This paper is included in the Proceedings of the 15th USENIX Symposium on Networked

Systems Design and Implementation (NSDI ’18).April 9–11, 2018 • Renton, WA, USA

ISBN 978-1-939133-01-4

Open access to the Proceedings of the 15th USENIX Symposium on Networked

Systems Design and Implementation is sponsored by USENIX.

Towards Battery-Free HD Video StreamingSaman Naderiparizi, Mehrdad Hessar, Vamsi Talla, Shyamnath Gollakota,

and Joshua R Smith, University of Washington

https://www.usenix.org/conference/nsdi18/presentation/naderiparizi

Towards Battery-Free HD Video StreamingSaman Naderiparizi, Mehrdad Hessar, Vamsi Talla, Shyamnath Gollakota and Joshua R. Smith

University of Washington

Abstract – Video streaming has traditionally been con-sidered an extremely power-hungry operation. Existingapproaches optimize the camera and communication mod-ules individually to minimize their power consumption.However, designing a video streaming device requirespower-consuming hardware components and computa-tionally intensive video codec algorithms that interfacethe camera and the communication modules. For exam-ple, monochrome HD video streaming at 60 fps requiresan ADC operating at a sampling rate of 55.3 MHz anda video codec that can handle uncompressed data beinggenerated at 442 Mbps.

We present a novel architecture that enables HD videostreaming from a low-power, wearable camera to a nearbymobile device. To achieve this, we present an “analog”video backscatter technique that feeds analog pixels fromthe photo-diodes directly to the backscatter hardware,thereby eliminating power-consuming hardware compo-nents, such as ADCs and codecs. To evaluate our design,we simulate an ASIC, which achieves 60 fps 720p and1080p HD video streaming for 321 µW and 806 µW, re-spectively. This translates to 1000x to 10,000x lowerpower than it used for existing digital video streamingapproaches. Our empirical results also show that we canharvest sufficient energy to enable battery-free 30 fps1080p video streaming at up to 8 feet. Finally, we designand implement a proof-of-concept prototype with off-the-shelf hardware components that successfully backscatters720p HD video at 10 fps up to 16 feet.

1 IntroductionThere has been recent interest in wearable cameras likeSnap Spectacles [16] for applications ranging from life-casting, video blogging and live streaming concerts, polit-ical events and even surgeries [18]. Unlike smartphones,these wearable cameras have a spectacle form-factor andhence must be both ultra-lightweight and cause no over-heating during continuous operation. This has resulted

Figure 1: Target application. Our ultra-low-power HD streamingarchitecture targets wearable cameras. We achieve this by performinganalog video backscatter from the wearable camera to a nearby mobiledevice (e.g., smartphone).

in a trade-off between the usability of the device and itsstreaming abilities since higher resolution video stream-ing requires a bigger (and heavier) battery as well aspower-consuming communication and processing units.For example, Snap Spectacle, while lightweight and us-able, cannot stream live video [16] and can record onlyup to one hundred 10-second videos (effectively less than20 minutes) [16, 13] on a single charge.

In this paper, we ask the following question: Can wedesign a low-power camera that can perform HD videostreaming to a nearby mobile device such as a smart-phone? A positive answer would enable a wearable cam-era that is lightweight, streams high quality video, andis safe and comfortable to wear. Specifically, reducingpower consumption would reduce battery size, which inturn addresses the key challenges of weight, battery lifeand overheating. Finally, since users typically carry mo-bile devices like smartphones that are comparatively notas weight and power constrained, they can relay the videofrom camera to the cloud infrastructure.

To understand this challenge, let us look at the dif-ferent components of a video-streaming device: image

USENIX Association 15th USENIX Symposium on Networked Systems Design and Implementation 233

(a) Conventional camera design

(b) Our camera design approach

Figure 2: The amplifier, AGC, ADC and compression module con-sume orders of magnitude more power than is available on a low-powerdevice. In our design, these power hungry modules have been delegatedto the reader, eliminating their power consumption overhead from thewireless camera.

sensor, video compression and communication. Imagesensors have an optical lens, an array of photo-diodesconnected to amplifiers, and an ADC to translate analogpixels into digital values. A video codec then performscompression in the digital domain to produce compressedvideo, which is then transmitted on the wireless medium.Existing approaches optimize the camera and communi-cation modules individually to minimize power consump-tion. However, designing an efficient video streamingdevice requires power-consuming hardware componentsand video codec algorithms that interface to the cameraand communication modules. Specifically, optical lensand photo-diode arrays can be designed to consume as lit-tle as 1.2 µW [22]. Similarly, recent work on backscattercan significantly lower the power consumption of commu-nication to a few microwatts [26, 23] using custom ICs.Interfacing camera hardware with backscatter, however,requires ADCs and video codecs that significantly add topower consumption.

Table 1 shows the sampling rate and data rate require-ments for the ADC and video codec, respectively. HDvideo streaming requires an ADC operating at a highsampling rate of more than at least 10 MHz. While theanalog community has reduced ADC power consumptionat much lower sampling rates [46, 20], state-of-the-artADCs in the research community consume at least a fewmilliwatts at the high sampling rates [34]. Additionally,the high data rate requires the oscillator and video codecto run at high clock frequencies, which proportionallyincreases power consumption. Specifically, video codecsat these data rates consume 100s of milliwatts to a fewwatts of power [2].

We present a novel architecture that enables videostreaming on low-power devices. Instead of indepen-dently optimizing the imaging and the communicationmodules, we jointly design these components to signifi-cantly reduce system power consumption. Our architec-

Figure 3: Sample HD video frame streamed with our analog videobackscatter design. Our prototype was placed 4 feet from the reader.

ture, shown in Fig. 2b, takes its inspiration from the GreatSeal Bug [5], which uses changes in a flexible metallicmembrane to reflect sound waves. Building on this idea,we create the first “analog” video backscatter system. Ata high level, we feed analog pixels from the photo-diodesdirectly to the backscatter antenna; we thus eliminatepower-hungry ADCs, amplifiers, AGCs and codecs.

Our intuition for an analog video backscatter approachis to shift most of the power-hungry, analog-to-digitalconversion operations to the reader. Because analog sig-nals are more susceptible to noise than digital ones, wesplit the ADC conversion process into two phases, oneperformed at the video camera, and one accomplishedat the reader. At the video camera, we convert analogpixel voltage into a pulse that is discrete in amplitude butcontinuous in time. This signal is sent via backscatterfrom the camera to the reader. Avoiding the amplituderepresentation in the wireless link provides better noiseimmunity. The reader measures the continuous lengthpulse it receives to produce a binary digital value. Philo-sophically, the continuous-time, discrete amplitude pulserepresentation used in the backscatter link resembles thatused in an extremely power-efficient biological nervoussystem, which encodes information in spikes that do notvary in amplitude but are continuous in time [45].

Specifically, our design synthesizes three key tech-niques. First, we show how to interface pixels directlyto the backscatter hardware without using ADCs. To do,this we transform analog pixel values into different pulsewidths using a passive ramp circuit and map these pulsesinto pixels at the reader. Second, we achieve intra-framecompression by leveraging the redundancy inherent intypical images. Intuitively, the signal bandwidth is pro-portional to the rate of change across adjacent pixels;since videos tend to be redundant, the bandwidth of theanalog signal is inversely proportional to the redundancyin the frame. Thus, by transmitting pixels consecutively,we can implicitly perform compression and significantlyreduce wireless bandwidth. Finally, to achieve inter-framecompression, we design a distributed algorithm that re-duces the data the camera transmits while delegating most

234 15th USENIX Symposium on Networked Systems Design and Implementation USENIX Association

Table 1: Raw digital video sampling and bitrate requirementFrame Rate: 60 fps Frame Rate: 30 fps Frame Rate: 10 fps

Video QualitySamplingRate(MHz)

Data Rate(Mbps)

SamplingRate(MHz)

Data Rate(Mbps)

SamplingRate(MHz)

Data Rate(Mbps)

1080p (1920x1080) 124.4 995.3 62.2 497.7 20.7 165.9720p (1280x720) 55.3 442.4 18.4 221.2 9.2 73.7480p (640x480) 18.4 147.4 9.2 73.7 3.1 24.58360p (480x360) 10.4 82.9 5.2 41.5 1.7 13.8

inter-frame compression functionality to the reader. At ahigh level, the camera performs averaging over blocks ofnearby pixels in the analog domain and transmits these av-eraged values using our backscatter hardware. The readercompares these averages with those from the previousframe and requests only the blocks that have seen a sig-nificant change in the average pixel value, thus reducingtransmission between subsequent video frames.

