基于卷积的稀疏跟踪算法

Convolution-Based Sparse Tracking Algorithm

传统的稀疏表示旨在通过字典的线性结合构建跟踪目标的表观模型,忽视了目标的分层结构特征,因此难以处理复杂的跟踪环境。针对该问题,提出一种新颖的基于卷积的稀疏跟踪算法(CSTA)。在目标区域中提取局部图像块作为局部描述子,依据稀疏表示从中选取一组图像块作为固定卷积核与输入的图像进行卷积运算,能够有效保留跟踪目标的层次化结构特征;同时提出一种新的选择性在线模型更新机制,有效避免错误模型更新导致跟踪结果漂移的问题。所提CSTA在公开数据集中与现有稀疏表示算法进行定量、定性的分析比较,结果表明,CSTA的准确度、鲁棒性均优于现有的稀疏跟踪算法。

Traditional sparse representation algorithms attempt to build a robust appearance model to track targets according to the linear combination of sparse dictionaries.However,such algorithms ignore the hierarchical structure features of the tracking object;thus,handling complex tracking scenery is difficult.In this paper,an innovative convolution-based sparse tracking algorithm(CSTA)is proposed to address this limitation.Local image patches extracted within the object region serve as local descriptors.According to the sparse representation theory,a group of sparse image blocks is selected as the fixed convolution kernel,and the results obtained by convoluting the convolution kernel with the input image demonstrate that the hierarchical structure of tracking objects has been preserved.In addition,a selective online updating mechanism is presented to avoid the drift problem caused by erroneous model updating.Quantitative and qualitative analyses are conducted,and the proposed CSTA and advanced sparse representation algorithms are compared using open datasets.The experimental results demonstrate that the proposed CSTA outperforms state-of-the-art sparse tracking algorithms in terms of accuracy and robustness.