中文摘要：

针对被跟踪目标在发生严重遮挡时采用基于自学习方法的在线Boosting算法易导致错误累积而产生＂漂移＂甚至目标丢失的问题,提出了一种基于子区域分类器的在线Boosting算法。首先,将特征池划分为多个子区域分类器对应的子区域特征池;然后,在跟踪过程中自适应地选取子区域分类器来组成强分类器以剔除被遮挡子区域对目标定位的影响;最后,采用对子区域特征池进行部分更新的方法有效解决了特征在线更新时的错误累积问题。对不同视频序列测试的结果表明,当目标大面积被遮挡时本算法能准确定位目标,目标大小为36pixel×40pixel时的处理帧率为15frame/s。与传统在线Boosting算法相比,本算法对发生严重遮挡的目标仍能进行准确跟踪。

英文摘要：

A new on-line boosting algorithm based on sub-regional classifiers was presented to solve the problem that traditional on-line boosting based tracking algorithm often leads to drifting or failure due to the accumulated error during on-line updating under serious occlusions.Firstly,the feature pool was divided into a number of sub-regional feature pools to correspond to their sub-regional classifiers.Then,the sub-regional classifiers were selected adaptively into a strong classifier to eliminate the influence of occluded sub-regions on the target location when occlusions took place.Finally,the sub-regional feature pools were updated partly to solve the problem of accumulated error during on-line learning.The proposed algorithm was tested with variant video sequences and results show that proposed algorithm achieves exact tracking for the object occluded,and the average computing frame rate is 15 frame/s when the object scale is 36 pixel×40 pixel.In conclusion,the algorithm can satisfy the requirements of stability under occlusion as compared with the original on-line boosting algorithm.