We implement a proof-of-concept prototype of our ana-log backscatter design on an ultra-low-power FPGA plat-form and a custom implementation of the backscattermodule. Because no HD camera currently provides ac-cess to its raw pixel voltages, we connect the output ofa DAC converter to our backscatter hardware to emulatethe analog camera and stream raw analog video voltagesto our backscatter prototype. Fig. 3 shows a sample framefrom an HD video streamed with our backscatter camera.More specifically, our findings are:

• We stream 720p HD video at 10 frames per secondup to 16 feet from the reader. The Effective Number ofBits (ENOB) received for each pixel at distances belowsix feet exceeds 7 bits. For all practical purposes, theseresults are identical to the quality of the source HD video.• Our inter and intra-frame compression algorithms re-duces total bandwidth requirements by up to two orders ofmagnitude compared to raw video. For example, for 720pHD video at 10 fps, our design uses a wireless bandwidthof only 0.98 MHz and 2.8 MHz in an average-case andworst-case scenario video, respectively.

We design and simulate an ASIC implementation forour system, taking into account power consumption forthe pixel array. Our results show that power consumptionfor video streaming at 720p HD is 321 µW and 252 µWfor 60 and 30 fps, respectively. Power consumption at1080p full-HD is 806 µW and 561 µW at 60 and 30 fps, re-spectively. We run experiments with RF power harvestingfrom the reader which shows that we can support 1080pfull-HD video streaming at 30 fps up to distances of 8 feetfrom the reader. Our results demonstrate that we caneliminate batteries and achieve HD video streaming onbattery-free devices using our analog video backscatterapproach.

2 A Case for Our Architecture

Fig. 2a shows the architecture of a traditional wirelesscamera. Photodiodes’ output is first amplified by a lownoise amplifier (LNA) with automatic gain control (AGC).The AGC adjusts the amplifier gain to ensure that theoutput falls within the dynamic range of the analog todigital converter (ADC). Next, the ADC converts theanalog voltage into discrete digital values. The videocodec then compresses these digital values, which aretransmitted on the wireless communication link.

Unfortunately, this architecture cannot be translatedto an ultra-low-power device. Although camera sensorsconsisting of an array of photodiodes have been shown tooperate on as low as 1.2 µW of power [22] at 128×128resolution, amplifiers, AGC, ADC and the compressionblock require orders of magnitude higher power. Fur-thermore, power consumption increases as we scale thecamera’s resolution and/or frame rate. A low resolution144p video recorded at 10 fps requires an ADC samplingat 368 KSPS to generate uncompressed data at 2.95 Mbps.With the latest advancements in ADCs [33] and backscat-ter communication [26], uncompressed data can be dig-itally recorded using low-power ADCs and transmittedusing digital backscatter while consuming below 100 µWof power, which is within the power budget of energyharvesting platforms. However, as we scale the resolu-tion to HD quality and higher, such as 1080p and 1440p,the ADC sampling rate increases to 10-100 MHz, anduncompressed data is generated at 100 Mbps to Gbps.ADCs operating at such high sampling rates consumeat least a few mW [34]. Further, a compression blockthat operates in real time on 100 Mbps to 1 Gbps of un-compressed video consumes up to one Watt of power [2].This power budget exceeds by orders of magnitude thatavailable on harvesting platforms. Thus, while existingarchitectures might operate on harvested power for low-resolution video by leveraging recent advancements inlow-power ADCs and digital backscatter, as we scale theresolution and/or the frame rate of the wireless camera,these architectures’ use of ADCs and compression blockdrives up power consumption to levels beyond the realmof battery-free devices.

USENIX Association 15th USENIX Symposium on Networked Systems Design and Implementation 235

Figure 4: Architecture of the PWM converter.

3 System DesignIn the rest of this section, we first describe our analogvideo transmission scheme followed by the intra-framecompression and finally the interactive distributed inter-frame compression technique.

3.1 Analog Video BackscatterProblem. At a high level, our design eliminates the ADCon the camera to significantly reduce power consumption.This limits us to working with analog voltage output fromthe photodiodes. A naive approach would leverage theexisting analog backscatter technique used in wirelessmicrophones [48] to implement an analog backscattercamera. In an analog backscatter system, sensor outputdirectly controls the gate of a field-effect transistor (FET)connected to an antenna. As the output voltage of thesensor varies, it changes the impedance of the FET whichamplitude modulates the RF signal backscattered by theantenna. The reader decodes sensor information by de-modulating the amplitude-modulated backscattered RFsignal. However, this approach cannot be translated toa camera sensor. Photodiode output has a very limiteddynamic range (less than 100 mV under indoor lightingconditions). These small voltage changes map to a verysmall subset of radar cross-sections at the antenna [19].As a result, the antenna backscatters a very weak signal.Since wireless channel and receivers add noise, this ap-proach results in a low SNR signal at the reader, whichlimits the system to both poor signal quality and limitedoperating range. One can potentially surmount this con-straint by introducing a power-hungry amplifier and AGC,but this would negate the power-saving potential of analogbackscatter.

Our solution. Instead of using amplitude modulation,which is typical in existing backscatter systems [52, 26],we use pulse width modulation (PWM) [24] to convertthe camera sensor’s analog output into the digital domain.At a high level, PWM modulation converts analog inputvoltage to different pulse widths. Specifically, the outputof PWM modulation is a square wave, where the duty cy-cle of the output square wave is proportional to the analogvoltage of the input signal. PWM signal harmonics do notencode any sensor information that is not already presentin the fundamental. The harmonics can be considered

a redundant representation of the fundamental. Whilethey may add robustness, they contain no additional in-formation. Thus, without causing any information loss,higher order components can be eliminated via harmoniccancellation techniques introduced in prior work [47].

As we show next, a PWM converter can implementedwith passive RC components and a comparator, therebyconsuming very low power. Fig. 4 shows the PWM con-verter architecture. The input is a square wave operating atfrequency f , which is determined by the camera’s framerate and resolution. First, the square wave is low-pass fil-tered by an RC network to generate a triangular waveform,as shown in the figure. This waveform is then comparedto the pixel value using a comparator. The comparatoroutputs a zero when the triangular signal is less than thepixel value and a one otherwise. Thus, the pulse widthgenerated changes with the pixel value: lower pixel val-ues have a larger pulse duration, while higher pixel valueshave a smaller pulse duration. We choose the minimumand maximum voltages for the triangular signal to ensurethat the camera’s pixel output is always within this limits.

The final component in our hardware design is sub-carrier modulation, which addresses the problem of self-interference. Specifically, in addition to receiving thebackscatter signal, the receiver also receives a strong in-terference from the transmitter. Since the sensor data iscentered at the carrier frequency, the receiver cannot de-code the backscattered data in the presence of a strong,in-band interferer. Existing backscatter communicationsystems, such as RFID and Passive Wi-Fi, address thisproblem using sub-carrier modulation. These systems usea sub-carrier to shift the signal from the carrier to a fre-quency offset ∆ f from the carrier frequency. The receivercan then decode the backscattered signal by filtering outof band interference. Another consideration in choosingsub-carrier frequency is to avoid aliasing; sub-carrier fre-quency should not be smaller than the effective bandwidthof the analog signal.

