摘要
大数据环境下,对行人检测的需求度不断提高,然而视频中的信息越来越丰富,视频中所获取的场景也愈加复杂。在如此背景下,目前大多使用卷积神经网络进行识别,但识别率不高。在原有的胶囊网络模型的基础上,增加了两层卷积层并将胶囊维度进行了扩展,同时使用了动态路由迭代算法,提出了一种基于改进胶囊网络的行人识别模型(PRM-ICN),该网络能够更有效地减少复杂背景中多余信息的干扰。实验在TensorFlow框架下使用三个国际知名且有一定难度的公开通用数据集CUHK01、CUHK03和Market-1501上进行验证,并将结果与PRM-AlexNet和PRM-VGG-16两个著名的行人识别网络相对比。实验结果表明在三个数据集上,所提出的网络模型在CMC曲线和MAP指标下都要优于其他两个网络,证明了所提模型在复杂场景下识别效果的优越性。
In the context of big data,the demand for pedestrian detection is constantly increasing.However,the information in the video is getting more and more abundant,and the scenes acquired in the video are also becoming more and more complicated.Under such background,convolutional neural network is mostly used for recognition at present,but the recognition rate is not high.Based on the original capsule network model,two convolutional layers are added and the capsule dimension is extended.At the same time,according to dynamic routing iteration algorithm,a pedestrian recognition model based on the improved capsule network(PRM-ICN)is proposed,which can more effectively reduce the interference of redundant information in the complex background.The experiments are verified under the TensorFlow framework using three publicly available datasets CUHK01,CUHK03 and Market-1501,and the results are compared with two famous pedestrian recognition networks,PRM-AlexNet and PRM-VGG-16.Experiment shows that on the three data sets,the proposed network model is greater than another two networks under the CMC curve and the MAP index,which proves its superiority in complex scene recognition.
作者
程换新
刘文翰
郭占广
张志浩
CHENG Huan-xin;LIU Wen-han;GUO Zhan-guang;ZHANG Zhi-hao(School of Automation and Electronic Engineering,Qingdao University of Science and Technology,Qingdao 266061,China)
出处
《计算机技术与发展》
2021年第2期75-79,共5页
Computer Technology and Development
基金
国家海洋局重大专项项目(国海科字[2016]494号No.30)。