基于CNN-SVM的行人活动识别方法  

A pedestrian activity recognition method based on CNN-SVM

作　　者：张帅 李召洋 陈建广 黄风华 ZHANG Shuai;LI Zhaoyang;CHEN Jianguang;HUANG Fenghua(School of Smart City,Chongqing Jiaotong University,Chongqing 400074,China;Key Laboratory of Spatial Information Perception and Intelligent Processing in Fujian Province(Sunshine College),Fuzhou 350015,China;Zhejiang Deqing Zhilu Navigation Research Institute Ltd.,Deqing,Zhejiang 313200,China)

机构地区：[1]重庆交通大学智慧城市学院,重庆400074 [2]福建省空间信息感知与智能处理重点实验室(阳光学院),福州350015 [3]浙江德清知路导航研究院有限公司,浙江德清313200

出　　处：《导航定位学报》2025年第1期87-93,共7页Journal of Navigation and Positioning

基　　金：重庆市自然科学基金面上项目(CSTB2022NSCQ-MSX1625);重庆市教委科学技术研究项目(KJQN202100734);重庆市博士“直通车”科研项目(CSTB2022BSXM-JSX0020);福建省空间信息感知与智能处理重点实验室开放基金项目(FKLSIPIP1008)。

摘　　要：针对传统行人活动识别方法过度依赖人工手动选择和提取特征,导致特征提取难度大及识别准确率低的问题,提出一种基于卷积神经网络结合支持向量机(CNN-SVM)的行人活动识别模型:将数据输入到卷积神经网络(CNN)与归一化指数函数(Softmax)层相结合的网络中进行训练直至网络收敛,收敛的CNN网络用于自动提取行人活动数据特征;然后利用支持向量机(SVM)取代CNN网络的归一化指数函数(Softmax)层来优化分类效果。实验结果表明,所提出的CNN-SVM模型可达到97.77%的识别准确率,优于对比实验模型,具有较好的行人活动识别效果。Aiming at the problem that traditional pedestrian activity recognition methods overly rely on manual feature selection and extraction,resulting in high difficulty in feature extraction and low recognition accuracy,the paper proposed a pedestrian activity recognition model based on convolutional neural network(CNN)combined with support vector machine(SVM):the data were input into a network that combines CNN and the normalized exponential function(Softmax)layer for training until the network converged,and the converged CNN network was used to automatically extract pedestrian activity data features;then,the Softmax layer of the CNN network was replaced by SVM to optimize the classification performance.Experimental result showed that the proposed CNN-SVM model could achieve a recognition accuracy of 97.77%,which would be better than the comparative experimental models with a good pedestrian activity recognition effect.

关 键 词：行人活动识别 卷积神经网络(CNN) 支持向量机(SVM) 惯性传感器 深度学习 

分 类 号：P228[天文地球—大地测量学与测量工程]