Sensor Assisted Wireless Communications

(Invited Paper)

Naveen Santhapuri∗ Justin Manweiler∗ Souvik Sen∗

Xuan Bao∗ Romit Roy Choudhury∗ Srihari Nelakuditi†∗Duke University †University of South Carolina

Abstract

The nature of human mobility demands thatmobile devices become agile to diverse operatingenvironments. Coping with such diversity requiresthe device to assess its environment, and triggerappropriate responses to each of them. While ex-isting communication subsystems rely on in-bandwireless signals for context-assessment and response,we explore a lateral approach of using out-of-bandsensor information. We propose a relatively novelframework that synthesizes in-band and out-of-bandinformation, facilitating more informed communica-tion decisions. We believe that further research in thisdirection could enable a new kind of device agility,deficient in today’s communication systems. Sincesuch a framework is located at the boundaries ofmobile sensing and wireless communications, we callit sensor assisted wireless communications.

1. Introduction

In everyday life, mobile devices transitionthrough a variety of environments. These transi-tions arise not only from changes in the physicalenvironment (due to mobility), but also due tofluctuations within and across different wirelesstechnologies. For instance, a mobile phone carriedby an office-goer may transition from a station-ary state (at home) to a walking state (on theway to the subway station) to a highly mobilestate (within the train). The wireless backgroundchanges as well, subjecting the phone through WiFi(at home), 3G (outdoors), and then a period ofcomplete disconnection, except perhaps when thetrain stops at the stations. Unsurprisingly, theseenvironmental changes strain the communicationsubsystems in these devices. Wireless networkingprotocols must constantly discern the “context” ofcommunication, and adapt with a new agility.

Context-aware device agility remains an elusiveresearch challenge because the contexts areoften difficult to discern based on wireless signalobservations alone. A microwave oven at a Star-bucks may induce the same kind of performancedegradation as observed in a classroom withcongested WiFi users. While the ideal response tothe microwave would be to switch to a differentfrequency channel, less-frequent transmissionattempts is most suitable for alleviating congestion.Clearly, such agility is difficult to attain withoutproper context assessment. Today’s mobile devicesoptimize for the common case, and sacrificeperformance where the context is atypical. Thispaper explores new opportunities for context-discrimination, thereby leading to improved deviceagility for truly pervasive communication.

Our main idea is simple. By using mobile phonesensors as an out-of-band context-assessmenttool, we show that a certain kind of deviceagility may be achieved. For example, the phone’saccelerometer measurements could identify themoving subway train, and switch off the WiFi/GSMsubsystems to save energy. When the phone stopsat each station, or when the user gets off thetrain, the accelerometer readings can pick the cueand switch on wireless access. As a generalization,the growing number of sensors on mobile devicespresents an out-of-band opportunity to discernthe communication context. While these contextshave been abundantly used in mobile computingapplications [5], there is limited work thatconnects them to MAC/PHY layer functions [9].This project attempts to make (and strengthen)this connection. Although an early work, our keycontributions can be summarized as follows.

(1) We identify the opportunity of utilizing out-of-band sensor information to optimize wireless

communication systems (Figure 1). Of particularinterest are the cases where the contextualinformation is implicitly present in the system.

(2) We propose a sensor-assisted wireless network-ing framework, and instantiate it with case-studies.

(3) We discuss our ongoing research on the gen-eralization of these ideas. Our thought experimentsshow that this relatively new research space couldbe a lateral approach to augment existing researchon wireless networking.

Application

TCP/IP

Sensing

MAC/PHY

Microphone,

Camera, Compass

Accelerometer

Weather

GPS,Sensor

Assistance

Network Stack

SAWC

ExistingUses

Figure 1. The Sensor Assisted Wireless Commu-nications (SAWC) framework.

2. Motivation and Opportunity

This section zooms into the scope of sensorassisted wireless communications, and asks a set ofnatural questions. The intent is to characterize theneed for sensor-assistance, gain an understandingof the opportunities, and therefrom, carve out theutility of the system. Three example applicationsare discussed as a verification of the generalideas in specific, real-world settings. Two of theseapplications are quantified through preliminarymeasurements and implementation.

