结合注意力机制的核相关滤波目标跟踪

针对目标跟踪中的旋转、快速运动、遮挡等问题 。提出了结合注意力机制的核相关滤波跟踪方法。该方法利用卷积 神经网络提取卷积特征；利用两个样本的相似度矩阵计算注意力权值,并结合注意力权值和 核相关滤波器；使用两个分类 器分别检测目标和背景,并依据两个分类器的响应值实现模型的自适应更新。选取公开数据 集上具有复杂场景的视频序列 进行测试,并与多种跟踪算法在跟踪精确度和成功率上进行定量分析,该算法与原核相关滤 波算法相比,精确度和成功率 分别提高了18.9%、58.7%。实验结果表明,添 加了注意力机制和自适应更新的核相关滤波,较好的解决了遮挡、旋转等 问题,相比其他算法具有更好的鲁棒性和适应性。

Kernelized correlation filter object tracking combined with attention mechanism

Visual tracking remains a challenging problem due to the scale change,deformatio n,occlusion,motion blur and environmental change of the target.A kernelized correlation filter object-tracking approach combined with atte ntion mechanism is proposed in this paper,utilizing attention model and convolution features to improve the robustness of the kernel ized correlation filter algorithm.The convolution neural network is introduced to extract the convolution features.T he attention weights are obtained using the similarity matrix of two samples,and combined with kerneliz ed correlation filters.Two classifiers are used to detect objects and background,and the adaptive update of the model is ach ieved according to the response values of the two classifiers.Compared with the original kernelized correlation filter algorit hm on typical video sequences in public data sets, the success rate and precision have been promoted by 18.9% and 58.7% in this pap er.The experimental results on public data sets show that the proposed approach maintains its good robustness and adaptability even in complex scenes.