基于在线多核boosting的鲁棒视觉跟踪

Robust visual tracking via online multiple kernel boosting

针对多核学习不满足视觉跟踪外观模型在线更新的要求,提出了一种基于在线多核boosting的鲁棒跟踪算法。采用boosting技术代替传统的全局寻优计算核函数权值,构建基于互补性特征集和核函数集的弱分类器池,将评估分类器判别性的置信度函数作为迭代计算中的目标函数,获取判别能力最好的弱分类器及其权值;引入基于当前帧候选样本信任度分布熵的修正因子,提高在快速变化环境下获取的权值精度;设计了"在线学习"方式代替传统的"批处理学习",通过基于l2范数子空间评估完成外观模型的自适应更新,避免因误差积累导致跟踪偏离。多组具有挑战性的视频序列的跟踪结果表明,本文算法的性能要好于多种现有的优秀跟踪算法。

Multiple kernel learning can＇t meet the requirements of appearance model online updating in visual tracking. To deal with the problem,a robust visual tracking algorithm via online multiple kernel boosting is proposed. We calculate the weights of kernel function with boosting technique instead of clas- sical global optimization. A weak classifier pool is constructed based on complementary features set and kernels set,and confidences of weak classifiers, which reflect the discriminative capacity, are utilized as objective function in iterative computation to select most discriminative classifiers and calculate weights. Then,the accuracy of weights in variable environment is developed via introducing correction factor cal- culated from the confidence distribution entropy of all candidates in the current frame. Classical patch learning in multiple kernel learning is replaced by online learning,and an appearance model can be adap- tively updated based on 12-norm subspace evolution strategy to avoid the drifting problem caused by error accumulation over time. The experiment results on extensive challenging sequences demonstrate that the proposed method has better performance than （MKL） the state-of-the-art trackers.