Our PWM-based design integrates subcarrier modula-tion. We implement this modulation with a simple XORgate. The sub-carrier can be approximated by a squarewave operating at ∆ f frequency. We input sub-carrier andPWM output to an XOR gate to up-convert the PWM sig-nal to a frequency offset ∆ f . Sub-carrier modulation ad-dresses the problem of self-interference at the reader, andas a result, the PWM backscattering wireless camera cannow operate at a high SNR and achieve broad operatingranges. We show in §5.1 that our PWM backscatter wire-less camera can operate at up to 16 feet for a 2.76 MHzbandwidth 10 fps monochrome video signal in HD resolu-tion. We also show in §6 that our camera system can workat up to 150 feet for a 50 KHz video signal in 112×112resolution, over a 4× improvement relative to the 3 kHzbandwidth analog backscatter wireless microphone [50].

236 15th USENIX Symposium on Networked Systems Design and Implementation USENIX Association

Table 2: Intra-frame compression for average/worst-case scenarios across 100 videos.

Frame Rate: 60 fps Frame Rate: 30 fps Frame Rate: 10 fps

VideoResolution

Raw DataRate(Mbps)

Analog BW(MHz)

Raw DataRate(Mbps)

Analog BW(MHz)

Raw DataRate(Mbps)

Analog BW(MHz)

1080p 995.3 10.8/30.5 497.6 5.6/18.7 165.8 1.9/9.3720p 442.3 6.3/15.6 221.1 8/3.1 73.7 0.98/2.8480p 147.4 3/6.7 73.7 1.6/3.3 24.5 0.53/1.4360p 82.9 1.9/3.9 41.4 0.94/2.3 13.8 0.32/0.78

3.2 Intra-Frame CompressionThere is significant redundancy in the pixel values of eachframe of uncompressed video. Redundancy occurs be-cause natural video frames usually include objects largerthan a single pixel, which means that the colors of nearbypixels are highly correlated. At the boundaries of objects(edges), larger pixel variations can occur; conversely, inthe interior of an object, the amount of pixel variation isconsiderably less than the theoretical maximum. The netresult of pixel correlations is a reduction in the informa-tion needed to represent natural images to levels far belowthe worst-case maximum. Traditional digital systems usea video codec to reduce pixel value redundancy. Specifi-cally, the ADC first digitizes the camera’s raw output atthe Nyquist rate determined by the resolution and framerate. The raw digital data stream is then fed to the videocodec, which implements compression algorithms.

In the absence of the ADC, our wireless camera trans-mits analog video directly. However, we note that thebandwidth of any analog signal is a function of the newinformation contained in the signal. Inspired by analogTV broadcast, which transmits pixel information in araster scan (left to right), we introduce and implementa zig-zag pixel scanning technique: pixels are scannedfrom left to right in odd rows and from right to left ineven rows. The intuition here is that neighboring pixelshave less variation, and the resulting signal would thusoccupy less bandwidth.

We evaluate how well zig-zag transmission performs interms of bandwidth reduction. We download one hundred60 fps Full-HD (1080p) videos from [17] to use as ourbaseline. These videos range from slow to fast movingand contain movie action stunts, running animals, racing,etc. To create video baselines at different resolutions andframe rates, we resize and subsample these Full-HD videostreams. For each of the resulting video resolutions andframe rates, we create an analog video by zig-zaggingthe pixels as described above. We then apply low-passfilters with different bandwidths on this analog signal andreport the minimum bandwidth at which the PSNR of theresulting signal exceeds 30 dB. The bandwidth require-ment reported in Table 2 shows the average/worst-casescenario, i.e., the bandwidth that ensures the recovery of

average/worst-case videos with minimal quality degrada-tion. We outline the uncompressed digital data rate forreference. Compared to the digital uncompressed datasampled at the Nyquist rate, the analog signal occupies17.7–32.6x less bandwidth for the worst-case and 43–92xfor the average-case, demonstrating our intra-frame com-pression technique’s capability. Fig. 5a shows the CDFof effective bandwidth for our 30 fps 720p video datasetand demonstrates that an average-case 30 fps 720p videoacheives up to a 71× improvement compared to raw digi-tal video transmission.

Finally, we note that, compared to raster, a zig-zagscan faces less discontinuity in pixel values: instead ofjumping from the last pixel in a row to the first pixel in thenext row, it continues at the same column in the next row,thereby taking greater advantage of the video’s inherentredundancy. This further lowers bandwidth utilizationin the wireless medium. As an example, on average,zig-zag pixel scanning occupies ∼120KHz and ∼60KHzless bandwidth than raster scanning in a 60 fps and 30 fps720p video stream, respectively. Fig. 5b shows the CDF ofbandwidth improvement for zig-zag scanning over rasterin our one hundred 30 fps 720p videos dataset. The plotmakes clear that use of zig-zag pixel scanning providesgreater bandwidth efficiency than its raster counterpart.

3.3 Distributed Inter-Frame Compression

In addition to the redundancy within a single frame, rawvideo output also has significant redundancy between con-secutive frames. Existing digital architectures use a videocodec to compress the raw video prior to wireless trans-mission to remove inter-frame redundancy and reducepower and bandwidth. Our approach to address this chal-lenge is to once again leverage the reader. Like backscat-ter communication, where we move power-hungry com-ponents (such as ADCs) to the reader, we also movecompression functionality to the reader. Specifically, ourdesign distributes the compression algorithm betweenthe reader and the camera. We delegate power-hungrycomputation to the reader and leverage the fact that ourcamera system can transmit super-pixels. A super-pixelvalue is the average of a set of adjacent pixel values. A

USENIX Association 15th USENIX Symposium on Networked Systems Design and Implementation 237

0 0.2 0.4 0.6 0.8

1

0 2 4 6 8

CD

F

Effective Bandwidth (MHz)

(a) Effective bandwidth

0 0.2 0.4 0.6 0.8

1

0 200 400 600

CD

F

Bandwidth (KHz)

(b) Zig-zag vs. raster improvement

Figure 5: CDF of effective bandwidth for zig-zag pixel scanning(a), and improvement it provides over raster scanning (b).

camera frame consists of an N ×N array of pixels, whichcan be divided into smaller sections of n× n pixels. Asuper-pixel corresponding to each section is the averageof all the pixels in the n× n frame section. The camerasensor, a photodiode, outputs a current proportional tothe intensity of light. The camera uses a buffer stage atthe output to convert the current into an output voltage.To compute the super-pixel, the camera first combinesthe current from the set of pixels and then converts thecombined current into a voltage output, all in the analogdomain. We note that the averaging of close-by pixels issupported by commercial cameras including the CentEyeStonyman camera [4].

Instead of transmitting the entire N ×N pixel frame,the camera transmits a lower resolution frame consistingof n×n sized super-pixels (the average value of the pixelsin the n×n block), called the low-resolution frame or Lframe. Doing so reduces the data transmitted by the cam-era by a factor of N2

n2 . The reader performs computationon L frames and implements a change-driven compres-sion technique. At a high level, the reader compares inreal time the incoming L frame with the previous L frame.If a super-pixel value differs by more than a predeter-mined threshold between frames, then the super-pixel hassufficiently changed, and the reader asks the camera totransmit all pixels corresponding to it. If the differencedoes not cross the threshold, then the reader uses the pixelvalues corresponding to the previous reconstructed frameto synthesize the new frame and does not request newpixel values. We call the frame that contains the pixelvalues corresponding to the sufficiently changed super-pixels, the super-pixel frame or S frame. In addition totransmitting the S and L frames, the camera periodicallytransmits uncompressed frames (I) to correct for potentialartifacts and errors that the compression process may haveaccumulated.

