Abstract: The compressive tracking algorithm in the tracking process can cause the failure when the target moves fast and has big changes in appearance. In order to solve these problems, a developed compressive tracking algorithm is proposed. Firstly, the sample sets are improved, besides the positive and negative sample sets, this paper also collects one additional candidate sample set, which can effectively reduce the error of gathering samples. Secondly, referencing the MIL algorithms, strong classifier is constituted by the assembled weak classifiers, and it will determine the target through calculating the candidate sample sets, it also proposes new methods which can update the learning rate and the strong classifier. Finally, it uses the probability values that the candidate sample set is positive sample set to adaptively adjust the size of the target window. Experiments show that the improved algorithm has an accurately tracking result for the fast moving and great changed targets, and it also has better robustness than the other algorithms.

Key words: compressive sensing, compressive tracking, Multiple Instance Learning（MIL）, samples set, classifiers

摘要： 针对压缩跟踪算法在跟踪目标过程中，目标快速运动和目标外观尺寸变化较大时引起的跟踪失败问题，提出了改进的压缩跟踪算法。首先，对正负样本的采集做了改进，除了采集正负样本集，还额外采集一个候选样本集，从而减少样本采集带来的误差。其次，借鉴MIL算法思想，利用弱分类器的组合构成强分类器，然后对候选样本进行计算来确定目标，并提出了模型学习率和强分类器更新方法。最后，根据候选样本集为正样本集的概率值来自适应调节跟踪目标窗口的大小。实验表明，改进后的算法能对快速运动和外观变化较大的目标进行准确的跟踪，改进算法比原算法具有更好的鲁棒性。

关键词: 压缩传感, 压缩跟踪, 多示例学习, 样本集, 分类器