基于自适应特征融合的多尺度相关滤波跟踪

为了减小目标跟踪中目标变形、光照影响、运动模糊以及目标旋转对跟踪效果的影响,在相关滤波KCF基础上,提出了一种基于自适应特征融合的多尺度相关滤波跟踪算法。首先,提取VGG19网络中conv2-2、conv3-4、conv5-4层的特征以及CN特征,并在conv2-2层加入CN特征;然后,将这3个特征分别代替HOG特征进行滤波学习,得到3幅响应图;进而对3幅响应图进行加权融合预测目标位置。最后,在尺度方面引入多尺度相关滤波器进行尺度的确定。该算法比KCF跟踪算法精确度和成功率分别提高了13.6%和11.8%。与现有的其他优异跟踪算法相比,该算法在应对运动模糊、背景杂乱、目标变形、平面旋转方面更具有较好的跟踪效果。

Multi-scale Correlation Filtering Tracking Based on Adaptive Feature Fusion

In order to reduce the influence of target deformation, illumination, motion blur and target rotation on the tracking effect in target tracking, a multi-scale correlation filtering tracking algorithm based on adaptive feature fusion is proposed on the basis of correlation filtering KCF. Firstly, the features of conv2-2, conv3-4 and conv5-4 layers and CN features in vgg19 network are extracted, and CN features are added to conv2-2 layer. Then, the three features are used to replace the hog feature for filtering learning, and three pairs of response diagrams are obtained. Then, the weighted fusion of the three response maps is carried out to predict the target position.Finally, multi-scale correlation filter is introduced to determine the scale.Compared with KCF tracking algorithm, the accuracy and success rate of this algorithm are improved by 13.6% and 11.8% respectively.Compared with several existing excellent tracking algorithms, the algorithm has better tracking effect in dealing with motion blur, background clutter, target deformation and plane rotation.