We begin by asking, why is in-band informa-tion inadequate for wireless protocol agility?Our argument is two-fold. (1) Environmentalfactors and user mobility perturb wireless com-munication in diverse and unpredictable ways.When all these factors are further combinedwith interferences from nearby transmissions,the net observable result is purely stochastic.Diagnosing the causes of perturbations fromthe wireless signals is difficult. Moreover, thesignal-level information exported by commoditywireless hardware is coarse-grained, making it

harder to segregate the causes. Nevertheless, theprotocol’s ideal reaction to the perturbation canvary depending on the cause of perturbation. Anin-band approach to may not always be optimal.

A simple example of this can be seen in wirelessrate adaptation. It has been well known that thetransmission rate of a packet must be reducedonly upon channel fading; the rate could remainthe same (or even increase) if the packet loss wasdue to collision. Yet, today’s wireless protocols areunable to reliably discriminate between fading andcollision, and blindly reduce the rate upon packetlosses. This is because in-band discriminationhas been a difficult problem [10], [12]. As weelaborate later, using out-of-band information mayfacilitate failure diagnosis in this case.

(2) Our second argument pertains to theoverhead of in-band approaches. Diagnosingthe cause of signal perturbations may requirebandwidth and energy investments, cutting backon the system’s throughput and/or battery life.When the microwave oven is turned on in thevicinity, the WiFi interface may need to probemultiple channels before switching to the best one.Worse, the interface may need to continue probing(like polling mechanisms) to determine when itmust switch back to the original channel. In-bandprobing will consume channel time, in addition toforcing frequent disassociations/reassociations toaccess points. Identifying the microwave throughan out-of-band sound sensor could be a moreeffective method of channel switching (much likeinterrupt based operations). Since these soundsensors may anyway be active for a variety ofother applications [1], [8], the cost of using themmay get amortized.

Along this thread of argument, the naturalfollow-up question is why would out-of-bandtechniques be any better for discriminatingthe context? If the microwave’s presence cannotbe discerned through RF signals, it may not bediscernible through sound either, i.e., other noisesmay drown the microwave-specific “hum”. Whilethat is true, we first clarify that in-band andout-of-band techniques are complementary, andcould be used to balance out mutual deficiencies.Second, we argue that contextual events mayhave multiple fingerprints scattered over multiplesensing dimensions. Moving in a subway trainmay be identifiable through accelerometers and

sound sensors (trains have a characteristic acousticsignature easy to identify). Since mobile devicesare equipped with a growing number of sensingdimensions, the context may be discernible overat least one of these dimensions. This diversityis likely to improve context-discrimination oversingle-dimensional, in-band techniques.

The next question then is what is the space ofopportunity for out-of-band techniques? Whatare example applications? Figure 2 shows onepossibility of classifying the broad topic of context-awareness from the perspective of wireless com-munications. The rows characterize the sourceof contextual information. When the contextualsource belongs to the same sensing dimension, itis termed in-band, while information derived fromother modes are out-of-band. Of course, one mayargue that GPS location is an out-of-band informa-tion and has been abundantly used in improvingcommunications [9], [5]. In light of this, we findthat information can also be classified into thosethat are explicitly produced for consumers, andothers that are implicitly present in the environ-ment. For example, RTS/CTS packets in 802.11are meant to explicitly alert nearby devices of animminent transmission; GPS also offers explicitlocation information to enable context-awareness.The microwave “hum”, on the other hand, is notmeant for a communication-related operation, andthereby an implicit source of information. A de-vice’s secure identifier, extracted from inherentclock skews and frequency offsets [3], is againimplicit for the same purposes. When viewed inthis manner, we find that existing work has not ad-equately explored the class of out-of-band, implicittechniques. Yet, the number of sensing dimensions,as well as the modes of communications, continueto proliferate on mobile devices. This paper identi-fies this opportunity of information synthesis, andsuggests a holistic approach to improve wirelessnetwork performance.

We now sample these higher level ideas throughthe following applications.

• Microwave-aware channel switching usingout-of-band sound sensing.

• MAC layer bit rate adaptation using mobilityand location classification.

• Augmenting end user’s communication expe-rience using out-of-band activity recognition.

The above applications belong to the <out-of-band, implicit> category, and are chosen to reflect

In band

Out of band

Implicit ExplicitRadioTelepathy[Mobicom 08]

Paradis[Mobicom 08]

AP Beacons,RTS/CTS

BreadCrumbs[Mobicom 08]

MWave InterferenceDetection

Accelerometer-basedRate Control

GPS-based RateControl [ICNP 08]

Blue-Fi[Mobisys 09]

Figure 2. Classification of information sourcesfrom the perspective of mobile communications.

potential types of sensor assistance. We discussthem next and present preliminary validations.

