基于时空上下文的视觉跟踪自适应超特征融合

论文提出了一种基于上下文感知相关滤波框架的鲁棒跟踪算法。为了提高特征表示的丰富性并充分利用多个不同特征的强度,论文线性加权融合了各种手工制作的特征(如HOG、颜色直方图)和分层深度卷积特征(如VGGNet)。论文将获得的特征定义为超特征,并通过输出约束变换方法优化最终输出响应图,以控制响应图遵循高斯分布,从而获得对目标外观变化的鲁棒性。此外,在模型更新方面,论文提出了一种有效的自适应模型更新方法,在一定程度上缓解了模型噪声。在主流数据集上的大量实验结果表明,所提出的算法在成功率、准确性和鲁棒性方面优于最先进的方法。

Adaptive Hyper-feature Fusion Based on Spatial-temporal Context for Visual Tracking

In this work,a robust tracking algorithm based on context-aware correlation filter framework is proposed.In order to improve the richness of the feature representation,a hyper-feature which contains linearly weighted mixture of hand-crafted fea?tures(such as HOG,color histogram)and hierarchical deep convolutional features(such as VGGNet)is proposed.The final output response map is optimized by the output constraint transformation method to control the response map follow the Gaussian distribu?tion,which gain the robustness to target appearance variations.In addition,in terms of model update,an effectively adaptive model updating method is proposed to suppress the model noises significantly.Extensive experimental results on popular tracking bench?mark datasets demonstrate that the proposed algorithm performs favorably against the state-of-the-art methods especially in compar?ison with several trackers follow deep learning paradigm in terms of success rate,accuracy,and robustness.