基于改进卷积神经网络的人体检测研究  

Research on pedestrian detection based on improved convolutional neural network

作　　者：左艳丽[1] 马志强[2] 左宪禹[1] ZUO Yanli;MA Zhiqiang;ZUO Xianyu(Computer and Information Engineering College,Henan University,Kaifeng 475000,China;Department of Electronic Information Engineering,Henan Vocational College of Agriculture,Zhengzhou 451450,China)

机构地区：[1]河南大学计算机与信息工程学院,河南开封475000 [2]河南农业职业学院电子信息工程系,河南郑州451450

出　　处：《现代电子技术》2017年第4期12-15,共4页Modern Electronics Technique

基　　金：国家自然科学基金资助项目:异构多核并行机上线性代数方程组的快速算法研究(61202098)

摘　　要：为了解决梯度方向直方图在复杂背景下行人检测性能不足的问题,引入深度学习算法进行人体特征提取和行人检测。为了减少卷积神经网络的训练样本数量需求,在保证原数据库背景分布和行人分辨率的基础上使用基于内容的图像检索方法进行数据扩充以便于训练。为了提高算法在复杂背景下的检测效率,在卷积神经网络反射传播权值更新时引入费舍尔约束准则,使用误差反向传播算法获取样本类内类间约束函数的权值,在考虑误差的同时保证算法的分类精度。对INIRIA数据库检测结果表明,改进后算法的漏检率、检测率等性能得到一定提高,在大多数复杂背景下可以成功检测出行人。The deep learning algorithm was introduced to execute the human body feature extraction and pedestrian detection because of the low performance of pedestrian detection histogram of oriented gradient in the complex background. The content?based image retrieval method is used for data expansion to reduce the quantity demand of the training samples of convolutional neural network. The method is able to ensure the original database background distribution and pedestrian resolution. The Fisher criterion is imported when the reflection propagation weights of the convolutional neural network are updated in order toimprove the detection efficiency of the algorithm. The back propagation algorithm is adopted to obtain the weight values of the inter-class constraint function in sample to ensure the classification accuracy while the errors exist. The test results on the INRIA database show that the omission rate and the detection rate of the improved algorithm have been improved,and can detect pedestrians in the most complex-backgrounds successfully

关 键 词：行人检测 深度学习 卷积神经网络 复杂背景 

分 类 号：TN711-34[电子电信—电路与系统] TP139[自动化与计算机技术—控制理论与控制工程]