3. Applications

3.1. Microwave-Aware Channel Switching

Opportunity: A running microwave oven cansubstantially degrade the performance of 2.4GHz802.11 networks because they interfere withchannels 6 through 11. However, given that mostAP deployments select non-overlapping channelswithin the spatial vicinity, it may be beneficialfor a device to switch to an AP that uses a non-overlapping channel (say, 1). Even though the linkto the new AP could have a lower signal strength(RSSI), the switch may still be worthwhile. Oncethe microwave turns off, the device could switchback to channel 6 or 11.

Context-change Detection: As mentioned ear-lier, detecting the microwave’s signature, in-band,may not be simple. Ongoing 802.11 packets fromhidden terminals, variable packet lengths, chan-nel noise, and non-802.11 interferences fromBluetooth/cordless-phones systems are likely tocomplicate diagnosis. These stochastic effectsmake context-assessment inherently ambiguous,resulting in false alarms. Even if the microwaveis somehow detected, the subsequent difficultyarises in knowing the optimal time to switch backto the original channel. The obvious techniquewould be to periodically probe for improved condi-tions on the original, microwave-affected channel.However, probing entails disassociation from thecurrent AP and re-association to the original AP.This is a costly operation and can be prohibitiveif triggered frequently. If infrequently triggered,the device may not realize that the microwave hasstopped, and will unnecessarily remain on the sub-optimal channel for the long probing period.

A sound sensor can provide out-of-band informa-tion regarding the presence of a nearby microwavein use. The characteristic background “hum” or thefamiliar microwave beeps can provide an acousticsignature. Figure 3 shows this signature on thefrequency domain. The microwave sounds wererecorded using a Nokia N95 phone (at distancesof 0, 1, 3, 5, and 15 meters) and subjected throughsimple analysis. Across all tests, we found reliabledetection accuracy, with false negatives/falsepositives at 1.5% and 4.6%, respectively. Webelieve this could enable the desired environment-aware agility in future mobile devices.

0 500 1000 1500 2000 2500 3000 3500 40000

0.001

0.002

0.003

0.004

0.005

0.006

0.007

0.008

0.009

0.01

Frequency (Hz)

Ampli

tude

MicrowaveDetection

Figure 3. Microwave oven’s acoustic signature onthe frequency domain

Throughput Improvement: Given that sound-based channel switching is feasible, we testedpotential performance gains from avoiding mi-crowave interference. Using iperf, we tested TCPthroughput across the commonly-used channels 1,6, and 11. We used three microwaves of differingage and vendor, operating at center frequency of2.45 GHz, between 802.11b/g channels 9 and 10.The microwave was placed at varying distancesfrom the TCP receiver, which was then switched todifferent channels. To maximize consistency acrosschannels, we conducted all experiments at night,minimizing network contention. Figure 4(a, b, c)show the impact of microwaves (MW) on through-put, when the laptops were switched to channels 1,6, and 11, respectively. For a switch from channel11 to 1, we find an average improvement of 83%and 87% under interference from MW1 and MW2,respectively. A switch from channel 6 to 1 offers87% and 75% under MW1 and MW2. Results fromMW3 are similar and omitted for visual clarity.We believe these gains are substantial, especiallyconsidering that only a simple channel change issufficient to realize them.

3.2. Sensor-aware Rate Control

Opportunity: Wireless rate adaptation is achallenging problem. Among other reasons, thedifficulty in estimating channel fluctuations is themost prominent one. The problem is exacerbatedbecause the nature of the fluctuations vary underdifferent mobility regimes. Research has shownthat the effects of path loss dominate in a vehicularnetwork scenario, while multipath effects arestrong in indoor low-mobility scenarios [11]. One-size-fits-all rate control protocols are difficult todesign, and hence, any given protocol is best suitedfor a subset of the mobility regimes. Specifically,SNR-based rate control algorithms are shownto perform well in urban outdoor environments[4]. In contrast, packet error-based schemeslike SampleRate are suitable for static indoorenvironments [12]. A simple way to distinguishindoor/outdoor locations is to obtain the lightsensor readings on the phone and classify thembased on time of the day. We also suggest thepossibility of identifying the mobility regime ofthe user (static, walking, vehicular) through on-phone accelerometers. This information could bevaluable in appropriately multiplexing betweenrate control algorithms (when feasible), or simplyoptimizing the given protocol for that condition.

