基于行走拓扑结构分析的行人检测

Pedestrian detection by walking topology structure analysis

针对计算机视觉应用中的行人检测,当运动目标受到非刚性形变、光照、色彩和遮挡等因素影响时如何建立高鲁棒性的特征描述,本文提出了一种基于语义级行走动作拓扑结构的行人特征,对上述噪声和运动参数不敏感,并且数据量小。算法首先由运动能量图像(MEI,motion energy image)建立行走动态数据,然后抽象为骨架拓扑结构,输入二级级联检测器完成检测。实验表明,该算法可以有效地在光照、形变和遮挡等情况下对行人进行检测。

Pedestrian detection has been a hot topic in computer vision and pattern recognition.To define a robust pedestrian descriptor against non-rigid deformation,illumination,color,clutter and partial occlusion,a novel semantic level feature is proposed based on walking topology structure.The feature is invariable to parameter changes and contains fewer data.Firstly skeleton graph is abstracted from a targets′ walking motion enrgy image（MEI）,then its topology is represented into a tree descriptor and fed into a cascaded detector to accomplish detection.Experiments results reveal that this feature descriptor performs robustly in those complicated scenarios.