红外目标跟踪在**和民用视频监控领域有重要的研究意义, 但受热成像原理限制, 红外目标分辨率低、对比度低、纹理信息缺失。针对红外目标特征信息量少导致跟踪性能较低的问题, 提出一种基于自适应响应融合的相关滤波跟踪算法。该算法基于连续卷积运算的相关滤波跟踪框架, 通过构造视觉显著性特征来增强目标外观描述, 并结合对冲决策理论对由不同特征计算得到的多个滤波响应进行自适应融合, 最终根据融合响应预测目标中心位置。此外, 通过尺度滤波器来实现目标的尺度预测, 得到完整的跟踪结果。在公开的红外视频数据集VOT-TIR2016进行测试, 实验结果表明: 与同类算法相比, 该算法表现出更高的跟踪精确度和鲁棒性。

Infrared target tracking is of great significance to the research in military and civil video surveillance. Due to the special thermal-imaging mechanism, infrared targets are often with low resolution, low contrast and in the lack of textures. Aiming at the deterioration of tracking performance caused by insufficient common features of infrared targets, a novel tracking algorithm was proposed based on adaptive fusion of correlation filter responses. The algorithm explored the framework of the correlation filter with continuous convolution operators. The saliency feature was comprised to enhance the object appearance description. The center location of a target was predicted by the fused responses that were calculated from an adaptive fusion of multiple correlation responses with Hedge decision-theoretic. Additionally, the final tracking result was obtained after multi-scale estimation based on scale filters. The experimental results show that the algorithm has better performance in tracking accuracy and robustness compared with other tracking methods on the public infrared video dataset VOT-TIR2016.