基于部件模型及颜色信息的行人检测

Pedestrian Detection Based on Part Model and Color Information

行人识别是人工智能与模式识别领域内一个新兴的研究方向,具有极其广泛的应用前景。但是由于人体是一个非刚性的运动体,相对普通物体的检测增加了不少难度。可变形部件模型算法对行人检测有着不错的效果,在此基础上提出了一种对传统的部件模型的改进方法,弥补了颜色特征在行人检测时的丢失。其基本思想是:使用传统的DPM方法对待检测窗口进行检测,然后判断检测的得分是否属于可疑区间,如果属于则进一步使用基于颜色特征的分类器对可疑区域进行检测,判断结果由两次的决策值共同决定。在INRIA数据库的检测结果表明,基于多决策的行人检测方法能够在几乎不影响检测速度的同时提髙检测准确率,为精准地对图片或视频中的行人做进一步的分析提供了有利的基础。

Pedestrian recognition is an emerging research in artificial intelligence and pattern recognition, and owns the extremely wide- spread application prospect. However,becanse the human body is a non-rigid body motion, it increases a lot of difficulty compared with ordinary objects detection. Deformable Part Model （DPM） algorithm has a good effect on pedestrian detection. On the basis of that,an improved algorithm for the traditional DPM is presented to makes up for the loss of color features in the pedestrian detection. Its thought is following： using the traditional DPM for detection of window, then judging whether the classification decision value belongs to the sus- picious interval or not. If it does, the classifier based on RGB feature will make the further classification on characteristics, and the results are decided by the two decision values jointly. The experimental results in INRIA database show that the proposed algorithm can raise the detection accuracy without impact on detection speed, and provide the basis for further analysis of pedestrians in pictures or videos.