基于级联和自适应子分类的目标检测方法

Object Detection Based on Cascade and Sub-categorization

在基于机器学习的目标检测中,检测速度和检测准确性是主要考虑的问题。通过所设计的两级级联和自适应子分类的方法,有效提升了检测速度和检测的准确率。在实际检测时,可以通过一些简单的特征,快速否定绝大多数非检测目标的探测窗口,因此设计了两级级联的方法获得较高的检测速度:在两级级联的第一级采用一种快速简单的检测方法,快速地否定绝大多数错误的探测窗口,并使得几乎所有的正确的探测窗口通过第一级检测。在实际场景下,同一类目标常常有不同的表现形态,如不同姿势、颜色等,因此设计了自适应子分类的方法来获得较高的检测准确率:在两级级联的第二级通过对一类目标使用自动子分类的方法训练多个识别模型,在子分类过程中自动寻找最优分类方法,提升了识别的准确性。在利用著名的UIUC数据集以及一些高清图像的检测实验结果表明,该算法显著提升了检测速度和准确性。

In machine learning based object detection problems, the detection speed and detection accuracy are of the major concerns. In this paper, a new 2-stage detector with self-adaptive sub-categorization is proposed to detect objects rapidly and accurately. At the first stage of our proposed method, most positive detection sub-windows and some negative detection sub-windows are quickly detected by applying a fast detection algorithm. At the second stage, by taking the inner difference like different colors, different directions of one type of objects into account, a strong classifier is build up by using the idea of sub-categorization to achieve a higher recognition rate. Experimental results on UIUC dataset and some high-resolution images have demonstrated that our proposed method can achieve both high detection speed and high detection accuracy in real scenes.