判别式相关滤波(Discriminative Correlation Filters,DCFs)跟踪器会受到目标周围背景噪声的干扰,使得DCF训练容易从上下文环境中来学习滤波器。为了解决上下文干扰难题,本文提出了一种新型的上下文干扰感知相关滤波(Contextual Disturbance-aware Correlation Filter,CDCF)目标跟踪算法,其建立背景上下文干扰模型来指导当前跟踪器的决策过程。首先,CDCF获取最新的上下文背景区域,通过空间裁切操作和相关运算特性产生大量负样本,并将其作为正则项实施到目标方程中。其次,CDCF引入畸变抑制正则项,通过引入历史帧中学习到的滤波器知识,避免形变、遮挡等造成的响应图畸变问题。最后,本文设计了基于交替方向乘子法的优化求解方法,实现裁切操作在空间域的实施和滤波器在频率域的高速求解。在OTB-2013和OTB-2015数据集上的定量评估和定性评估证明了提出算法的有效性。

Discriminative Correlation Filters (DCFs) based trackers are subject to interference from background noises surrounding the target, which makes the DCF training easy to learn the filter from the contextual environment. To resolve this, this paper proposes a novel Contextual Disturbance-aware Correlation Filter (CDCF) target tracking algorithm, which establishes a background contextual disturbance model to guide the decision-making process of the tracker in the current frame. Firstly, CDCF employs the latest contextual background patches to generate a large number of negative samples through spatial cropping and correlation operations, followed by implementing them as a regularization term into the objective function. Secondly, CDCF incorporates an aberration repressed regularization term, which avoids the response map aberration problem caused by deformation, occlusion, etc., via the reuse of the filter knowledge learnt from the historical frames. Finally, an optimization method based on Alternating Direction Method of Multipliers (ADMM) is designed to achieve the implementation of the cropping operation in the spatial domain and the high-speed solution of the filter in the frequency domain. Quantitative and qualitative evaluations on OTB-2013 and OTB-2015 datasets demonstrate the effectiveness of the proposed algorithm.

TP391.4

国家自然科学基金青年项目(62306049)；重庆市自然科学基金面上项目(CSTB2023NSCQ-MSX0665)。