Saman Naderiparizi Mehrdad Hessar Vamsi Talla Shyamnath Gollakota Joshua R Smith University of Washington University of Washington University of Washington University of Washington University of Washington 1. INTRODUCTION take two forms: intra-frame redundancy and inter-frame redundancy. Why would you need low-power wearable cameras? There has been recent interest in wearable cameras like the Snap Intra-frame redundancy is induced by unnecessary pixel Spectacles for applications like life casting, video blogging values, and it is very high in uncompressed videos. It occurs and live streaming concerts, political events and surgeries. because, normally, video frames contain objects larger than a Being worn by a user in the form of glasses, it is imperiously single pixel, thus, little to no variation in color is present in necessary that these cameras be lightweight, comfortable and neighboring pixels. Big variations in pixel values occur mostly not prone to overheating: this makes battery size and life span at the borders of the objects depicted in-frame. With this the most important factors to consider. observation in mind, the solution uses a zigzag pixel scanning technique in order to transmit analog video. This technique Video streaming is an extremely power-hungry operation. ensures the resulting signal gets rid of information from The existing approaches optimize the camera module and the neighboring pixels, thus occupying less bandwidth. Tests have communication system individually to somewhat reduce the shown that a 30 fps 720p video achieves up to a 71× power consumption. This however, does not take into improvement compared to raw digital video transmission. consideration the issues coming from a lower level: the use Inter-frame redundancy in raw videos is also significant, of video codec algorithms and analog-digital converters is still and it happens between consecutive video frames. This required in such approaches, and they are very power- redundancy is eliminated by the inter-frame compression intensive. functionality, which is split between the camera and the reader. The camera mostly transmits low resolution frames (L frames), 2. THE SOLUTION comprised of „super-pixels” (a super pixel corresponds to a Naderiparizi et al. have developed a novel architecture that block of pixels, and it contains the average value of the pixels enables HD video streaming from a low-power wearable in that block). The reader compares the incoming low-res camera to a nearby mobile device. They have eliminated the frame with the previous one: if at least one super-pixel has need for power-hungry ADCs and codecs by introducing an changed sufficiently between frames, the reader asks the analog video backscatter technique that transmits analog pixels camera to send the full resolution frame (I frames); If the from photo-diodes directly to the backscatter antenna. difference between frames passes below the treshold, the reader reconstructs a full-res picture based on the received The signal used in the backscatter link resembles that used pixel values. A Full-resolution frame is also sent periodically in an extremely power-efficient biological nervous system, by the camera to the reader in order to cleanse artifacts and which translates data in spikes that are continuous in time but errors that may accumulate over time. constant in amplitude. This has led to as much as 10,000x lower power consumption than it is norm for existing video 4. IMPLEMENTATION streaming services. Two hardware prototypes have been built for this paper: This technology enables battery-free HD video streaming, one for the high definition (HD) version and the other for the with RF (radio frequency) harvested energy being sufficient to low-resolution version of the camera. power 1080p video streaming at 30fps up to distances of 8 feet from the reader. The low resolution version uses a 112 × 112 grayscale random pixel access camera taken from CentEye. This provides access to individual analog pixel values. 3. SYSTEM DESIGN With the ADC removed from the camera, it is up to the For the HD version, a real implementation has been reader to convert the camera sensor’s analog output into the deemed impossible. As no HD camera on the market gives digital domain. It does this by using pulse width modulation access to raw analog pixel values, the camera module proposed (PWM). A PWM converter can be implemented with passive by this solution cannot be truly implemented yet. To overcome RC components and a comparator, thereby consuming very this, the HD camera has been emulated using a laptop that low power. transmits 1 minute long HD YouTube videos to a DAC (digital to analog convertor) through a USB interface. The analog The reader is also designed to use ASIC (Application- convertor simulates the voltage levels an HD camera sensor specific integrated circuit), which integrates the image sensor, would provide. PWM converter, digital core, oscillator and backscatter modulator onto a small silicon chip. This ensures maximum The reader is implemented on the X-300 USRP software- optimization for low power usage, avoiding the pitfalls off defined radio platform by Ettus Research. It uses a bi-static components such as FPGAs which are designed for flexibility radar configuration with two antennas, one used as a and ease of use, with little to no consideration for power transmitter and the other one as a receiver. consumption. 