In streaming camera applications, the communicationoverhead of the reader requesting specific pixel values isminimal and is implemented using a downlink similar toprior 100-500 kbps designs [26, 44], where the receiverat the camera uses a simple envelope detector to decodethe amplitude modulated signal from the reader. We notethat prior designs can achieve Mbps downlink transmis-sions [42] as well as full-duplex backscatter [31], whichcan be used to remove the downlink as a bottleneck.

Figure 6: Distributed compression. The sequence of frames andpixels transmitted by the camera to the reader.

Fig. 6 shows the sequence of frames and pixels trans-mitted by our camera. Between two I frames, the cameratransmits M low-resolution L frames and M super-pixelS frames which contain pixel values corresponding tosuper-pixels whose value differences exceed the thresh-old between consecutive frames. The number of L andS frames (M) transmitted between consecutive I framestrades off between the overhead associated with transmis-sion of full resolution frames and the artifacts and errorsthe compression algorithm introduced. In our implemen-tation of 10 fps HD video streaming, we transmit an Iframe after a transmission of every 80 L and S frames.

4 ImplementationWe built our wireless camera using off-the-shelf compo-nents on a custom-designed Printed Circuit Board (PCB).We use the COTS prototype to evaluate the performanceof the wireless camera in various deployments. We thenpresent the application-specific integrated circuit (ASIC)design of the wireless camera that we used to quantify thepower consumption for a range of video resolutions andframe rates.

COTS implementation. Our wireless camera designeliminates the power-hungry ADCs and video codecsand consists only of the image sensor, PWM converter, adigital block for camera control and sub-carrier modula-tion, a backscatter switch and an antenna. We built twohardware prototypes, one for the high definition (HD) andanother for the low-resolution version of the camera.

We built the low-resolution wireless camera using the112× 112 grayscale random pixel access camera fromCentEye [4], which provides readout access to individualanalog pixels. We implement the digital control blockon a low-power Igloo Nano FPGA by Microsemi [7].The analog output of the image sensor is fed to thePWM converter built using passive RC components and aMaxim NCX2200 comparator [10]. We set R1 = 83KΩ,R2 = 213KΩ and C = 78pF in our PWM converter designto support video frame rates of up to 13 fps. PWM con-verter’s output acts as input to the FPGA. The FPGA per-forms sub-carrier modulation at 1.024 MHz using an XORgate and outputs the sub-carrier modulated PWM signalto the Analog Devices ADG919 switch which switches a2 dBi dipole antenna between open and short impedancestates. The FPGA injects frame and line synchronization

238 15th USENIX Symposium on Networked Systems Design and Implementation USENIX Association

Figure 7: HD wireless camera prototype. A laptop emulating ananalog output camera feeds raw pixel data into our backscatter prototype.Then the PWM encoded pixels are backscattered to the reader.

patterns into the frames data before backscattering. Weuse Barker codes [3] of length 11 and 13 for our frameand line synchronization patterns, respectively. Barkercodes have a high-autocorrelation property that helps thereader more efficiently detect the beginning of the framein the presence of noise.

We use Verilog to implement the digital state machinefor camera control and sub-carrier modulation. Verilogdesign can be easily translated into ASIC using industrystandard EDA tools. We can further reduce our system’spower consumption by using the distributed compressiontechnique. As described in §6, a camera deployed in a nor-mal lab space can achieve an additional compression ratioof around 30×, which proportionately reduces wirelesstransmissions.

To develop an HD-resolution wireless camera, we needaccess to the raw analog pixel outputs of an HD cam-era. Currently, no camera on the market provides thataccess. To circumvent this constraint, we download fromYouTube HD-resolution sample videos lasting 1 minuteeach. We output the recorded digital images using aUSB interface to an analog converter (DAC) to simu-late voltage levels corresponding to an HD quality imagesensor operating at 10 fps. Given the USB speeds, weachieve the maximum frame rate of 10 fps. We feed thevoltage output to our PWM converter. Fig. 7 shows thehigh-level block diagram of this implementation. Forthe high-resolution version of the wireless camera, weset R1 = 10KΩ, R2 = 100KΩ and C = 10pF and use anLMV7219 comparator by Texas Instruments [8] in ourPWM converter. The digital block and other system com-ponents were exactly the same as for the low-resolutionwireless camera described above except that sub-carrierfrequency is set to ∼10 MHz here to avoid aliasing.

ASIC Design. As noted, our design eliminates the power-hungry LNA, AGC and ADC at the wireless camera bydelegating them to the reader to reduce wireless cam-era power consumption by orders of magnitude. How-ever, since commercially available Stonyman cameras(like CentEye) and components (such as FPGA) are de-signed for flexibility and ease of prototyping and are notoptimized for power, our COTS implementation cannotachieve the full power savings from our design. There-fore, we analyze the power consumption of an application-specific integrated circuit (ASIC) implementation of our

Figure 8: Signal processing steps at the reader. Recov-ering video from a PWM backscatter signal.

design for a range of video resolutions and frame rates.An ASIC can integrate the image sensor, PWM converter,digital core, oscillator and backscatter modulator onto asmall silicon chip. We implement our design in a TSMC65 nm LP CMOS process.

We use Design Compiler by Synopsis [14] to synthe-size transistor level from the behavioral model of ourdigital core, which is written in Verilog. We custom-design the PWM converter, oscillator and backscattermodulator described in §3 in Cadence software and useindustry standard simulation tools to estimate power. Tosupport higher resolution and higher frame rate video, wesimply increase the operating frequency of the oscillator,PWM converter and digital core. As an example, 360p at60 fps requires a 10.4 MHz input clock, which consumesa total of 42.4 µW in the digital core, PWM converterand backscatter switch; a 1080p video at 60 fps requiresan ∼124.4 MHz input clock, which consumes 408 µW inthe digital core, PWM converter and backscatter switch.To eliminate aliasing in all cases, we choose a sub-carrierfrequency equal to the input clock of each scenario. Notethat sub-carrier frequency cannot be lower than the effec-tive bandwidth of the signal reported in Table 2.

We use existing designs to estimate the power consump-tion of the image sensor for different video resolutions.State-of-the-art image sensors consume 3.2pW/( f rame×pixel) [51] for the pixels-only image sensor, which re-sults in 33.2 µW for 360p resolution; this increases to398 µW for 1080p resolution video at 60fps. Table 3shows the power consumption of the ASIC version ofour wireless camera for different video resolution andframe rates. Note that these results show power consump-tion before inter-frame compression distributed across thereader and camera, which could further reduce wirelessbandwidth and power consumption.

Reader Implementation. We implement the reader onthe X-300 USRP software-defined radio platform by EttusResearch [15]. The reader uses a bi-static radar config-uration with two 6 dBi circularly polarized antennas [1].Its transmit antenna is connected to a UBX-160 daughter-board, which transmits a single-tone signal. USRP outputpower is set to 30 dBm using the RF5110 RF power am-plifier [11]. The receive antenna is connected to anotherUBX-160 daughter board configured as a receiver, whichdown-converts the PWM modulated backscattered RF sig-nal to baseband and samples it at 10 Msps. The digitalsamples are transmitted to the PC via Ethernet.

USENIX Association 15th USENIX Symposium on Networked Systems Design and Implementation 239

Table 3: Power consumptionFrame Rate: 60 fps Frame Rate: 30 fps Frame Rate: 10 fps

Video Quality Power (µW) Power (µW) Power (µW)1080p (1920x1080) 806.50 560.63 167.77

720p (1280x720) 320.94 252.10 78.31480p (640x480) 126.88 106.78 36.71360p (480x360) 75.63 65.68 25.11

Fig. 8 shows block diagram of the signal processingsteps required to recover the transmitted video. For exam-ple, for low-resolution video, the received data is centeredat an offset frequency of the 1.024 MHz; therefore, wefirst filter the received data using a 500 order bandpassfilter centered at 1.024 MHz. Then, we down-convert thesignal to baseband using a quadrature down-conversionmixer. Next, we correlate the received data with 13 and 11bit Barker codes to determine the frame sync and line sync.After locating frame and line sync pulses, we extract thetime periods corresponding to a row of PWM-modulatedpixel values and low-pass filter the signal with a 500 or-der filter to remove out of band noise. We divide the rowin evenly spaced time intervals that corresponded to thenumber of pixels in a single row of the image sensor. Werecover the pixel value by calculating the average voltageof the signal, which corresponds to the duty cycle of thePWM-modulated signal. We sequentially arrange recov-ered pixel values into rows and columns to create videoframes. Finally, we run a histogram equalization algo-rithm on the video to adjust frames intensity and enhanceoutput video contrast [21].

