利用判别字典学习的视觉跟踪方法

针对复杂背景及遮挡等引起目标跟踪性能显著下降的问题,提出一种目标跟踪方法。该方法首先根据目标时空局部相关性获取目标及背景样本。而后建立字典学习模型:基于误差项捕获遮挡等产生的异常值,利用极大极小凹加函数惩罚稀疏编码及误差矩阵,且对字典施加不一致约束项以提高字典的鲁棒性和判别性。针对所构建的非凸字典学习优化问题,利用优化最小化方法对其求解以获得较好的收敛性。最后,由所得判别字典计算候选目标的重构误差以构建目标观测模型,并基于贝叶斯推理框架实现目标精确跟踪。仿真结果表明,与现有主流算法相比,所提方法在复杂环境下可显著地提高目标跟踪的精度及鲁棒性。

Visual tracking method using discriminant dictionary learning

Focusing on the issue of the great decrease in object tracking performance induced by complex background and occlusion, a visual tracking method is proposed. The object and background samples are first obtained according to the local correlation of the object in the temporal-spatial domain. In what follows, a dictionary learning model is established: the outliers generated by occlusion are captured by error terms, and the sparse encoding matrix and error matrix are punished by nonconvex minimax concave plus functions. In addition, inconsistent constraints are imposed on the dictionaries to improve the robustness and discriminability of dictionaries. Concerning the established nonconvex dictionary learning optimization issue, the majorization-minimization (MM) optimization method can be exploited to get better convergence. Finally, the reconstruction errors of the candidate object are computed from the learned discriminative dictionary to construct the object observation model, and after that, the object tracking is realized accurately based on the Bayesian inference framework. As compared to the existing state-of-the-art algorithms, simulation results show that the proposed algorithm can improve the accuracy and robustness of the object tracking significantly in complex environments.