5. EVALUATION To further optimize the solution, compression is used to eliminate redundancies. Redundancy in video streaming can The following evaluations have been conducted for the proposed solution: A. Operating range 6. SECURITY CAMERA WITH LOWER RESOLUTION The purpose of this evaluation is to label the video quality Because wireless cameras are increasingly popular in according to the distance between the wireless camera and the security and smart home applications, the previously reader. To accomplish this, the distance was increased from 4 exemplified design has been expanded to have lower resolution to 16 feet, and the camera was configured to repeatedly (enough for applications such as face detection) but to operate transmit a 15-second video using PWM backscatter at longer distances. Thus, the camera can detect people's faces, communication. Thus, noise (SNR) was measured and the and this system can also set an alarm when an intruder is effective number of bits reaching the reader (ENOB) was detected. The distance you can operate is up to 100 feet away calculated. from the reader, and the accuracy is 95%. Following the evaluation, it was found that at a distance of less than 6 feet, an ENOB of more than 7 was obtained, 7. RELATED WORK indicating a negligible video quality degradation. The higher the distance, the higher the noise, the ENOB drops, so the A. Backscatter Communication video quality decreases more and more. However, a distance The first backscatter devices were spy devices [1]; they of 16 feet is enough for a wearable wireless camera because it were used for listening in, using a resonant cavity. The human can transmit video streams to a nearby mobile device, and the voice moved the diaphragm to shape the cavity resonance video quality is not degraded too much. frequency, which could be detected by analyzing the RF signals reflected by the cavity. B. Head position and movement All backscatter systems were microphone models, none of To make this evaluation, the participant wore the antenna them used video streams. Moreover, they had a low rate of of the wireless camera and made different head movements. audio data transmitted. Thus, the wearable wireless camera is He was at 5 feet away from the reader, whose position was the first video backscatter system. always a fixed one. The movements analyzed were: sitting straight, moving his head to the left and right, turning his head Additionally, this wireless camera sends 4 times more to one side, moving his head up and down, talking. video data than audio data transmitted by existing audio In this case, the noise and the effective number of bits were backscatter systems. measured and it was found that, as a result of these movements, the video quality is higher than the original B. Low-power cameras quality of the video. The first example [2] is a camera whose photodiodes can switch between power extraction and photo capture mode. C. Extracting RF signals One disadvantage of this system is that it is not a wireless one. The purpose of this assessment is to study the feasibility of Later, the low-resolution wireless cameras [3, 4, 5, 6] developing a battery-free HD camera that works by extracting appeared, but they could not transmit video streams: they could the RF signals transmitted by the reader. send a new frame at every 10 seconds (when the distance from It has been observed that at a distance of 4 and 6 feet there the RF power source is about a foot) or once in a few tens of will be enough amount of available energy from RF extraction minutes (when the distance is greater). in order to operate the wireless camera at a resolution of A low-power system [7] replaced a regular camera because 1080p. At higher distances, the power is reduced, which will the regular one limited the operating time to less than 2 hours. lead to a decrease in resolution. At a distance of 16 feet, the However, the battery-free wearable camera optimized the resolution will reach 480p; beyond this distance, the power power savings by eliminating ADCs and amplifiers. obtained by extracting RF energy is not enough to continue to supply the wireless camera. 8. CONCLUSIONS D. Distributed inter-frames compression This solution does not take security into account, but in Two videos were used in this test: an HD video transmitted order to secure the link between the reader and the wireless by a fixed monitoring camera and another one transmitted by a camera, each camera could use a security key, which is known camera mounted on a person riding a motorcycle. only by the reader. The pixel sent by the camera will be The compromise between the compression ratio and the encoded using this random generated key and decoded by the reader. Peak Signal to Noise Ratio (PSNR) for different compression rates is obtained by changing the threshold at which we The use of ASIC instead of off-the-shelf devices ensures consider the super pixel to have changed significantly (from maximum optimization in regards to power consumption, as 20 to 20 pixels). multi-purpose devices like FPGAs often overlook this aspect. It has been observed that, for static cameras, a In order to use a smartphone as a reader, the mobile device compression rate of at least 35x is achieved, maintaining a needs to support a backscatter reader or a RFID reader [8], PSNR of over 30 dB. For dynamic cameras, the compression which can be plugged into the headjack. ratio will be 2x for a PSNR of more than 30 dB. [4] S. Naderiparizi, Z. Kapetanovic, and J. R. Smith. Wispcam: An rf- powered smart camera for machine vision applications. In Proceedings of the 4th International Workshop on Energy Harvesting and Energy- Neutral Sensing Systems, pages 19–22. ACM, 2016. [5] S. Naderiparizi, Z. Kapetanovic, and J. R. Smith. Rf-powered, backscatter-based cameras. In Antennas and Propagation (EUCAP), 2017 11th European Conference on pages 346–349. IEEE, 2017. [6] S. Naderiparizi, A. N. Parks, Z. Kapetanovic, B. Ransford, and J. R. Smith. Wispcam: A battery-free rfid camera. In RFID (RFID), 2015 IEEE International Conference on pages 166–173. IEEE, 2015. REFERENCES [7] S. Naderiparizi, P. Zhang, M. Philipose, B. Priyantha, J. Liu, and D. Ganesan. Glimpse: A programmable early-discard camera architecture for continuous mobile vision. In Proceeding of the 15th annual [1] The great seal bug. http://counterespionage.com/the-great-seal-bug-part- international conference on Mobile systems, applications, and services 1.html. ACM, 2017 [2] S. K. Nayar, D. C. Sims, and M. Fridberg. Towards self-powered [8] Head jack pluggable rfid reader for smartphones. cameras. In Computational Photography (ICCP), 2015 IEEE http://www.rfidtagworld.com/products/Ear-Jack-reader- International Conference on, pages 1–10. IEEE, 2015. UHF_1056.html. [3] S. Naderiparizi, Z. Kapetanovic, and J. R. Smith. Battery-free connected machine vision with wispcam. GetMobile: Mobile Computing and Communications, 20(1):10–13, 2016.