序列中运动人体目标检测技术进行研究。首先采用基于改进的混合高斯模型（Gaussian mixture model，GMM）的背景减除法对人体目标进行分割，通过多个带有权值的高斯过程来描述复杂变化的背景，对模型个数、权值、学习率进行更新。然后对分割得到感兴趣区域（Region of interest，ROI）采用融合边缘方向累加和特性的梯度方向直方图（Accumulation of oriented edge and histogram of oriented gradient，AOE-HOG）进行特征描述，利用支持向量机（Support vector machine，SVM）实现对人体目标分类检测。实验表明，本文算法能够在复杂场景下正确检测出人体目标，对于多目标距离较近甚至有部分粘连的情形，也具有较好效果。

Infrared images have disadvantages such as low signal-to-noise ratio and contrast, a lack of color texture information, and a halo effect around target and blurry edges. These factors pose challenges for detecting humans in infrared images. This study focuses on human detection technology used for infrared image sequences in complicated environments. Specifically, we use a background subtraction method to segment a human-body target based on a modified Gaussian mixture model. First, we use multiple Gaussian processes to simulate the complex changes in the background with the appropriate weight values. These processes also update the number, weight values, and learning rate of the Gaussian model. We then use the fusion of the accumulated oriented edges and a histogram of oriented gradient characteristics to describe the region of interest. Finally, we employ a support vector machine to classify the human targets precisely. Experiments show that the algorithm can detect human targets accurately in complex backgrounds and that it generates good results on multiple objects, those near in distance, and even some of having adhesion multiple objects, near distance, and even some of the adhesion.