5 EvaluationWe now evaluate various aspects of our wireless camerasystem. We start by characterizing the received videoquality from the camera as a function of its distance tothe reader. Next, we evaluate the performance of ourwireless camera while it is worn by a user under differenthead positions and orientations. We then evaluate thepower available by harvesting energy from RF signalstransmitted by the reader to demonstrate the feasibility ofa battery-free wireless camera. Finally, we evaluate thedistributed interactive compression algorithm under twodifferent scenarios.

5.1 Operational RangeWe deploy our high-definition wireless camera device ina regular lab space. We use the USRP-based reader imple-mentation ( §4), which we set to transmit 23 dBm into a6 dBi patch antenna. This is well below the 30 dBm max-imum transmit power permitted by FCC in the 900 MHzISM band. We vary the distance between the reader and

0 1 2 3 4 5 6 7 8

4 6 8 10 12 14 16

EN

OB

Distance (in ft)

Figure 9: ENOB of the Received Video. ENOB of the receivedvideo versus distance of our wireless camera prototype from the reader.

the wireless camera prototype from 4 to 16 feet and con-figure the camera to repeatedly stream a 15-second videousing PWM backscatter communication. The 720p res-olution video is streamed at 10 fps, and the pixel valuesare encoded as 8-bit values in monochrome format. Werecord the wirelessly received video at the reader andmeasure the Signal to Noise Ratio (SNR). From that, wecalculate the Effective Number of Bits (ENOB) [27] atthe receiver.

We plot the ENOB of the video digitized by the readeras a function of distance between the reader and the wire-less camera in Fig. 9. The plot shows that up to 6 feetfrom the reader, we achieve an ENOB greater than 7,which indicates negligible degradation in quality of thevideo, streamed using PWM backscatter. As the distancebetween the reader and the camera increase the SNRdegrades, indicates a decrease in ENOB. Beyond the dis-tance of 16 feet, we stop reliably receiving video frames.A separation of 16 feet between the reader and wear-able camera is more than sufficient for wearable cameras,typically located a few feet away from readers such assmartphones. For reference, Fig. 10 shows frames froma 720p HD color video backscattered with our prototypeat different distances from the reader; this resulted indifferent ENOB values.

We conclude the following: our analog video backscat-ter approach is ideal for a wearable camera scenario sincethe video is streamed to a nearby mobile device such as asmartphone. In this scenario, the SNR is high; hence, qual-ity degradation due to an analog approach is not severe.Further, we gain significant power reduction benefits.

240 15th USENIX Symposium on Networked Systems Design and Implementation USENIX Association

(a) Original Image (b) 7 Bits

(c) 4 Bits (d) 3 Bits

Figure 10: Effective Number of Bits (ENOB) versus video quality.The frame corresponding to an ENOB of 3 (d) shows video qualitydegradation.

5.2 Effect of Head Pose and MotionWe ask a participant to wear the antenna of our wirelesscamera on his head and perform different head move-ments and poses while standing around five feet froma reader with fixed location on a table. The poses in-cluded standing still, moving head left and right, rotatinghead to the side, moving head up and down, and talking.Hence, our evaluation includes scenarios with relativemobility between the reader and camera; in fact, it alsoincludes cases where no line of sight communication ex-ists between the reader and camera. We next, evaluateour wireless camera for in-situ applications and assesshow movements and antenna contact with a body affectvideo quality. We record the streaming video with thereader and measure the SNR and ENOB as we did in §5.1.Fig. 11 plots the ENOB of the received video for fivedifferent poses and movements. This plot shows that wecan achieve sufficient ENOB while performing most ofthese gestures, resulting in a high-quality video comparedto the original source video.

0 1 2 3 4 5 6 7 8

Still Moving Left and Right

Rotation Talking Moving Up and Down

EN

OB

Figure 11: ENOB of received video under different head motions.

5.3 RF Power HarvestingNext, we evaluate the feasibility of developing a battery-free HD streaming camera that operates by harvestingRF signals transmitted by the reader. We build an RFharvester for the 900 MHz ISM band based on a state-of-

0

500

1000

1500

2000

2500

3000

4 6 8 10 12 14 16

Har

vest

ed P

ower

(in u

W)

Distance (in ft)

1080p @ 60fps

1080p @ 60fps 1080p @

30fps 720p @ 60fps

720p @ 30fps 360p @

30fps480p @ 10fps

Figure 12: Power harvesting. We plot the average power harvestedby the battery-free hardware over different distances from the reader.

the-art RF harvester design [43]. The harvester consistsof a 2 dBi dipole antenna and a rectifier that convertsincoming RF signals into low-voltage DC output. Thelow-voltage DC is amplified by a DC-DC converter togenerate the voltage levels to operate the image sensorand digital control block. We measure the power availableat the output of the DC-DC converter.

We configure the reader to transmit a single tone at30 dBm into a 6 dBi patch antenna and move the RFharvester away from the reader. Fig. 12 plots availablepower at the harvester as a function of distance. Basedon available power, we also plot in Fig. 12 the maximumresolution and frame rate of the video that could be trans-mitted by an RF-powered ASIC version of our wirelesscamera. At close distances of 4 and 6 feet, we see suffi-cient power available from RF energy harvesting to oper-ate the wireless camera at 60 fps 1080p resolution. As thedistance increases, available power reduces, which lowersresolution of video being continuously streamed from thewireless camera. At 16 feet, the wireless camera contin-uously streams video at 10 fps 480p resolution; beyondthis distance, the harvester does not provide sufficientpower to continuously power the wireless camera. Notethat Fig. 12 shows camera performance without using thedistributed inter-frame compression algorithm describedin §3.3. That algorithm, distributed across the wirelesscamera and reader, reduces camera transmissions, whichlowers power consumption and consequently increasesoperating distances.

5.4 Distributed Inter-Frame CompressionEvaluation

We consider two scenarios to evaluate our distributedinter-frame compression ( subsection 3.3). We analyzeHD video streamed from a fixed highway monitoring cam-era and from an action camera mounted on a user ridinga motorcycle [9]. We evaluate the trade-off between thecompression ratio and PSNR under both static and dy-namic video feeds using our design. We measure the PeakSignal to Noise Ratio (PSNR) for different compressionratios by varying the threshold at which we consider thesuper-pixel (20 by 20 pixels) to have significantly changed.

USENIX Association 15th USENIX Symposium on Networked Systems Design and Implementation 241

10

20

30

40

50

60

0 5 10 15 20 25 30 35 40 45 50

PS

NR

(dB

)

Compression Ratio

Moving Camera on Motorcycle

Static Camera

Figure 13: Evaluation of distributed inter-frame compression al-gorithm for HD video. The plots show PSNR at different compressionratio for two HD videos.

A higher threshold would result in higher compressionratios but at the cost of a degraded PSNR. Fig. 13 showsPSNR of the compressed video as a function of the com-pression ratio for our distributed inter-frame compressiontechnique. For static cameras, we achieve a compressionratio of at least 35× while maintaining a PSNR above30 dB. For dynamic videos recorded from a motorcycle,we achieve a compression ratio of 2× for a PSNR greaterthan 30 dB. This is expected since mobile scenarios sig-nificantly change majority of pixel values between frames,resulting in lower compression using our approach. Wecould address this by implementing a more complex com-pression algorithm at the reader to track moving objects inthe frame and request specific pixel locations, achievingcompression levels similar to video codecs. Implement-ing such complex compression algorithms, however, isbeyond the scope of this paper.

