基于1-norm SVM权值学习的多示例目标跟踪

针对复杂场景下目标跟踪存在鲁棒性低，容易发生跟踪漂移的问题，提出一种改进的多示例目标跟踪算法。该算法针对多示例跟踪算法在包概率计算过程中忽略样本间的差异，对所有样本赋予相同权值，造成分类器性能下降及弱分类器选择存在复杂度高的问题，通过1-norm SVM计算各样本对包概率的重要程度，并在弱分类器选择过程采用内积的方法计算包概率的似然函数，从而减小算法的复杂度和计算时间。实验结果表明，该算法在目标发生遮挡、姿势变化、场景光照发生较大变化以及出现相似目标等较强干扰的情况下仍能较好地跟踪目标，具有较强的鲁棒性和抗干扰能力。

Multiple instance object tracking algorithm based on 1-norm SVM weight distribution

For the poor robustness and target drift problem of the most existing tracking algorithms in complex environment, an improved target tracking algorithm based on multiple instance learning is proposed. The MIL tracker ignores the differences of each sample in the process of computing the bag probability, which declines the performance of classifier, and there exists complex problem in choosing the weak classifier. This paper solves these problems by computing the importance of each sample to bag probability based on the 1-norm SVM method. Then, it adopts inner product method to compute the log-likelihood of bag in the process of choose weak classifier, which is benefit to reduce the computing complexity. Experimental results show that the proposed algorithm performs well with strong robustness and high tracking accuracy under the complicated environments such as occlusion, rotation, pose and illumination change.