一种时空特征融合的鲁棒视觉跟踪算法

针对视觉跟踪在复杂背景下因外观特征表征不足等原因造成的目标丢失问题，结合深度光流网络估计的运动特征，文中提出了一种基于时序信息和空间信息自适应融合的视觉跟踪算法。该算法在相关滤波跟踪框架基础上，引入递归全对场变换（RAFT）深度网络估计光流以获取目标的时序信息，提取目标的CN特征和HOG特征获取空间信息，然后融合目标时序信息和空间信息，以增强对目标时空特征的表征能力；其次，建立了一种跟踪结果质量判别机制，实时调整时序信息在融合过程中的权重, 有效提升了算法在复杂动态环境下的泛化能力。为评估算法的有效性，在OTB100和VOT2019两个数据集上进行了测试，实验结果表明，与主流视觉跟踪算法相比，所提算法的跟踪性能获得了显著提升，尤其在运动模糊、快速运动等属性的视频中，具有明显优势。

A Robust Visual Tracking Algorithm Based on Temporal and Spatial Feature Fusion

Aimed at the problems that visual tracking is loss caused by insufficient appearance feature representation under complex backgrounds in combination with the motion features estimated by deep optical flow network, a visual tracking algorithm is proposed based on adaptive fusion of temporal information and spatial information. On the basis of correlation filtering tracking framework, the Recurrent All Pairs Field Transforms (RAFT) deep network is utilized for estimating the optical flow to obtain the timing information of the target, and extract the CN feature and HOG feature to obtain spatial information, and fuse the target timing information and spatial information to enhance the characterization ability of the target''s temporal and spatial characteristics, and then a mechanism for discriminating the reliability of tracking results is established, and the weight of time sequence information in the fusion process is adjusted in real time, effectively improving the generalization ability of the algorithm in a complex dynamic environment. In order to evaluate the effectiveness of the algorithm in this paper, the tests are carried out on two data sets, OTB100 and VOT2019 respectively. The experimental results show that compared with the mainstream visual tracking algorithms in recent years, the tracking performance is improved by the algorithm, especially in motion blur, fast motion and other attributes of the video. And this algorithm has obvious advantages.