摘要:目的 当前,目标跟踪问题常常会通过在线学习、检测的方法来解决。针对在线学习过程中,分类器训练需要花费大量时间以提高其识别准确率的问题,提出使用Adaboost算法级联弱分类器,在训练一定帧数后仅进行检测的方法来达到实时和准确的折中。方法 首先针对跟踪问题简化了haar特征,以降低特征计算量。同时考虑到经典的Adaboost算法可能并不适合跟踪过程中存在的正负样本不均衡问题,提出在样本权重更新公式中引入一个新的调整因子项并且结合代价敏感学习来提高目标识别率的方法。最终给出使用简化的haar特征作为描述子,改进的代价敏感Adaboost作为分类器的目标跟踪算法。结果 对20组视频进行跟踪实验,本文算法的平均代表准确率高于压缩跟踪算法约26%,高于原始代价敏感算法约11%;本文算法的视频处理平均帧率高于压缩跟踪算法约38%。结论 本文提出的新代价敏感Adaboost算法对目标的识别、跟踪具有较高的准确率及较快的处理速度,并具有一定的抗干扰能力。特别对人等非刚性目标能够进行较好跟踪。

关键词:目标跟踪  代价敏感Adaboost算法  简化haar特征  调整权重因子项  在线学习  非刚性目标

AbstractObjective Visual tracking is one of the most active computer vision research topics because of its wide range of applications. Currently, target tracking problems are often solved through online learning and detection methods. A tracking task can be considered a binary classification problem solved using online learning method. However, in the process of online learning, the classifier training takes a considerable amount of time to improve its recognition accuracy. In this study, a method using the Adaboost algorithm is proposed to solve this problem. The algorithm initially trains weak classifiers in a certain number of beginning frames and will subsequently perform only as a detector without training to address the issues related to real time and accuracy. Method The Haar feature needs to be simplified because its computational cost remains a burden for real-time tracking. Thus, we remove the Haar orientation to facilitate calculation. Positive samples, i.e., samples containing the target, are always the minority in tracking; as a result, the training samples are imbalanced. Accordingly, the algorithm needs to focus more on the positive targets to achieve higher detection rate. The equal treatment of false positives and false negatives by Adaboost may no longer be appropriate. In this case, we choose a cost-sensitive Adaboost to achieve higher detection rate for the positives. Furthermore, given that misclassified samples appear more often during a scenario because of the complex environment in visual tracking, we add a new parameter in the sample weight-updating formula of the cost-sensitive Adaboost to provide more weight to the misclassified samples, which consequently will be given more focus by the classifier. Finally, we propose a tracking method based on the simplified Haar feature as descriptor and the improved cost-sensitive Adaboost as classifier with online learning strategy. Result In our experiments, we compared our method with two state-of-the-art algorithms and the original cost-sensitive method in both accuracy and processing speed. We tested the different methods on 20 benchmark video sequences. In terms of accuracy, the average representative precision of our method is approximately 26% higher than that of the compressive tracking method and approximately 11% higher than that of the original cost-sensitive method. In terms of processing speed, the average frame rate of our method is approximately 38% faster than that of the compressive tracking method. Conclusion Our method is based on a modified cost-sensitive Adaboost that focuses more on the minority positive samples to improve detection rate. The proposed method performs well in terms of accuracy and speed, especially for non-rigid objects, such as human bodies.

Key wordsobject tracking  cost-sensitive Adaboost  simplified haar feature  a parameter for weight updating  online learning  non-rigid objects

文章编号： 0258_7106 (2016) 01_0018_15 中图分类号： P618.41 文献标志码：A

投稿时间：2015-11-13修订日期：2016-01-04 改回日期：2015_07_11

基金项目

**通讯作者耿新霞， 女， 1979年生， 助理研究员， 成矿规律研究方向。 Email： gen gxinxia@cags.ac.cn