Context-change Detection: To verify the pos-sibility of detecting the user’s mobility regime,we drove a sedan in a 25-mile loop at an av-erage speed of 28 MPH. The route consisted ofa representative mix of both well and poorly-paved roads. Three Nokia N95s continually recoredaccelerometer readings. Fig. 5 shows the standarddeviation of acceleration across a portion of onetrace. Clear patterns of peaks and valleys emerge,reflecting movement and stillness, respectively. Wealso tested the accelerometer signature when auser walked with the phone in her pocket. Asmall portion of the trace shows distinct rhythmicpatterns (Fig. 5). Such pattern [8] has beenattempted before but we only show these resultsto illustrate the opportunity. We believe that thiscan provide the out-of-band information necessaryfor informed rate control decisions.

3.3. Activity Assisted Communications

Taking a broad view of mobile communication,we consider the human element in maintaining

0 5 10 15 20 25 300

0.2

0.4

0.6

0.8

1

Throughput (Mbps)

Empi

rical

CD

F

Channel 1

No MWMW 1MW 2

0 5 10 15 20 25 300

0.2

0.4

0.6

0.8

1

Throughput (Mbps)

Empi

rical

CD

F

Channel 6

No MWMW 1MW 2

0 5 10 15 20 25 300

0.2

0.4

0.6

0.8

1

Throughput (Mbps)

Empi

rical

CD

F

Channel 11

No MWMW 1MW 2

Figure 4. Throughput comparison across 802.11b/g channels 1, 6, 11. The impact of microwaveinterference is substantially less on channel 1.

0

50

100

150

200

200 201 202 203 204 205 206 207

Stan

dard

dev

iatio

n

Time in sec

ZYX

Figure 5. An acceleration profile distinguishingvehicular motion from stopped periods; Clearlydistinguishable walking patterns

a wireless channel. Ultimately, end-user experi-ence is the metric-of-interest. Pervasive activityrecognition might improve the convenience ofwireless communications. Suppose a phone callis inopportunely received while driving. Insteadof ringing, the agile device might instead confirmwith the caller first, “The person you’ve dialedis driving. Do you wish to continue?” Moreover,if the caller declines, the question has provideda helpful cue as to how long the caller shouldwait before trying later. We implemented thissimple application on a Google Nexus One Phone.If the callee is in a moving car and the callwas not answered, our preliminary applicationsenses the mobility of the user and sends a textmessage to the caller. Clearly, many situations canbenefit from out-of-band contexts (e.g., watchinga movie, attending a seminar, jogging). Officephone systems provide a rudimentary attempt,but require human participation to setup (e.g.,“in a meeting,” “on vacation”). Advanced activitysensing combined with pervasive localization(e.g., [8] and [1]) can provide a deeper contextawareness that would be difficult to extract in-band. Sharing this context with the callee helps tooptimize the experience of the communication.

Of course, there are privacy implications for au-tomatically sharing contexts between users. How-

ever, as has been already shown in a number ofexisting systems (e.g., Facebook, Google Latitude),straightforward configurability can balance utilitywith risks. We believe that the integration of hu-man activity recognition and direct feedback canprovide useful contexts, improving the end-userexperience.

4. Related Work

Several works use sensors to characterize theambience [1], [7]. We consider sensor assistanceto infer the RF environment, allowing optimizedwireless performance. In-band schemes, such aschannel hopping [6], have been proposed to avoidnon-compliant interference (e.g., from cordlessphones, microwaves). Our work is complemen-tary. Inferred context from out-of-band channelscan enhance known remedies. Using sensor in-formation for improving wireless performance isrelatively unexplored. Context-aware rate controluses GPS location and history to infer pathloss,thereby adjusting wireless bitrate [11]. This isan explicit out-of-band inference mechanism. Weare not aware of any work utilizing out-of-bandimplicit channels for improving wireless perfor-mance. However, [9] has considered the use ofcontext-sensing in optimizing the energy use of amobile device, switching between WiFi and GSMconnectivity as appropriate.

5. Limitations and Discussion

Out-of-band context may not always be dis-cernible and accurate, and may incur additionalcost and latency. This section discusses some guide-lines to employ sensor assisted communication.

