基于SVM-LeNet模型融合的行人检测算法

发布时间：2018-03-28 02:36

本文选题：行人检测　切入点：权重模板　出处：《计算机工程》2017年05期

【摘要】：在方向梯度直方图(HOG)联合支持向量机(SVM)算法(HOG-SVM)和Le Net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-Le Net)。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Le Net中后验概率为依据判断行人区域。实验结果表明,与HOGSVM和Le Net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。
[Abstract]:In the direction of the gradient histogram (HOG) combined with support vector machine (SVM) algorithm (HOG-SVM) and Le Net network model based on HOG is proposed and the convolutional neural network (CNN) pedestrian detection algorithm fusion (SVM-Le Net). The multi-scale HOG feature extraction and sliding window into the SVM classifier, according to posterior probability judgment the candidate region, then use CNN algorithm to remove the false alarms. In order to solve the problem of single target is frame multiple candidate regions, the use of non maxima suppression algorithm (NMS) for rectangular fusion, keep checking probability maximum window suppression and overlap detection area. In the process of classification, the candidate region in SVM and Le Net in the posterior probability to judge the pedestrian area. The experimental results show that compared with HOGSVM and Le Net pedestrian detection algorithm, this algorithm has obvious advantages in accuracy and recall rate.

【作者单位】：武汉工程大学计算机科学与工程学院;武汉工程大学智能机器人湖北省重点实验室;华中科技大学自动化学院;
【基金】：国家自然科学基金(61103136) 武汉工程大学创新基金(CX2015057)
【分类号】：TP391.41

【相似文献】