融合深度特征与梯度特征的红外目标跟踪算法

摘要

Based on the complementarity of these two features,this paper presents an infrared target tracking algorithm in combination with the deep feature and the gradient feature.The deep feature and gradient feature are used to represent the semantic information and the local structure information in this paper,enhancing the ability to characterize any target.Next,the tracking models based on different features can further improve the tracking performance.Finally,the paper builds up a model mutual mechanism based on the combination of deep feature tracking model and gradient feature tracking model,implementing positioning precisely to the target.The article selects the latest infrared tracking database and uses the database to verify the effectiveness of the algorithm.The results show that the algorithm in this paper achieves a 3.8％ improvement in accuracy and obtains a 4.3％ improvement in success rate,enabling to effectively handle the effects of the background similarly and deformation.%由于红外序列图像目标与背景的对比度低,图像的边缘模糊并且灰度级动态范围小,采用何种特征描述目标成为跟踪的关键.深度特征和梯度特征是目前大部分跟踪算法采用的主要特征,然而深度特征提取的目标语义信息关注类间分类(Intra-Class),忽略类内差别,容易受到相似背景(Distractor)干扰;梯度特征作为局部区域特征不易受背景干扰,但不能适应目标的剧烈形变.基于这2种特征的互补性,提出一种融合深度特征和梯度特征的红外目标跟踪算法.深度特征与梯度特征被分别用来表征目标的语义信息与局部结构信息,增强了对任意目标的表征能力;利用不同特征建立的跟踪模型进一步提高了跟踪的鲁棒性.通过建立模型互助机制,利用深度特征跟踪模型与梯度特征跟踪模型的互补性,对目标实施了精准的定位.实验中,选取了最新的红外视频跟踪数据库(VOT-TIR2016)用来验证文中算法的有效性,结果表明:和当今主流跟踪算法相比,算法在精确度上获得了3.8％的提升,在成功率上获得了4.3％的提升,能够有效处理跟踪中相似背景与形变的影响.