6 Low-Resolution Security CameraSo far, we have demonstrated how high-resolution videooffers a useful paradigm for a wearable camera system.However, various other applications, such as security sys-tems and smart homes, do not require high-resolutionvideo; lower resolution video would suffice for applica-tions such as face detection. Specifically, wireless cam-eras are increasingly popular for security and smart homeapplications. In contrast to wearable camera applications,these cameras require low-resolution but much longeroperating distances. To show that our design can extendto such applications, we evaluate the operation range at13 fps 112×112 resolution. Our IC design at this resolu-tion consume less than 20 µW without accounting for anydistributed inter-frame compression saving.

To evaluate the range, we use a 13 fps 112×112 reso-lution, gray-scale random pixel access camera from Cent-Eye [4] as our image sensor. The camera has a photodiodeimage sensor array, a trans-impedance buffer stage (toconvert photodiode current to voltage) and a low-noiseamplifier. It is extremely flexible, and the user can modifyvarious settings, such as gain of the amplifier stage and,if desired, completely bypass the amplifier. We use thisunique camera feature to prototype our wireless camera,

Figure 14: Prototype of our low-resolution, video streaminghome security camera. Image of the analog camera, FPGA digitalcore and pulse width modulated (PWM) backscatter, all implementedusing COTS components. The overall board measures 3.5 cm by 3.5 cmby 3.5 cm.

which directly transmits analog values from the image sen-sor (sans amplification) using PWM backscatter. Fig. 14shows photographs of our wireless camera prototype. Thecamera allows random access, i.e., any pixel on the cam-era can be accessed at random by setting the correspond-ing value in the row and column address registers. It canalso be configured to output a single pixel, two adjacentpixels, or a super-pixel with sizes ranging from 2×2 to7×7. We use the camera’s random access and super-pixelfunctionality to implement our distributed inter-framecompression algorithm. The power consumption of ourlow-resolution, off-the-shelf analog video streaming pro-totype is 2.36 mW. We emphasize that this off-the-shelfcamera is used only to demonstrate operational range; toachieve the tens of microwatts power budget, we need touse our ASIC design.

Deployment results. We deployed our wireless camerasystem in the parking lot of an apartment complex. Weuse the USRP-based reader implementation set to transmit30 dBm into a 6 dBi patch antenna. We vary the distancebetween the wireless camera prototype, and, the readerand at each separation, we stream 20 seconds of videofrom the camera to the reader. Simultaneously, we recordcamera output using a high input impedance National In-strument USB-6361 DAQ as the ground truth. We choosethe popular PSNR metric commonly used in video appli-cations to compare the video wirelessly streamed to thereader using PWM backscatter to the ground truth videorecorded at the camera using an NI DAQ. PSNR computesthe difference between the ground truth and wirelesslyreceived video.

We measure the PSNR of the received video to evalu-ate the performance of the wireless camera under normallighting conditions (below 300 lux) at a frame rate of13 fps. To evaluate how much our sole wireless commu-nication method affects the quality of received video, weconsider PWM converter output as the ground truth forPSNR measurement. Also, to isolate the impact of AGC,which occurs at the reader, unaltered video received bythe reader prior to applying any AGC is compared to theground truth for PSNR measurement. Fig. 15 plots thePSNR of the received video at the reader as a function of

242 15th USENIX Symposium on Networked Systems Design and Implementation USENIX Association

20

25

30

35

40

0 15 30 45 60 75 90 105 120 135 150-90

-80

-70

-60

-50P

SN

R (

dB)

RS

SI (

dBm

)

Distance (in ft)

PSNR RSSI

Figure 15: Operational range of low-resolution security camera.

the separation between the reader and the wireless camera.The plot shows that wireless camera streamed video atan average PSNR greater than 24 dB to a reader up to adistance of 150 feet away. Beyond 150 feet, the readerdoes not reliably decode the sync pulses, which limitsthe operating range of our wireless camera system. Thus,our analog backscatter approach achieves significantlylonger range for low-resolution video compared to theHD version of the camera due to the trade-off betweenbandwidth/data rate and operating distances.

Applying our distributed inter-frame compression al-gorithm to low resolution videos. We deploy this secu-rity camera in a normal lab space and then implementour distributed inter-frame compression technique on thevideos received. We evaluate the performance of our al-gorithm for three different super pixels sizes, 3×3, 5×5and 7×7 pixels, and plot results in Fig. 16. We achievea 29.4× data reduction using our distributed inter-framecompression technique while maintaining a PSNR greaterthan 30 dB.

20

30

40

50

4 8 12 16 20 24 28 32

PS

NR

(dB

)

Compression Ratio

3x3 5x5 7x7

Figure 16: Distributed interactive compression with our low-resolution security camera. PSNR of the compressed video as a func-tion of compression parameters in a typical lab setting.

Face detection accuracy. Next, we demonstrate that thequality of the video streamed from our low-resolutionCOTS implementation is sufficient for detecting humanfaces. Such a system can be used to detect human oc-cupancy, grant access (such as Ring [12]), or set off analarm in case of an intruder. To evaluate the system, weplace the wireless camera at five different distances rang-ing from 16 to 100 feet from the reader. We ask ten usersto walk around and perform gestures within 5 feet of thecamera. We stream a 2 minutes video at each locationat 13 fps and use the MATLAB implementation of theViola-Jones algorithm to analyze the approximately fourthousand video frames. Fig. 17 shows the accuracy of

40

60

80

100

24 26 28 30 32 34

Acc

urac

y (%

)

PSNR (dB)

Figure 17: Face detection with our low-resolution security cam-era. We show the accuracy of face detection on the video streamed fromour wireless camera.

face detection as a function of the PSNR of the receivedvideo: as the quality (PSNR) of the video improves, theaccuracy of face detection increases. We accurately detectup to 95% of human faces when the PSNR exceeds 30 dB.

7 Related WorkPrior work falls in two different categories.

Backscatter communication. An early example of ana-log backscatter was a gift by the Russians to the US em-bassy in Moscow, which included a passive listening de-vice. This spy device consisted of a sound-modulatedresonant cavity. The voice moved the diaphragm to mod-ulate the cavity’s resonance frequency, which could bedetected by analyzing the RF signals reflected by the cav-ity. [5]. A more recent example of analog backscatter isa microphone-enabled, battery-free tag that amplitude-modulates its antenna impedance using microphone out-put [50, 48]. In contrast, we design the first analog videobackscatter system. Further, prior microphone designshad a low data-rate compared to video streaming. Ourcamera at 10 fps transmits about 9.2M pixels per second;for a microphone, a few kilo-samples of audio transmis-sion is sufficient to fully recover the voice. In addition,our 13 fps 163K pixels per second camera operates atmore than four times the range of the microphone in [50]due to pulse width modulation.

Ekhonet [53] optimizes the computational blocks be-tween the sensor and the backscatter module to reduce thepower consumption of backscatter-based wireless sensors.Our design builds on this work but differs from it in mul-tiple ways: 1) prior work still uses ADCs and amplifierson the cameras to transform pixels into the digital domainand hence cannot achieve streaming video on the limitedharvesting power budget. In contrast, we provide the firstarchitecture for battery-free video streaming by designingan analog video backscatter solution.

Recent work on Wi-Fi and TV-based backscatter sys-tems [25, 26, 23, 30, 42] can achieve megabits per secondof communication speed using a backscatter technique.Integrating these designs with our video backscatter ap-proach would prove a worthwhile engineering effort.

