基于卷积特征选择的红外目标跟踪

对红外图像中的目标跟踪时，复杂的背景信息以及目标像素数较少等因素增加了红外目标跟踪难度，目标区域的图像块缺乏特征信息使得普通跟踪算法较易产生跟踪偏移问题。为解决此问题，提出了一种基于粒子滤波框架下的卷积特征选择的红外目标跟踪算法。首先，在初始目标块上提取少量图像块作为滤波器，进而获得表征能力更强的卷积特征。然后，采用在线提升算法对该特征进行选择，增加跟踪算法的精度和执行效率。最后，将贝叶斯分类器的响应作为粒子权值估计出目标状态。实验结果验证了所提算法的跟踪性能优于其他几种传统算法。

Infrared target tracking based on selective convolution features

Infrared target tracking is heavily influenced by illumination variation, small size and complex background, and the lack of target information makes the algorithm lose targets easily. Therefore, an algorithm based on convolution features and feature selection method is presented in this paper to track IR targets. First, several filters in target patches of the first frame are used to obtain strong features. Then, the boosting method is utilized to train the features with redundant information, thus, the algorithm performance of accuracy and execution efficiency can be improved. Finally, particle weights are represented by the response of the native Bayes classifier. Experimental results show that the presented algorithm obtains good performance.