基于多尺度协同的人头检测方法

Head Detection Method Based on Multi-scale Collaboration

摘要:

针对HOG特征本身不具有尺度不变性,在实际应用中仅能检测出与样本图片大小相差不大的目标对象这一弊端,提出多尺度窗口融合的头部检测的方法;利用线性支持向量机在分类决策方面的优势,与提取的HOG特征结合作分类器的离线训练；在实时的目标检测阶段,采用高斯金字塔式缩放对输入的视频序列作多尺度处理,得到对应的不同分辨率下的待检测帧,在不同的尺度空间作人头的扫描检测并存储结果;之后融合各尺度的检测结果并在相应位置决策标定;实验对某监控视频作检测分析,结果表明,该方法在检出率、召回率、准确度等方面均有较大提升。

Abstract:

The HOG feature was not scale invariance and could only detect the targets which had similar size with sample image in practical application, proposed a method of head detection based on multi-scale windows fusion. The HOG features were extracted and the support vector machinewas used as classifier. In real-time detection of moving targets, the Gaussian Pyramid was used to make multi-scale decomposition for a sequence of input video frames, and got frames of different resolutions, and then run head detection at different scales and storage result. To improve the detection accuracy and efficiency, all detection results of each scale space were fused and got their corresponding locations signs. One monitoring video was tested, and the experiment results showed that the proposed method could improve the detection rate, recall and detection accuracy.