|
Abstract
|
|
|---|---|
| This paperproposestheadaptationofwell-knownstrategiessuccessfullyusedinspeech processing: MelFrequencyCepstralCoefficients (MFCCs)andPerceptualLinearPrediction (PLP) coefficients. AdditionallycharacteristicslikeRASTA filtering ordeltacoefficients are also consideredandevaluatedforinertialsignalprocessing.Theseadaptationshavebeen incorporated intoaHumanActivityRecognitionandSegmentation(HARS)systembased on HiddenMarkovModels(HMMs)forrecognizingandsegmentingsixdifferentphysical activities: walking,walking?upstairs, walking-downstairs,sitting,standingandlying. All experimentshavebeendoneusingapubliclyavailabledatasetnamedUCIHuman ActivityRecognitionUsingSmartphones,whichincludesseveralsessionswithphysical activity sequencesfrom30volunteers.Thisdatasethasbeenrandomlydividedintosix subsets forperformingasix-foldcrossvalidationprocedure.Foreveryexperiment, averagevaluesfromthesix-foldcross-validationprocedureareshown. The resultspresentedinthispaperovercomesignificantly baselineerrorrates,con- stitutingarelevantcontributioninthe field. AdaptedMFCCandPLPcoefficients improve human activityrecognitionandsegmentationaccuracieswhilereducingfeaturevector size considerably.RASTA-filtering anddeltacoefficients contributesignificantly toreduce the segmentationerrorrateobtainingthebestresults:anActivitySegmentationError Rate lowerthan0.5%. | |
|
International
|
Si |
|
JCR
|
Si |
|
Title
|
Signal Processing |
|
ISBN
|
0165-1684 |
|
Impact factor JCR
|
2,063 |
|
Impact info
|
|
|
Volume
|
120 |
|
|
10.1016/j.sigpro.2015.09.029 |
|
Journal number
|
|
|
From page
|
359 |
|
To page
|
372 |
|
Month
|
MARZO |
|
Ranking
|
Journal Rank in Category 66/255; Quartile in Category Q2) |