摘要：

在方向梯度直方图(HOG)联合支持向量机(SVM)算法(HOG-SVM)和LeNet网络模型基础上,提出了HOG与卷积神经网络(CNN)融合的行人检测算法(SVM-LeNet)。采用多尺度滑动窗口提取HOG特征并送入SVM分类器,根据后验概率判断候选区,随后运用CNN算法剔除误检窗口。为解决单个目标被多个候选区域框定的问题,使用非极大值抑制算法(NMS)进行多矩形融合,保留检测区域中后验概率最大的窗口抑制与其重叠的检测窗口。分类过程中,以候选区域在SVM和LeNet中后验概率为依据判断行人区域。实验结果表明,与HOG-SVM和LeNet行人检测算法相比,该算法在准确率和召回率上有明显优势。

关键词: 行人检测, 权重模板, 支持向量机, 非极大值抑制算法, 卷积神经网络

Abstract:

On the basis of Histogram of Oriented Gradient with Support Vector Machine(HOG-SVM) algorithm and LeNet network model, a pedestrian detection algorithm which is the combination of HOG and Convolutional Neural Network(CNN) is proposed. Firstly, multi-scale sliding windows are used to extract the HOG features which are then sent to SVM classifier to find the candidate regions.The regions are judged according to the posterior probability. And the CNN algorithm is used to eliminate the false detection window. In order to solve the problem that a single target is framed by multiple candidate regions, the Non-maximum Suppression(NMS) algorithm is used to fuse the multi-rectangles, remaining the largest posterior probabilitywindow and suppressing the overlapped windows.In the classifying process, the candidate region is judged as pedestrian region based on the posterior probability in SVM and LeNet. Experimental results show that this algorithm can get higher recognition rate and recall rate compared with HOG-SVM and LeNet algorithms.

中图分类号: 

• TP391.41