step counting investigation with smartphone sensorsstep counting investigation with smartphone...

Rochester Institute of Technology

Department Of Computer Science

FALL 2016

Step Counting Investigation withSmartphone Sensors

Author:Rohit Dilip Giyanani

SupervisorDr. Leon Reznik

December 12, 2016

Step Counting Investigation with Smartphone Sensors Rohit Dilip Giyanani

Contents

1 Introduction 2

2 Background Research 32.1 Data collection for the Step Counter Accuracy Study . . . . . . . . . 32.2 Results from the Step Counter Accuracy Study . . . . . . . . . . . . 42.3 Inferences from the study . . . . . . . . . . . . . . . . . . . . . . . . . 4

3 Experimental Procedure 53.1 Data Acquisition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63.2 Approaches considered . . . . . . . . . . . . . . . . . . . . . . . . . . 73.3 Step Counting Mechanism . . . . . . . . . . . . . . . . . . . . . . . . 8

4 Results 11

5 Conclusion 13

1


Abstract

Over the past 25 years, cellphones have gone from being devices withouteven a display which are only used for calling, to all-in-one devices capableof doing everything. They can not only do everything, but they are able toperform so well that they have singlehandedly eradicated product lines likeIPods, Pagers, most digital cameras and for that matter, even reduced theneed for a laptops. On the other hand, they are still lagging in terms of fitnesstracking, more specifically step counters. Step Counters, in some form, havebeen around since at least the last 400 years [1]. They used to be and are stillavailable as standalone devices whose sole or major responsibility was to trackthe number of steps taken by a user. Today’s dedicated fitness trackers seemto have a really high accuracy rate of around 97% [8]. Smartphone sensorson the other hand, are still playing catch up. Even though the quality of theMEMS on smartphone sensors have improved dramatically over the past 10years or so, the step counts from even the flagship phones on the market todaystill have really high error rates. There is a need to improve the quality ofthe step counters in smartphones and this is the problem being tackled in thisproject.

1 Introduction

Step Counters have gained popularity in the past few years. While heath consciouspeople have always been interested in keeping a track of their step counts, caloriesthey burn and the distances they cover during their workout sessions or during theirdaily life, even the average person now tracks the same metrics more regularly. Thishas been attributed to the rise of readily available and built in applications in theirsmartphones which facilitate this behavior.

Smartphones on both Android and iOS have had step counters since the past 6years, and their accuracy as well as the general structure and purpose the appli-cations has improved gradually over time. The have evolved into relatively moreaccurate step counters with a feature set that now also includes entire slew of fitnesstracking metrics like distance walked, steps climbed and the new Samsung phoneseven having a heartrate sensor built right into the hardware of the phone. Whilethese advancements improve the package as a whole, the step counting is still notaccurate enough.

There is a growing trend of people switching to FitBit’s line of step counting wear-ables and even some lesser known brands. This shift of not relying on smartphoneapplications to dedicated fitness trackers is due to the perception that dedicated

2


fitness trackers generally are more accurate as compared to the smartphone ones.This is supported to by the studies conducted across the board like in the case of [2]and also in a background research conducted as part of preparation for this project.

The goal of the project is to do an experimental study to try and improve the qualityof smartphone sensors. The reason is that since we already depend on smartphonesfor pretty much everything we do, why not have a accurate step counter on it too.Thiscan be accomplished by utilizing either one or multiple smartphone sensors. Forthis, the sensors such as the accelerometer, gyroscope and the magnetometer wereconsidered. Another aspect of this project is to find out how the positioning ofsmartphones with respect to the body affects the step counting.

2 Background Research

Before concluding on whether smartphone step counters are actually inaccurate andthat the fitness trackers in dedicated wearables were better at step counting asdiscussed in [2] and [8], it made sense to actually conduct a study to verify whetherthis assertion. Hence, an analysis of smartphone sensor data was conducted, whichincluded research about sensors like gyroscope, accelerometer and magnetometer,but mainly focused on a study to compare the quality of smartphone sensors insmartphones versus a dedicated fitness tracking wearable.

2.1 Data collection for the Step Counter Accuracy Study

For this study[3], the smartphones considered were the Samsung Galaxy S7 and theNexus 6P and the dedicated fitness tracking wearable chosen was the FitBit Alta. Tohave a broader analysis of the quality of sensors and the impact the final algorithmhas on the quality of the step counting a combination of third party and built inapplications were installed on the two smartphones. For the Samsung Galaxy S7, theGoogle Fit, Noom Walker, Pacer and Samsung’s own S Health apps were installed.As for the Nexus 6P, the Google Fit, Pacer and Noom Walker apps were installed.

