自适应嵌套级联的在线集成学习方法研究

Research on online ensemble learning methods based on adaptive nesting-strutted cascade

针对视频目标检测问题,提出一种新的在线集成学习方法。该方法把目标检测看成两类分类问题,首先用少量已标注样本离线训练一个初始集成分类器,然后在检测目标的同时通过跟踪过滤虚警目标,并通过样本置信度作进一步验证自动标注样本,最后通过在线集成学习方法更新级联分类器。该方法通过在线调整级联分类器,提高分类器对目标环境变化的适应能力,在大量视频序列上进行实验验证,并与现有在线集成学习方法进行比较,结果表明,通过该方法训练得到的检测器不但能够很好地应对目标特征的变化,也能在出现目标遮挡及背景干扰下稳定地检测出目标,具有较好的适应性及鲁棒性。

A new online ensemble learning method is proposed for object detection in video. In this method, object detec-tion is considered as two-class classification problem. Firstly, an off-line primed ensemble classifier should be trained with a few labeled samples, and then the false alarm targets will be filtered by tracking while detecting the objects, at the same time, the automatically labeled samples will be further validated by the sample confidence, finally the cascade classi-fier can be updated by the online ensemble learning algorithm. The adaptability of the proposed method is improved by online adjusting the cascade classifier. Based on the detection results of video sequences, comparing with existing online ensemble learning methods, the detector trained by the proposed approach is adaptive and robust. It can adapt to features changes of the objects, detect objects in partial occlusion or cluttered background.