Accelerometer-based Transportation Mode Detection on Smartphones Samuli Hemminki, Petteri Nurmi, Sasu Tarkoma Helsinki Institute for Information Technology.

wine.korea.ac.kr

WINE

Accelerometer-based Transportation Mode Detection on Smartphones

Samuli Hemminki, Petteri Nurmi, Sasu TarkomaHelsinki Institute for Information Technology

Proceedings of the 11th ACM Conference on Embedded Net-worked Sensor Systems, 2013

배문규

wine.korea.ac.kr

WINE 2

Introduction Communication management Protocol description Validation Conclusion

Contents

wine.korea.ac.kr

WINE 3

Contributions An improved algorithm for estimating the gravity component of ac-

celerometer measurements A novel set of accelerometer features that are able to capture key charac-

teristics of vehicular movement patterns A hierarchical decomposition of the detection task

Results It is able to improve transportation mode detection by over 20% com-

pared to current accelerometer-based systems

Introduction

wine.korea.ac.kr

WINE 4

A positive impact on many research fields Automatically monitor the transportation behavior of individuals Unban planning Monitoring and addressing the spread of diseases and other hazards Providing emergency responders information of the fastest route Localization and positioning algorithms applications: -footprint, level of physical activity User profiling

Challenge To distinguish information pertaining to movement behavior from other

factors that affect the accelerometer signals

Introduction

wine.korea.ac.kr

WINE 5

Transportation mode detection: overview

wine.korea.ac.kr

WINE 6

Three dimensional acceleration measurement Preprocessing the raw measurements by applying a low-pass filter that

retains 90% of energy Aggregating the measurements using a sliding window with 50% overlap

and a duration of 1.2 seconds Projecting the sensor measurements to a global reference frame by esti-

mating the gravity component along each axis and calculating gravity eliminated projections of vertical and horizontal acceleration

Preprocessing and gravity estimation

wine.korea.ac.kr

WINE 7

Algorithm for estimating the gravity component of accelerome-ter measurements

Preprocessing and gravity estimation

wine.korea.ac.kr

WINE 8

Preprocessing and gravity estimation

wine.korea.ac.kr

WINE 9

Frame-based features Statistical features, time-domain metrics, frequency-domain metrics These are able to effectively capture characteristics of high-frequency mo-

tion

Peak-based features These characterize acceleration and deceleration period to capture fea-

tures from key periods of vehicular movement Extract these features only during stationary and motorized periods peak areas

Feature extraction

wine.korea.ac.kr

WINE 10

Segment-based features These characterize patterns of acceleration and deceleration periods

Feature extraction

wine.korea.ac.kr

WINE 11

Full list of the features

Feature extraction

wine.korea.ac.kr

WINE 12

Instance-based classifier Decision tree, SVM

Hidden Markov Model (HMM) For the kinematic motion classifier

Classification

wine.korea.ac.kr

WINE 13

Adaptive boosting (AdaBoost) To iteratively learn weak classifiers that focus on different subsets of the

training data and to combine these classifiers into one strong classifier This paper uses decision trees with depth of one or two as the weak clas-

sifier Boosting rounds T was determined using the scree-criterion

Decision tree

Classification

wine.korea.ac.kr

WINE 14

Segment-based classification It is performed for the stationary and the motorized classifiers Acquiring classification results from the two information source

Aggregating classification results of frame- and peak-based features over the observed segment (simple voting scheme)

Computing the classification result of the segment-based features over the observed segment

Obtaining the final classification by combining the results of the two clas-sifier outputs

Average of the two source

Classification

wine.korea.ac.kr

WINE 15

Kinematic motion classifier (99%) It utilizes the frame-based accelerometer features extracted from each

window to distinguish between pedestrian and other modalities It uses decision trees with depth one

Classification

It captures the repetitive nature of walking, typically with 1 – 3s interval

wine.korea.ac.kr

WINE 16

Stationary classifier (90%) It uses both the peak features and the frame-based features for distin-

guishing between stationary and motorized periods 15 weak learners, each comprising a decision tree of depth two

Classification

wine.korea.ac.kr

WINE 17

Motorized classifier (80%) It is responsible for distinguishing between different motorized trans-

portation modalities 20 weak leaners, decision trees of depth two

Classification

wine.korea.ac.kr

WINE 18

Motorized classifier (80%) Car, bus, tram vs. train, metro

The frequency of acceleration and breaking peaks car vs. other

the intensity, length of the acceleration and breaking period tram vs. other

the intensity, volume of acceleration and breaking periods Frame-based features to distinguish vehicles on roads and rails

it can capture characteristics of vertical movement as well as overall noisiness of the measurements

Classification

wine.korea.ac.kr

WINE 19

Evaluation

wine.korea.ac.kr

WINE 20

A novel accelerometer techniques for transportation mode de-tection on smartphones

This generalizes well across user and geographic locations

Critics Performance evaluation (estimation delay) is nowhere The detailed algorithms or information about classifier does not exist It is needed to compare other accelerometer-based mode detection algo-

rithm

Conclusion