针对监控场景的行人检测应用,提出一种结合改进的支持向量机和卷积神经网络的行人检测方法.首先,通过运动检测初步定位感兴趣的可疑目标区域;然后,计算这些区域图像块的灰度共生矩阵,并采用主成分分析方法提取纹理特征向量,采用支持向量机进行纹理分类,滤除干扰区域;最后,对余下区域构建多尺度图像子块,采用LeNet5卷积神经网络架构进行行人分类.在Caltech数据集上的测试结果表明,该方法的真正率指标高,假正率指标低.%We present an improved pedestrian detection method for video monitoring application, using support vector machines and convolutional neural networks.First, we located initial suspicious targets of interested area by motion detection, and then calculated the gray level co-occurrence matrix of image patches of these areas.Moreover, the principal component analysis method was used to extract the texture feature vector, and the support vector machine was used to classify the texture and filter out the interference region.Finally, multi-scale image blocks were constructed for the remaining area, the LeNet5 architecture of convolutional neural network was used to execute pedestrian classification.Experimental results on Caltech dataset show that this method has a high true positive rate and low false positive rate.
展开▼