Two subjects were each assigned a FitBit Alta and one of the smartphones and askedto use them daily whenever they were walking. Precaution was taken to ensure thatboth the subjects were using both the devices together always so that the data wasas accurate as possible. Data for 20 days was collected from each of them. So, thestep count data from various applications from both smartphones and the FitbitAlta were collected.

3


2.2 Results from the Step Counter Accuracy Study

Now, the step count data from the various FitBit devices has been known to beaccurate within a 2-3% error rate margin. Hence, for this study, the step count datafrom them was considered as the baseline. The raw data from subject 1 using theNexus 6P and subject 2 using the Samsung Galaxy S7 along with their respectiveFitbit data is represented in Figures 1 and 2 [3] as tables. Also a graph representingthe step data for each day is displayed in Figures 3 and 4[3].

Figure 1: Step count table for Samsung Galaxy S7 and the assigned FitBit Alta

From the step data, it is clear that there is a large discrepancy between the datacollected in the step counts from the smartphones. The percent error of the datacollected from the smartphones when the step data from the Fitbit is consideredabsolute is represented in Figures 5 and 6[3].

2.3 Inferences from the study

Also, another interesting analysis can be made by deriving the cumulative errorvalues when the data from both smartphones is used together. We can gauge theoverall performance of the smartphone in general and also identify the error rates ofeach application. This is represented in Figure 7 [3]. It is clear from this, that the

4


Figure 2: Step count table for Nexus 6P and the assigned FitBit Alta

step counter quality of the Google Fit application is really good. It ranges in theregion of roughly 3-4% error rate.

Another factor is that overall, the error rates on the flagship devices from Google andSamsung was still averaging between 7-10% error rate. This implies that on deviceswith not flagship like qualities, especially ones which don’t even have a dedicatedstep counter MEMS, the quality will be somewhere in the range of 20% error rate.This is because, Noom Walker does not use the data from the step counter on asmartphone to calculate the number of steps. Hence, there is a need to develop abetter method to get the step count on smartphones.

3 Experimental Procedure

For this project, Android platform was chosen. While researching for the best ap-proach to improve the accuracy of smartphone sensors I researched techniques fromvarious papers which included using the various built in sensors in Android devices.[7], [4] were a few papers which provided techniques to use Accelerometer data forstep counting. A few spoke about using magnetometer for step counting whereas [5]and [6] represented gyroscope based smartphone sensors.

While all of them provided with compelling reasons to choose any of the above

5


Figure 3: Step count graph for Samsung Galaxy S7 and the assigned FitBit Alta

mentioned sensors, based on the techniques mentioned in the papers and the givenaccuracy and techniques, the gyroscope based technique mentioned in [5] seemed tobe the optimal choice.

3.1 Data Acquisition

The data collection for this project involved creating an Android application. Thisapplication is responsible to collect the information from various smartphone sensors.In this case, since this was an experimental analysis and there was a possibilitythat the technique mentioned in [5] may not be as optimal as represented and ifsome modifications to that technique still did not yield any positive results, theaccelerometer and gyroscope data both were collected. This application has a startand stop button. Figure 8 represents the layout of this application.

The accelerometer and gyroscope sensors on smartphones are really sensitive. Eventhe smallest movements that you would not even notice, could trigger the accelerom-eter and gyroscope to record a change. The application was designed in such a waythat all these movements were recorded.

The procedure to do the data collection was carefully designed. Five subjects were

6


Figure 4: Daily step count for Nexus 6P and the assigned FitBit Alta

chosen. Each of these subjects had a different height and body structure so that thedata could be as diverse as possible. Each subject was given 2 smartphones, one toplace in the pocket of their pants or jeans and the other one in their shirt pocket.Each of the 5 subjects had to walk 10 times. Starting with 50 steps, then 100 steps,then 150 steps, all the way to 500 steps at intervals of 50 steps. Two Google Pixelsmartphones were used for each and every walk so that there is no issue of differenthardware affecting the accelerometer and gyroscope data discrepancies. For eachwalk, the start time, end time, actual number of steps and all the accelerometer andgyroscope data was recorded.

