动态场景红外图像的压缩感知域高斯混合背景建模

针对动态场景下红外图像的背景模型构建问题，提出一种基于压缩感知（Compressed sensing，CS）域高斯混合模型（Gaussian mixture model，GMM）的背景建模方法.该方法不是对图像中的每个像素建立高斯混合模型，而是对图像局部区域的压缩感知测量值建立高斯混合模型.1）通过提取红外图像轮廓的角点特征，估计相邻帧图像间的相对运动参数以对图像进行校正与配准；2）将每帧图像网格化为适当数目的局部子图，利用序列图像构建每个局部子图的压缩感知域高斯混合背景模型；3）采用子空间学习训练稀疏字典，通过子空间追踪对可能含有目标的局部子图进行选择性稀疏重构；4）通过背景减除实现前景目标检测.以红外图像数据集CDnet2014和VIVID PETS2005进行实验验证，结果表明:该方法能建立有效的动态场景红外图像背景模型，对成像过程中所受到的场景动态变化、背景扰动等具有较强的鲁棒性，其召回率、精确率、F-measure等性能指标及处理速度较之于同类算法具有明显优势.

Background Modeling of Infrared Image in Dynamic Scene With Gaussian Mixture Model in Compressed Sensing Domain

For the problem in background modeling of infrared image in dynamic scene, a new approach to background modeling based on Gaussian mixture model (GMM) in the compressed sensing (CS) domain is presented. The Gaussian mixture model is not for each pixel in the image but for the compression sensing measurement of local regions in the image. Firstly, correction and registration of images are carried out with the motion parameters between adjacent frames estimated by utilizing corner feature of image contour. Then, each frame in the infrared image sequence is meshed into an appropriate number of local sub-images, and the background model of each local sub-image is constructed with Gaussian mixture model in the compressed sensing domain. Furthermore, the local sub-images which may contain target are selectively reconstructed by employing subspace pursuit algorithm with sparse dictionary trained by the subspace learning method. Finally, the foreground targets are detected by background subtraction. Experiments on two datasets of infrared images, CDnet2014 and VIVID PETS2005, are conducted to verify the performance of the proposed algorithm. The results show that the proposed algorithm can establish efficient background model for infrared image in dynamic scene, and has strong robustness to dynamic changes of scene and background disturbance during imaging. The performance evaluations such as recall, precision and F-measure as well as processing speed have obvious advantages over the comparison algorithms.