基于多部位多示例学习的人体检测

Human Detection Method Based on Multi-Part Detector and Multi-Instance Learning

基于部位的检测方法能处理多姿态及部分遮挡的人体检测,多示例学习能有效处理图像的多义性,被广泛应用于图像检索与场景理解中.文中提出一种基于多示例学习的多部位人体检测方法.首先,根据人体生理结构将图像分割成若干区域,每个区域包含多个示例,利用AdaBoost多示例学习算法来训练部位检测器.然后利用各部位检测器对训练样本进行测试得到其响应值,从而将训练样本转化为部位响应值组成的特征向量.再用SVM方法对这些向量进行学习,得到最终的部位组合分类器.在INRIA数据集上的实验结果表明该方法能改进单示例学习的检测性能,同时评价3种不同的部位划分及其对检测性能的影响.

Part-based detection methods can deal with large articulated pose variations partial occlusions. Multi-instance learning is employed in content-based image understanding, because it is good at handling the inherent ambiguity of images of human target and retrieval and scene A human detection method based on multi-part and multi-instance learning methods is presented. Firstly, the training samples are partitioned into several regions containing multi-instance according to body structure. Then, the part detectors are trained by using multiple instance learning method based on AdaBoost algorithm. After that the responding scores from the training samples tests are obtained by using the individual part detector when predicting on the positive and negative training bags. Therefore, the training samples are converted to feature vectors composed of part scores. The final assemble detector is learned using a linear SVM method. The experimental results on INRIA database show that the proposed approach improves the detection performance in single instance learning and the influence of the three different multi-part divisions on detection performance is evaluated.