| 摘要: |
| 针对传统的非足部基于人员轨迹盲推的室内定位（Pedestrian Dead Reckoning,PDR）方法仅适合单个行为条件下的定位，无法适应真实定位场景的问题，提出了一种基于深度行为分类的人员轨迹盲推方法。该方法首先利用基于循环神经网络（Recurrent Neural Network,RNN）的深度学习进行人员行为分类，并且根据分类结果（手持条件下、口袋中、自由晃动）设计不同的模型进行迈步检测、步长估计和航向估计，从而对人员的位置进行估计。实测实验验证了所提方法的有效性，其行为识别的准确率达到98%，并且相比于两种传统的单一行为定位方法，定位精度分别提高了1.7 m和2.3 m。 |
| 关键词: 室内定位 轨迹盲推（PDR） 行为分类 深度学习 |
| DOI： |
|
| 基金项目:重庆市教委科学技术研究计划项目青年项目(KJQN202102403)；基于物联网健康管理系统的异构无线网关研究项目（KJQN201802404）；新工科背景下基于学科交叉融合视角下的人工智能专业课程体系研究（20352） |
|
| Pedestrian dead reckoning based on deep motion recognition |
| ZHENG Qiuju,CHEN Xin |
| (Institute of Communication and Information Engineering,Chongqing College of Mobile Communication,Chongqing 401520，China) |
| Abstract: |
| Traditional none-foot-mounted pedestrian dead reckoning(PDR) methods only consider situations with single motion activities,and are not appropriate for adoption in real positioning scenarios.For this problem,a method for PDR based on deep activity recognition is proposed.Firstly,a recurrent neural network(RNN) is adopted for classification of the pedestrian’s motion types(herein denoted as hand-held,pocket-mounted and free swing).Then according to the motion types,different models are designed for the processes of step detection,step length estimation and heading estimation,which are basically the procedures of PDR.Real scenario experiments validate the proposed method.The result shows that the correct motion recognition rate can reach up to 98%,and the positioning accuracy has increased by 1.7 m and 2.3 m respectively compared with two single motion PDR methods. |
| Key words: indoor positioning pedestrian dead reckoning(PDR) motion classification deep learning |