USENIX Association 15th USENIX Symposium on Networked Systems Design and Implementation 243

Low-Power cameras. [41] introduces a self-poweredcamera that can switch its photo diodes between energyharvesting and photo capture mode. Despite being self-powered, these cameras do not have wireless data trans-mission capabilities. [49, 38, 40, 35, 36, 37] show that us-ing off-the-shelf, low-resolution camera modules, one canbuild battery-free wireless cameras that will capture stillimages using the energy they harvested from RF waves,including Wi-Fi and 900 MHz transmissions. Despitetheir ability to transmit data wirelessly, they are heavilyduty cycled and cannot stream video. In particular, thesedesigns can send a new frame at most every ten secondswhen they are very close to the RF power source (withinabout a foot) and once every few tens of minutes at longerdistances [49].

[32] presents a 90× 90 pixels image sensor with pix-els that are sensitive to changes in the environment. If apixel receives sufficient illumination variation, the pixeladdress will be stored in a FIFO, thus compressing theimage to the pixels that have significantly changed. De-spite enabling image compression to occur at the imagesensor, this system does not stream live video. In addi-tion, at this low resolution, it burns about 3 mW of powerwhen running at 30 fps. [28] introduces a 128×128 pixelevent-driven image sensor that emphasizes low latencyfor detecting very fast moving scenes, so its power con-sumption is orders of magnitude higher than our system’s.

[29] addresses the problem of conventional image sen-sors’ power consumption not scaling well as their resolu-tion and frame rate increases. In particular, the authorspropose to change the camera input clock and aggres-sively switch the camera to standby mode based on de-sired image quality. However, streaming video requiresaddressing the power consumption of multiple compo-nents, including camera, communication, and compres-sion. Our work jointly integrates all these components toachieve the first battery-free video streaming design.

Finally, [39] shows that a regular camera on wearabledevices burns more than 1200 mW, which limits the cam-era’s operation time to less than two hours on a wearabledevice. They instead design a low-power wearable vi-sion system that looks for certain events to occur in thefield of view and turns on the primary imaging pipelinewhen those events happen. The power and bandwidthsavings, however, are limited to the application and donot address communication. In contrast, we present thefirst battery-free video streaming application by jointlyoptimizing both backscatter communication and cameradesign and by eliminating power-consuming interfacessuch as ADCs and amplifiers.

8 Limitations and ConclusionThis paper takes a significant first step in designing videostreaming for battery-free devices. In this section, wediscuss limitations and a few avenues for future research.

Security. Our current implementation does not accountfor security. However, to secure the wireless link betweenthe camera and reader, we can leverage the fact that ourdigital core processes the PWM signal. Each wirelesscamera can be assigned a unique pseudo random securitykey. Based on this key, the camera’s digital core canmodulate the width of the PWM-encoded pixel valueusing an XOR gate. The reader, which knows the securitykey, can map the received data to the desired pixel valuesby performing the analogous operation.

ASIC versus off-the-shelf. While existing works onbackscatter cameras focus on using off-the-shelf compo-nents, they treat cameras and backscatter independentlyand just interface the two. Thus, these works cannotachieve video streaming and the low-power demonstratedin this paper. Our key contributions are to make a case fora joint camera and backscatter architecture and to designthe first analog video backscatter solution. However, thisholistic architecture cannot be achieved with off-the-shelfcomponents. The cost of building ICs in a research envi-ronment is prohibitively high. We believe that we specout the IC design (in §4) in sufficient detail using industrystandard EDA tools to take it from the lab to industry.

Mobile device as reader. To support our design for HDvideo streaming from wearable cameras, the smartphonemust support a backscatter reader. We can use RFIDreaders that can be plugged into the headjack of the smart-phone [6] to achieve this. In the future, we believe thatbackscatter readers would be integrated into smartphonesto support video and other applications [23].

Enabling concurrent streaming. Our analog backscat-ter camera design can use frequency division multiplexingtechniques to share the wireless medium across multipledevices. The reader can coordinate communication byassigning different frequency channels to each camera.Multiple cameras can simultaneously backscatter on dif-ferent channels by using different (assigned) frequencyoffsets in our sideband modulation. However, evaluatingthis approach is beyond the scope of this paper.

9 Acknowledgments.We thank Kyle Jamieson and the anonymous reviewersfor their helpful feedback on the paper. Also, we thankYe Wang for his efforts in building a real-time demon-stration of our low-resolution security camera. This workwas funded in part by NSF awards CNS-1407583, CNS-1305072, CNS-1452494, CNS-1420654, Google FacultyResearch Awards and Sloan Fellowship.

244 15th USENIX Symposium on Networked Systems Design and Implementation USENIX Association

References

[1] 6-dBi gain patch antenna. https://www.arcantenna.com/index.php/product_documents/get/document/id/492/. Accessed: March 2017.

[2] Ambarella a9 ultra hd 4k camera soc productbrief. http://www.ambarella.com/uploads/docs/A9-product-brief.pdf.

[3] Barker code. http://mathworld.wolfram.com/BarkerCode.html. Accessed: September 2017.

[4] CentEye Stonyman image sensor datasheet.http://www.centeye.com/products/current-centeye-vision-chips/. Accessed: March 2017.

[5] The great seal bug. http://counterespionage.com/the-great-seal-bug-part-1.html. Accessed:March 2017.

[6] Head jack pluggable rfid reader for smart-phones. http://www.rfidtagworld.com/products/Ear-Jack-reader-UHF_1056.html.

[7] IGLOO Nano FPGA datasheet. https://www.microsemi.com/document-portal/doc_view/130695-ds0110-igloo-nano-low-power-flash-fpgas-datasheet. Accessed: March 2017.

[8] Lmv7219 comparator. http://www.ti.com/lit/ds/symlink/lmv7219.pdfl. Accessed: September2017.

[9] Moving camera on a motorcycle video clip.https://www.youtube.com/watch?v=sHj3xSG-R_E&t=376s.

[10] NCX2200I low power comparator datasheet.http://www.nxp.com/documents/data_sheet/NCX2200.pdf. Accessed: March 2017.

[11] RF5110 amplifier. http://www.rfmd.com/store/downloads/dl/file/id/30508/5110g_product_data_sheet.pdf. Accessed: March 2017.

[12] Ring video doorbell. https://ring.com/. Ac-cessed: March 2017.

[13] Spectacles battery life. http://www.barrons.com/articles/snapchat-spectacles-review-the-good-the-bad-the-revolutionary-1487846715.

[14] Synopsis design complier. http://www.synopsys.com/Tools/Implementation/RTLSynthesis/DesignCompiler/Pages/default.aspx.

[15] USRP X-300. https://www.ettus.com/product/details/X300-KIT. Accessed: March 2017.

[16] Wearable spectacles. http://www.techradar.com/reviews/snap-spectacles.

[17] Youtube website. http://www.youtube.com.

[18] H. R. Beard, A. J. Marquez-Lara, and K. S. Hamid.Using wearable video technology to build a point-of-view surgical education library. JAMA surgery,151(8):771–772, 2016.

[19] S. R. Best and B. C. Kaanta. A tutorial on the re-ceiving and scattering properties of antennas. IEEEAntennas and Propagation Magazine, 51(5), 2009.

[20] L. Chen, X. Tang, A. Sanyal, Y. Yoon, J. Cong, andN. Sun. A 0.7-v 0.6-uw 100− ks/s low-power saradc with statistical estimation-based noise reduction.IEEE Journal of Solid-State Circuits, 52(5):1388–1398, 2017.

[21] R. C. Gonzalez and R. E. Woods. Image processing,volume 2. 2007.

[22] S. Hanson, Z. Foo, D. Blaauw, and D. Sylvester. A0.5 v sub-microwatt cmos image sensor with pulse-width modulation read-out. IEEE Journal of Solid-State Circuits, 45(4):759–767, 2010.

[23] V. Iyer, V. Talla, B. Kellogg, S. Gollakota, andJ. Smith. Inter-technology backscatter: Towardsinternet connectivity for implanted devices. InProceedings of the 2016 Conference on ACM SIG-COMM 2016 Conference, SIGCOMM ’16, pages356–369, New York, NY, USA, 2016. ACM.

