结合梯度直方图贝叶斯理论视频行人检测

Pedestrian detection for surveillance video based on gradient histogram Bayesian theory

面向监控视频的行人检测应用需求,提出一种结合贝叶斯理论的行人检测方法。采用Vibe算法提取前景目标区域,得到二值掩膜图像,降低背景区域引起的虚警现象和时间耗费;在前景区域提取方向梯度直方图特征,采用支持向量机分类器检测行人目标,得到行人目标矩形窗口集;基于二值掩膜图像和行人目标矩形窗口集,计算各像素点属于行人目标的先验概率和似然,基于贝叶斯理论估计后验概率,生成概率图像;采用OTSU方法自适应分割概率图像,得到每一帧图像最终的行人目标检测结果。在Caltech数据集的行人检测结果表明,该方法的真正率和检测帧率高,假正率极低。

A pedestrian detection method based on Bayesian theory was proposed for facing pedestrian detection application need of surveillance video.The Vibe algorithm was used to extract foreground object regions,and a binary mask image was obtained,for reducing false alarm phenomena and time consuming caused by these background regions.Features of histogram of oriented gradients on foreground regions were extracted,and support vector machines classifier was used to detect pedestrian objects and a set of pedestrian objects’ rectangles was obtained.Prior and likelihood of each pixel belonging to the pedestrian were computed based on the binary mask image and the set of pedestrian objects’ rectangles,and posterior probability was estimated and a probability image was generated based on Bayesian theory.OTSU method was used to segment the probability image adaptively,and final pedestrian detection results of each frame were obtained.The pedestrian detection results on Caltech dataset show that,the proposed method has high true positive and detection frame rate,and very low false positive.