基于高斯混合模型与PCA-HOG的快速运动人体检测

摘要

Since a traditional human object detection system searched the human object in a huge region with the method of slide window and used a high dimension features to represent the human objects for a better classification taking high memory resources and time consummations, this paper proposed an improved method. An application of Gaussian mixture model (GMM) can coarsely extract the foreground and reduce the region, then the system searching human objects only in the foreground area, which speeded up the detection velocity and reduced the false positives error rate in a practical application. Owing to the high dimension of histogram of oriented gradient ( HOG) feature vector, it proposed a novel descriptor which was called PCA-HOG. This descriptor used a classical dimension reduction method called principal component analysis ( PCA) to reduce the dimension of raw HOC. The novel descriptor had a lower dimension feature vector than that of HOG, so it gave a much faster speed of the detect window classification. The detect rate of PCA-HOG was almost as high as that of HOG. The experiment shows that the human object detection system combined with the PCA-HOG descriptor and Gaussian mixture model gives a remarkable performance of detection rate.%针对传统人体检测系统中由于检测窗口标扫描区域过大,帧的特征维度过高使其在实际应用中内存消耗量大且检测速度慢的情况,提出了改进的运动人体检测方法.该方法利用高斯混合模型进行背景建模剔除掉大部分图像背景,减少了侦测扫描区域,从而在减少负例样本误检率的同时提升了检测速度.同时对处理HOG的高维度,提出了一种基于主成分分析(PCA)降维的梯度方向直方图(HOG)的描述子,即PCA-HOG描述子,它在不降低识别率的前提下,很大程度地提升了侦测窗口的分类速度.实验验证了混合高斯模型与PCA-HOG相结合显著提升了人体检测速度.