Automated time activity classification based on global positioning system (GPS) tracking data

Environmental Health

Research

Dean Baker¹  Ralph Delfino²  Douglas Houston³  Chengsheng Jiang⁴  Jun Wu⁵ 

[1] Center for Occupational & Environmental Health, University of California, Irvine, USA;Department of Epidemiology, School of Medicine, University of California, Irvine, USA;Department of Planning, Policy and Design, School of Social Ecology, University of California, Irvine, USA;Program in Public Health, University of California, Irvine, USA;Program in Public Health, University of California, Irvine, USA;Department of Epidemiology, School of Medicine, University of California, Irvine, USA;

关键词: Global Position System;    Random Forest;    Static Cluster;    Global Position System Data;    Random Forest Model;   

DOI  :  10.1186/1476-069X-10-101

 received in 2011-06-07, accepted in 2011-11-14,  发布年份 2011

来源: Springer

【 摘 要 】

BackgroundAir pollution epidemiological studies are increasingly using global positioning system (GPS) to collect time-location data because they offer continuous tracking, high temporal resolution, and minimum reporting burden for participants. However, substantial uncertainties in the processing and classifying of raw GPS data create challenges for reliably characterizing time activity patterns. We developed and evaluated models to classify people's major time activity patterns from continuous GPS tracking data.MethodsWe developed and evaluated two automated models to classify major time activity patterns (i.e., indoor, outdoor static, outdoor walking, and in-vehicle travel) based on GPS time activity data collected under free living conditions for 47 participants (N = 131 person-days) from the Harbor Communities Time Location Study (HCTLS) in 2008 and supplemental GPS data collected from three UC-Irvine research staff (N = 21 person-days) in 2010. Time activity patterns used for model development were manually classified by research staff using information from participant GPS recordings, activity logs, and follow-up interviews. We evaluated two models: (a) a rule-based model that developed user-defined rules based on time, speed, and spatial location, and (b) a random forest decision tree model.ResultsIndoor, outdoor static, outdoor walking and in-vehicle travel activities accounted for 82.7%, 6.1%, 3.2% and 7.2% of manually-classified time activities in the HCTLS dataset, respectively. The rule-based model classified indoor and in-vehicle travel periods reasonably well (Indoor: sensitivity > 91%, specificity > 80%, and precision > 96%; in-vehicle travel: sensitivity > 71%, specificity > 99%, and precision > 88%), but the performance was moderate for outdoor static and outdoor walking predictions. No striking differences in performance were observed between the rule-based and the random forest models. The random forest model was fast and easy to execute, but was likely less robust than the rule-based model under the condition of biased or poor quality training data.ConclusionsOur models can successfully identify indoor and in-vehicle travel points from the raw GPS data, but challenges remain in developing models to distinguish outdoor static points and walking. Accurate training data are essential in developing reliable models in classifying time-activity patterns.

【 授权许可】

CC BY   
© Wu et al; licensee BioMed Central Ltd. 2011