3.2 Approaches considered

One of the first approaches considered was one that was directly described in [5].This approach was described in the paper as one which would even work when aperson would be walking slowly. To achieve this, they used a sixth-degree low passButterworth filter on the gyroscope data. The data collection technique in the appli-cation needed to be changed for this testing. This is because the Butterworth filterrequired the gyroscope data to be collected at fixed intervals. When the filter wasthen applied to the gyroscope data, the number of steps that were finally calculated

7


Figure 5: Daily Error Percentage for the Samsung Galaxy S7 w.r.t to FitBit Alta

was very low. This could be attributed to the fact that since the data was collectedat fixed intervals to facilitate the possibility of using the Butterworth low pass filter,it left out some of the steps and hence the step counter was off.

Another approach to counter this problem was to use a Gaussian low pass filterinstead of the Butterworth filter, but even this approach yielded slightly better butnonetheless inaccurate results.

3.3 Step Counting Mechanism

Out of the two approaches considered above one of them is used in [5]. The Butter-worth filter and the Gaussian filter are both preprocessing techniques. These andmany such approaches are sometimes used in the analysis of gyroscope data becausethe data from the gyroscope is usually very noisy. The reason for this apart fromother reasons is that this sensor as well as most Android smartphone MEMS arevery sensitive. This means that they are triggered by even the smallest movements.Hence, it is necessary to make sure that the data collected from these sensors ishandled carefully and by keeping this into consideration. So, while none of thestandard filtering techniques are not applied to the data, an alternate approach to

8


Figure 6: Daily Error Percentage for the Nexus 6P w.r.t to FitBit Alta

identification of real significant data versus redundant data was devised.

The data collection phase in its current phase has one major drawback. The appli-cation as displayed in Figure 8, has a start and stop button to help facilitate thestep recording process. After clicking the start button, the phones need to be placedin the shirt and pant pockets. This means that during this time, the recording isalready on, and that this may interfere with the step counting process. Hence, theinitial and final few readings are discarded from the analysis. This is done by ametric which checks the total time, and accordingly discards the first and last 5seconds of the data collected. Since, each reading from the sensor is tagged with atimestamp, this can be accomplished. This cleaning takes place every time a newwalk session is started.

Now, based on the method mentioned in [5], almost all smartphones in the pantpocket and the shirt pocket are usually placed vertically or slightly inclined. Itis also mentioned that most humans walk in such a way that the motion of theirbody while walking is sinusoidal. This can be easily identified by the pattern inwhich ones leg moved while walking. Thus, according to this, the X axis motionon the gyroscope would be in sinusoidal motion. This is the basic principle used inthis project’s proposed technique too. So since the motion is sinusoidal, every time

9


Figure 7: Percentage error rate w.r.t to FitBit Alta

there is a zero crossing, a step can be recorded and in that way, step count can becalculated. In reality, if we use this approach on the cleaned gyroscope data, weget more than double the actual number of steps, even triple in some cases. Henceanother important factor to be able to count steps is that we need to be able toidentify the real steps from huge amount of data collected from the sensors. Thedata collected from the walks is used for this purpose.

The fact that raw gyroscope data estimates two to three times more steps than theactual number of steps indicates that there are a large number of false positives.This can be identified easily by observing Figure 9. This figure represents a graphof gyroscope data. As you can observe, there are a lot of places where there is azero crossing, but it represents data points which are not conclusive enough to besteps. Hence, as opposed to filtering of such points in the form of a Butterworthor a Gaussian filter, a more granular approach of dealing with actual gyroscopecutoff values was determined. This was done, by analysis of actual steps versus thesteps calculated when the cutoff values were modified. Since the data collected wassignificant in size, suitable lower and upper thresholds for cutoff values were selected.Obviously, since the data was collected in both the pant pocket and the shirt pocket,cumulative cutoff points were selected. Also individual cutoff values for both pantand shirts were calculated separately too. The values represent the distance between

10


Figure 8: Layout of data collection application

the crest and trough values. The cumulative, pant pocket and shirt pocket cutoffvalues for the upper and lower thresholds are displayed in Figure 10.

With the help of these cutoff values, the real steps can be determined. During theanalysis of the gyroscope data after a walk, whenever there is a zero crossing, thecorresponding crest and trough data is maintained and if the difference betweenthese values is in between the upper and higher threshold values, a step is counted.During the span of a walk, every time this condition is satisfied, the number of stepsare incremented. Hence, the total step count is calculated using this approach.

4 Results

As mentioned in the previous section, 3 separate metrics were used to calculatethe total number of steps taken. These are based on the cutoff values specified forshirt, pant and a common metric which is a cumulative cutoff value which shouldideally work for both. The results from the experiments are specified in Figure 11.

