基于l0正则化的增量低秩特征学习目标跟踪

为了提高生成型目标跟踪算法在遮挡、背景干扰 等复杂条件下的性能,在稀疏编码模型中引入l₀范数正 则化约束,以减少冗余编码信息并改善目标表观重构效果。同时提出一种新的基于非凸近端 加速梯度的快速迭代算法, 解决由此产生的非凸非光滑优化问题。设计了一种增量低秩学习策略,和传统方法需 要将目标观测数据作为 一个整体进行低秩学习不同,本文方法通过l₀正则化稀疏编码能够有效地对目标低秩特 征子空间进行在线学习和更 新。在多个视频序列上的实验表明:基于l₀正则化的增量低秩学习方法能有效提高目标 跟踪算法的准确率和鲁棒性； 和8种优秀的跟踪算法相比,本文算法在中心误差稳健性和重叠率稳健性两个指标上都取得 了最好结果。

Target tracking based on the l0 regularized incremen tal low-rank features learning

In order to improve the performance of generative visual tracking unde r complex environment such as occlusion and background clutter,firstly, l₀ regularized constraint is introduced to sparse coding model to reduce the redun dant encoding information and improve the effect of objective apparent reconstruction.As a no ntrivial byproduct,a novel fast iterative algorithm based on the non-convex accelerated proximal gradient is proposed to solve the resulting non-convex and non-smooth optimization problems.Secondly,a incremental low-rank features le arning strategy is designed.Unlike the traditional methods which need to do low-rank learning on the whole objective o bservation data matrix,the strategy proposed in this paper can effectively learn and update objective low-rank features subs pace online by l₀ regularized sparse coding.Experimental results on multiple video sequences show that the method based on the l₀ regularized incremental low-rank features learning can effectively improve the accuracy and robustness of target tracking algorithm.Compared with the 8state-of-the-art target tracking algorithms,the proposed algorithm achieves the best results both in location error robustness and overlap rate robustness.