[24] K. Kapucu and C. Dehollain. A passive uhf rfidsystem with a low-power capacitive sensor inter-face. In RFID Technology and Applications Confer-ence (RFID-TA), 2014 IEEE, pages 301–305. IEEE,2014.

[25] B. Kellogg, A. Parks, S. Gollakota, J. R. Smith, andD. Wetherall. Wi-fi backscatter: Internet connectiv-ity for rf-powered devices. In Proceedings of the2014 ACM Conference on SIGCOMM, 2014.

[26] B. Kellogg, V. Talla, S. Gollakota, and J. Smith.Passive wi-fi: Bringing low power to wi-fi transmis-sions. In Usenix NSDI, 2016.

[27] W. Kester. Understand sinad, enob, snr, thd,thd+ n, and sfdr so you don’t get lost in thenoise floor. MT-003 Tutorial, www. analog.com/static/importedfiles/tutorials/MT-003. pdf, 2009.

USENIX Association 15th USENIX Symposium on Networked Systems Design and Implementation 245

[28] J. A. Leñero-Bardallo, T. Serrano-Gotarredona,and B. Linares-Barranco. A 3.6us latency asyn-chronous frame-free event-driven dynamic-vision-sensor. IEEE Journal of Solid-State Circuits,46(6):1443–1455, 2011.

[29] R. LiKamWa, B. Priyantha, M. Philipose, L. Zhong,and P. Bahl. Energy characterization and optimiza-tion of image sensing toward continuous mobilevision. In Proceeding of the 11th annual interna-tional conference on Mobile systems, applications,and services, pages 69–82. ACM, 2013.

[30] V. Liu, V. Talla, and S. Gollakota. Enabling in-stantaneous feedback with full-duplex backscatter.In Proceedings of the 20th Annual InternationalConference on Mobile Computing and Networking,MobiCom ’14.

[31] V. Liu, V. Talla, and S. Gollakota. Enabling instan-taneous feedback with full-duplex backscatter. InProceedings of the 20th annual international confer-ence on Mobile computing and networking, pages67–78. ACM, 2014.

[32] U. Mallik, M. Clapp, E. Choi, G. Cauwenberghs,and R. Etienne-Cummings. Temporal change thresh-old detection imager. In Solid-State Circuits Confer-ence, 2005. Digest of Technical Papers. ISSCC. 2005IEEE International, pages 362–603. IEEE, 2005.

[33] F. Michel and M. Steyaert. A 250mv 7.5 µw 61dbsndr cmos sc δσ modulator using a near-threshold-voltage-biased cmos inverter technique. In Solid-State Circuits Conference Digest of Technical Pa-pers (ISSCC), 2011 IEEE International, pages 476–478. IEEE, 2011.

[34] B. Murmann. Adc performance survey 1997-2017. http://web.stanford.edu/~murmann/adcsurvey.html.

[35] S. Naderiparizi, Z. Kapetanovic, and J. R. Smith.Battery-free connected machine vision with wisp-cam. GetMobile: Mobile Computing and Communi-cations, 20(1):10–13, 2016.

[36] S. Naderiparizi, Z. Kapetanovic, and J. R. Smith.Wispcam: An rf-powered smart camera for ma-chine vision applications. In Proceedings of the4th International Workshop on Energy Harvestingand Energy-Neutral Sensing Systems, pages 19–22.ACM, 2016.

[37] S. Naderiparizi, Z. Kapetanovic, and J. R. Smith. Rf-powered, backscatter-based cameras. In Antennasand Propagation (EUCAP), 2017 11th EuropeanConference on, pages 346–349. IEEE, 2017.

[38] S. Naderiparizi, A. N. Parks, Z. Kapetanovic,B. Ransford, and J. R. Smith. Wispcam: A battery-free rfid camera. In RFID (RFID), 2015 IEEE In-ternational Conference on, pages 166–173. IEEE,2015.

[39] S. Naderiparizi, P. Zhang, M. Philipose, B. Priyan-tha, J. Liu, and D. Ganesan. Glimpse: A pro-grammable early-discard camera architecture forcontinuous mobile vision. In Proceeding of the 15thannual international conference on Mobile systems,applications, and services. ACM, 2017.

[40] S. Naderiparizi, Y. Zhao, J. Youngquist, A. P. Sam-ple, and J. R. Smith. Self-localizing battery-freecameras. In Proceedings of the 2015 ACM Interna-tional Joint Conference on Pervasive and Ubiqui-tous Computing, pages 445–449. ACM, 2015.

[41] S. K. Nayar, D. C. Sims, and M. Fridberg. Towardsself-powered cameras. In Computational Photogra-phy (ICCP), 2015 IEEE International Conferenceon, pages 1–10. IEEE, 2015.

[42] A. N. Parks, A. Liu, S. Gollakota, and J. R. Smith.Turbocharging ambient backscatter communication.In Proceedings of the 2014 ACM Conference onSIGCOMM, 2014.

[43] A. N. Parks, A. P. Sample, Y. Zhao, and J. R. Smith.A wireless sensing platform utilizing ambient rf en-ergy. In Biomedical Wireless Technologies, Net-works, and Sensing Systems (BioWireleSS), 2013IEEE Topical Conference on, pages 154–156. IEEE,2013.

[44] A. P. Sample, D. J. Yeager, P. S. Powledge, A. V.Mamishev, and J. R. Smith. Design of an rfid-basedbattery-free programmable sensing platform. IEEETransactions on Instrumentation and Measurement,57(11):2608–2615, 2008.

[45] R. Sarpeshkar. Analog versus digital: extrapolatingfrom electronics to neurobiology. Neural computa-tion, 10(7):1601–1638, 1998.

[46] M. D. Scott, B. E. Boser, and K. S. Pister. An ultra–low power adc for distributed sensor networks. InSolid-State Circuits Conference, 2002. ESSCIRC2002. Proceedings of the 28th European, pages 255–258. IEEE, 2002.

[47] V. Talla, M. Hessar, B. Kellogg, A. Najafi, J. R.Smith, and S. Gollakota. Lora backscatter: Enablingthe vision of ubiquitous connectivity. Proceedingsof the ACM on Interactive, Mobile, Wearable andUbiquitous Technologies, 1(3):105, 2017.

246 15th USENIX Symposium on Networked Systems Design and Implementation USENIX Association

[48] V. Talla, B. Kellogg, S. Gollakota, and J. R. Smith.Battery-free cell phone. ACM UBIComp, 2017.

[49] V. Talla, B. Kellogg, B. Ransford, S. Naderiparizi,S. Gollakota, and J. R. Smith. Powering the nextbillion devices with wi-fi. In Proceedings of the11th ACM Conference on Emerging Networking Ex-periments and Technologies, CoNEXT ’15, pages4:1–4:13, New York, NY, USA, 2015. ACM.

[50] V. Talla and J. R. Smith. Hybrid analog-digitalbackscatter: A new approach for battery-free sens-ing. In RFID (RFID), 2013 IEEE InternationalConference on, pages 74–81. IEEE, 2013.

[51] F. Tang and A. Bermak. An 84 pw/frame per pixelcurrent-mode cmos image sensor with energy har-

vesting capability. IEEE Sensors Journal, 12(4):720–726, 2012.

[52] A. Wang, V. Iyer, V. Talla, J. R. Smith, and S. Gol-lakota. Fm backscatter: Enabling connected citiesand smart fabrics. 14th USENIX Symposium on Net-worked Systems Design and Implementation (NSDI17).

[53] P. Zhang, P. Hu, V. Pasikanti, and D. Ganesan.Ekhonet: High speed ultra low-power backscat-ter for next generation sensors. In Proceedings ofthe 20th annual international conference on Mobilecomputing and networking, pages 557–568. ACM,2014.

USENIX Association 15th USENIX Symposium on Networked Systems Design and Implementation 247