Out-of-band information should supplement,not supplant, in-band information. While diverse

sensors for sound, light, speed, etc. provide out-of-band information in multiple dimensions, itmay still not be adequate to discern context insome environments. For example, white spacenetworking [2] permits secondary users to reusea spectrum, provided they do not interfere withprimary users, say microphones. We can facilitatesuch an opportunistic spectral reuse if the presenceof microphones can be detected through soundsensors. However, inaudible microphones can stillinterfere with secondary users. Moreover, whenout-of-band information is available, it should becoupled with in-band information to determinethe appropriate course of action. In other words,out-of-band information should be treated as ahelpful hint to address a problem rather than acomplete solution.

Out-of-band information should provide properand timely context. One could argue that theinformation obtained from out-of-band channelsmust have high fidelity. Otherwise, it could causethe protocol operations to be tuned for an incorrectcontext. Discerning the context in a timely manneris also necessary to respond to frequent contextualtransitions. Our case studies have shown that wecan identify an active microwave and a movingcar with reasonable accuracy and latency. Notethat even when the out-of-band information is notquite precise, it may still provide useful hints forprotocol adaptation.

Overhead of out-of-band information shouldideally be minimal. The ability to discernthe context of communication with out-of-bandinformation does not warrant its use in allsituations. Even if the information is helpful, thecost of obtaining it in the out-of-band channel mustbe less than the additional cost of obtaining thesame information in-band. But in many instances,the sensory information may be available at noadditional cost since the sensors are typicallyalways on to serve several other applications.Only when the desired sensory information is notimmediately available, we need to assess the prosand cons of activating sensors. We believe thatthis cost can often be amortized over many otherapplications that benefit from context awareness.

6. Conclusion

Mobile devices must continually cope with chal-lenging and diverse operating environments. With

their integration of sensing, computation, andwireless connectivity, modern mobile devices areuniquely positioned to characterize their surround-ings. We believe that sensing can provide a neces-sary contextual awareness, allowing these devicesto become truly agile. Ultimately, wireless systemsmay be made more robust to their environments,yielding an enhanced end-user experience.

References

[1] M. Azizyan, I. Constandache, and R. R. Choudhury.Surroundsense: Mobile phone localization via am-bience fingerprinting. In MOBICOM, 2009.

[2] P. Bahl, R. Chandra, T. Moscibroda, R. Murty, andM. Welsh. White space networking with wi-fi likeconnectivity. In SIGCOMM, 2009.

[3] V. Brik, S. Banerjee, M. Gruteser, and S. Oh. Wire-less device identification with radiometric signa-tures. In Proc. ACM Mobicom, 2008.

[4] J. Camp and E. Knightly. Modulation Rate Adapta-tion in Urban and Vehicular Environments: Cross-layer Implementation and Experimental Evalua-tion. In MobiCom, 2008.

[5] G. Chen and D. Kotz. A survey of context-awaremobile computing research. Dartmouth ComputerScience Technical Report TR2000-381, 2000.

[6] R. Gummadi, D. Wetherall, B. Greenstein, andS. Seshan. Understanding and mitigating theimpact of rf interference on 802.11 networks. InProc. ACM Sigcomm, 2007.

[7] H. Lu, W. Pan, N. D. Lane, T. Choudhury, and A. T.Campbell. SoundSense: Scalable Sound Sensingfor People-Centric Sensing Applications on MobilePhones. In Mobisys, 2009.

[8] E. Miluzzo, N. Lane, K. Fodor, R. Peterson, H. Lu,M. Musolesi, S. Eisenman, X. Zheng, and A. Camp-bell. Sensing meets mobile social networks:The design, implementation and evaluation of thecenceme application. In SENSYS, 2008.

[9] A. Rahmati and L. Zhong. Context-for-wireless:Context-sensitive energy-efficient wireless datatransfer. In MOBISYS, 2007.

[10] S. Rayanchu, A. Mishra, D. Agrawal, S. Saha, andS. Banerjee. Diagnosing Wireless Packet Losses in802.11: Separating Collision from Weak Signal. InINFOCOM, 2008.

[11] P. Shankar, T. Nadeem, J. Rosca, and L. Iftode.CARS: Context Aware Rate Selection for VehicularNetworks. In ICNP, 2008.

[12] M. Vutukuru, H. Balakrishnan, and K. Jamieson.Cross-layer wireless bit rate adaptation. In ACMSIGCOMM, 2009.