基于多任务最小软阈值回归方法的目标跟踪

Object Tracking via Multi-task Least-Soft-Threshold Squares Regression

在视频跟踪中,模型表示是直接影响跟踪效率的核心问题之一.在随时间和空间变化的复杂数据中学习目标外观模型表示所需的有效模板,从而适应内在或外在因素所引起的目标状态变化是非常重要的.文中详细描述较为鲁棒的目标外观模型表示策略,并提出一种新的多任务最小软阈值回归跟踪算法(MLST).该算法框架将候选目标的观测模型假设为多任务线性回归问题,利用目标模板和独立同分布的高斯-拉普拉斯重构误差线性表示候选目标不同状态下的外观模型,从而跟踪器能够很好地适应各种复杂场景并准确预测每一时刻的真实目标状态.大量实验证明,文中在线学习策略能够充分挖掘目标在不同时刻的特殊状态信息以提高模型表示精度,使得跟踪器保持最佳的状态,从而在一定程度上提高跟踪性能.实验结果显示,本文算法体现较好的鲁棒性并优于一些目前较先进的跟踪算法.

In visual object tracking, model representation is one of the core issues directly affecting tracking efficiency. The object model representation needs an efficient template learning from the online complicated data to adapt to the variations caused by intrinsic or extrinsic factors during the tracking process. In this paper, the detailed descriptions of robust object model representation algorithm is provided, and a novel multi-task least-soft-threshold squares tracking framework （MLST） is proposed. In the proposed scheme, the observation model is treated as a multi-task linear regression problem, and the appearance models of the candidate target under different states are represented by using object templates and additive independently and identically distributed Gaussian-Laplacian noise assumption, so that the tracker is well adopted to some complex scenes and is able to predict the accurate state of the object in every frame. Extensive experiments are carried out to validate that the online learning scheme improves the representation accuracy by exploring some specific properties of the target in every task, and the tracker maintains the best proposed algorithm ability to achieve favorable performance. The experimental results show that the performs favorably against several state-of-the-art tracking algorithms.