基于多特征级联筛查的在线boosting快速跟踪算法

A FAST ON-LINE BOOSTING TRACKING ALGORITHM BASED ON MULTI-FEATURE CASCADE SCREENING DETECTION

传统基于Haar-like特征的在线boosting跟踪算法(HBT)采用局部穷举搜索目标的方式,不能很好地应对运动速度较快的目标以及目标被完全遮挡的情形。当目标状态和周围背景发生变化时,传统HBT算法会产生累积错误。对此系统进行改进,提出一种基于多特征级联筛查的在线boosting快速跟踪算法:将每帧视频网格化,依次根据目标运动方式、网格方差、目标模型、颜色分布以及重叠情况等多种特征级联筛选出有可能成为目标的网格。将这些候选网格交给boosting分类器得到最终的置信度,从而得到目标位置信息,实现快速的在线目标跟踪。用朴素贝叶斯分类器代替简单的阈值分类器,提高算法的准确性。实验结果表明,所提出的方法在鲁棒性、准确性和实时性上都有很大提升。

Conventional on-line boosting tracking algorithm （HBT） based on Haar-like feature adopts the objects local exhaustive search pattern, which cannot perfectly respond to the object with faster motion speed and the situation that the object is fully occluded. In addition, conventional HBT algorithm will generate cumulative errors when the state of the object and the environment background changes. In this paper, we make the improvement on the system, and propose a fast on-line boosting tracking algorithm which is based on multi-feature cascade screening and detection. The algorithm makes every frame of the video to gridding form, according to multi-feature in turn including the manner of object motion, grid variance, object model, colour distribution and overlap situation, it screens and detects in cascade those grids possibly to be the object and passes these candidate grids onto boosting classifier to acquire final confidence degree, and therefore obtains the object location information, achieves fast on-line object tracking. To replace the original simple threshold classifier with naive Bayes classifier improves the accuracy of the algorithm. Experimental results demonstrate that the robustness, accuracy and real-time property of the proposed algorithm are all greatly improved.