11


Figure 9: X-Axis Gyroscope data

Figure 10: Cut Off Values

It is evident that if the specific metrics for shirt and pant are used, the accuracy ofthe step counting is close to 87%. Whereas, if the common metric is used for stepcounting filtration, the accuracy is reduced to an average of around 79%. So clearly,this approach works best when the metrics specific to the position of the smartphonesensor are utilized. The data is also visualized in a graphical manner in Figure 12.

Another interesting trend in the results, which can be observed in Figures 11 and12 is that, based on the current metrics, the position of the phone impacts the finalstep count tremendously. There is a significant difference between the step countsobserved for the same walk, where the step count from the phone placed in the pantpocket is on average higher than the original step count, and when placed in theshirt pocket, is lower than the step count. This can be attributed to the fact thatthe number of significant sinusoidal movements will be lower when the phone is in

12


Figure 11: Step Count Values for Different Positions and Metrics

the shirt pocket, as compared to when the phone is in the pant pocket.

A table representing the percentage error for different measures of actual steps isrepresented in Figure 13. It indicates that on an average, the accuracy of the stepcounting remains roughly the same, improving slightly as the number of steps in-crease.

5 Conclusion

With a dedicated metric for positioning of the smartphone in both pant and shirtpocket, the technique used in this paper was able to achieve an accuracy of 87%on an average. If we consider, only the positioning of the smartphone in the pantpocket, which is the most common place, the accuracy is improved even further.Based on the results, it is clear that, as compared to other smartphone applicationsthat don’t use the built in step counter MEMS, this project provides an improvementin the step counting accuracy.

While the technique specified in this project does improve the accuracy of stepcounters over applications that don’t use the dedicated step counter MEMS, thereis still an effort needed to improve on the ones that do. A few improvements like,a larger set of data samples for the analysis of the cut off values would improve theaccuracy of the step counting even further.

13


Figure 12: Graphical Representation of Step Counts

Figure 13: Percentage Errors of Step Counts for Different Positions and Metrics

14


The proposed technique could also be extended by working on technique for adaptivecutoff values. Essentially, if the person walking starts walking faster, lengtheningtheir stride, the threshold value should be adapted accordingly to accommodate thischange in walking pattern. Such improvements could improve this project even moreproviding us with a perfect step counting technique which could be used in even themost basic smartphone.

15


References

[1] Who Invented the Pedometer, 2016. http://eachstepyoutake.com/

who-invented-the-pedometer/.

[2] Meredith A Case, Holland A Burwick, Kevin G Volpp, and Mitesh S Patel.Accuracy of smartphone applications and wearable devices for tracking physicalactivity data. Jama, 313(6):625–626, 2015.

[3] Rohit Dilip Giyanani. A Review of Mobile Sensors, 2016. unpublished.

[4] W. Hongman, Z. Xiaocheng, and C. Jiangbo. Acceleration and orientation mul-tisensor pedometer application design and implementation on the android plat-form. In 2011 First International Conference on Instrumentation, Measurement,Computer, Communication and Control, pages 249–253, Oct 2011.

[5] S. Jayalath and N. Abhayasinghe. A gyroscopic data based pedometer algorithm.In 2013 8th International Conference on Computer Science Education, pages551–555, April 2013.

[6] Y. P. Lim, I. T. Brown, and J. C. T. Khoo. An accurate and robust gyroscope-based pedometer. In 2008 30th Annual International Conference of the IEEEEngineering in Medicine and Biology Society, pages 4587–4590, Aug 2008.

[7] Martin Mladenov and Michael Mock. A step counter service for java-enableddevices using a built-in accelerometer. In Proceedings of the 1st InternationalWorkshop on Context-Aware Middleware and Services: Affiliated with the 4thInternational Conference on Communication System Software and Middleware(COMSWARE 2009), CAMS ’09, pages 1–5, New York, NY, USA, 2009. ACM.

[8] Judit Takacs, Courtney L. Pollock, Jerrad R. Guenther, Mohammadreza Bahar,Christopher Napier, and Michael A. Hunt. Validation of the fitbit one activitymonitor device during treadmill walking. Journal of Science and Medicine inSport, 17(5):496–500, 09 2014. Copyright - Copyright Copyright Agency Limited(Distributor) Sep 2014; Document feature - Tables; ; Graphs; Last updated -2016-06-18.

16

step counting investigation with smartphone sensorsstep counting investigation with smartphone...

Documents