基于增量子空间自适应决策的目标跟踪

基于增量子空间的目标跟踪算法多数不加选择地将检测到的目标作为模板训练的样本, 并以固定频率更新模板, 这种无反馈闭环机制使得算法在目标外观模型发生变化、 光照变化等复杂条件下难以鲁棒跟踪目标, 一旦跟踪失败很难从错误中恢复. 为此, 我们提出一种反馈闭环跟踪算法, 在增量子空间粒子滤波跟踪框架下, 引入跟踪状态判决作为后续模板更新依据. 通过判决反馈信息选择合适的样本适时更新模板, 有效克服目标外观模型的变化, 持续跟踪目标. 实验结果表明, 由于引入跟踪状态判决, 在目标外观变化、光照变化等情况下, 本算法能够以与环境相适应的频率及时更新模板, 提高跟踪精度, 实验结果验证了本文算法的鲁棒性和有效性.

Robust Object Tracking Based on Adaptive and Incremental Subspace Learning

The traditional target tracking algorithm usually trains the template with detected samples and updates the template at a fixed frequency. This close-loop mechanism lacks feedback and often makes it impossible to track targets robustly when target appearance or illumination changes. Besides, it can not recover from tracking failure easily. Therefore, we propose a feedback-loop tracking framework by bringing in the tracking state judgement. In this framework, the tracking state judgement works as the basis of the following template updating. According to the tracking state judgement, we can choose suitable samples to update the template at appropriate time so as to track targets continuously. Experimental results show that our method can get the current template immediately and correctly due to the tracking state judgement and decision mechanism. We can upate the template at an adaptive frequency and meanwhile track targets correctly even in the case of target appearance or illumination changing.