MobiCom 2018 Papers

A rich body of work has focused on motion tracking techniques using inertial sensors, namely accelerometers, gyroscopes, and magnetometers. Applications of these techniques are in indoor localization, gesture recognition, inventory tracking, vehicular motion, and many others. This paper identifies room for improvement over today's motion tracking techniques. The core observation is that conventional systems have trusted gravity more than the magnetic North to infer the 3D orientation of the object. We find that the reverse is more effective, especially when the object is in continuous fast motion. We leverage this opportunity to design MUSE, a magnetometer-centric sensor fusion algorithm for orientation tracking. Moreover, when the object's motion is somewhat restricted (e.g., human-arm motion restricted by elbow and shoulder joints), we find new methods of sensor fusion to fully leverage the restrictions. Real experiments across a wide range of uncontrolled scenarios show consistent improvement in orientation and location accuracy, without requiring any training or machine learning. We believe this is an important progress in the otherwise mature field of IMU-based motion tracking.

Visible Light Positioning (VLP) has attracted much research effort recently. Most existing VLP approaches require special designed light or receiver, collecting light information or strict user operation (e.g., horizontally holding mobile phone). This incurs a high deployment, maintenance and usage overhead. We present RainbowLight, a low cost ambient light 3D localization approach easy to deploy in today's buildings. Our key finding is that light through a chip of polarizer and birefringence material produces specific interference and light spectrum at different directions to the chip. We derive a model to characterize the relation for direction, light interference and spectrum. Exploiting the model, RainbowLight calculates the direction to a chip after taking a photo containing the chip. With multiple chips, RainbowLight designs a direction intersection based method to derive the location. We implement RainbowLight and extensively evaluate its performance in various environments. The evaluation results show that RainbowLight achieves an average localization error of 3.3 cm in 2D and 9.6 cm in 3D for light on, and an error of 7.4 cm in 2D and 20.5 cm in 3D for light off scenario in daytime.

We present CrossSense, a novel system for scaling up WiFi sensing to new environments and larger problems. To reduce the cost of sensing model training data collection, CrossSense employs machine learning to train, off-line, a roaming model that generates from one set of measurements synthetic training samples for each target environment. To scale up to a larger problem size, CrossSense adopts a mixture-of-experts approach where multiple specialized sensing models, or experts, are used to capture the mapping from diverse WiFi inputs to the desired outputs. The experts are trained offline and at runtime the appropriate expert for a given input is automatically chosen. We evaluate CrossSense by applying it to two representative WiFi sensing applications, gait identification and gesture recognition, in controlled single-link environments. We show that CrossSense boosts the accuracy of state-of-the-art WiFi sensing techniques from 20% to over 80% and 90% for gait identification and gesture recognition respectively, delivering consistently good performance - particularly when the problem size is significantly greater than that current approaches can effectively handle.

Network connectivity is often one of the most challenging aspects of deploying sensors. In many countries, cellular networks provide the most reliable, highest bandwidth, and greatest coverage option for Internet access. While this makes smartphones a seemingly ideal platform to serve as a gateway between sensors and the cloud, we find that a device designed for multi-tenant operation and frequent human interaction becomes unreliable when tasked to continuously run a single application with no human interaction, a seemingly counter-intuitive result. Further, we find that economy phones cannot physically withstand continuous operation, resulting in a surprisingly high rate of permanent device failures in the field. If these observations hold more broadly, they would make mobile phones poorly suited to a range of sensing applications for which they have been rumored to hold great promise.

We present DeepCache, a principled cache design for deep learning inference in continuous mobile vision. DeepCache benefits model execution efficiency by exploiting temporal locality in input video streams. It addresses a key challenge raised by mobile vision: the cache must operate under video scene variation, while trading off among cacheability, overhead, and loss in model accuracy. At the input of a model, DeepCache discovers video temporal locality by exploiting the video's internal structure, for which it borrows proven heuristics from video compression; into the model, DeepCache propagates regions of reusable results by exploiting the model's internal structure. Notably, DeepCache eschews applying video heuristics to model internals which are not pixels but high-dimensional, difficult-to-interpret data. Our implementation of DeepCache works with unmodified deep learning models, requires zero developer's manual effort, and is therefore immediately deployable on off-the-shelf mobile devices. Our experiments show that DeepCache saves inference execution time by 18% on average and up to 47%. DeepCache reduces system energy consumption by 20% on average.

Mobile virtual reality (VR) headsets, such as Google Cardboard and Samsung GearVR, can reuse a smartphone as near-eye display to create immersive experience. But such devices barely support any user interaction, even for simple tasks such as menu selection and single-character input. In this paper, we design Inkput, a simple passive interface attached to the unexploited backside of the headset to enable touch sensing. Inkput is a piece of paper substrate with carbon ink patterns printed atop. It leverages the column of electrodes near the edge of the smartphone touchscreen to sense multi-touch on the 2D space, and is even able to locate finger hovering. Our experiments demonstrate that Inkput can precisely detect touch positions with mm-level precision. Our case studies in actual VR applications also verify that Inkput can support common VR interactions and can even outperform high-end handheld controllers in terms of efficiency.

This paper explores the possibility of detecting the hidden presence of wireless eavesdroppers. Such eavesdroppers employ passive receivers that only listen and never transmit any signals making them very hard to detect. In this paper, we show that even passive receivers leak RF signals on the wireless medium. This RF leakage, however, is extremely weak and buried under noise and other transmitted signals that can be 3-5 orders of magnitude larger. Hence, it is missed by today's radios. We design and build Ghostbuster, the first device that can reliably extract this leakage, even when it is buried under ongoing transmissions, in order to detect the hidden presence of eavesdroppers. Ghostbuster does not require any modifications to current transmitters and receivers and can accurately detect the eavesdropper in the presence of ongoing transmissions. Empirical results show that Ghostbuster can detect eavesdroppers with more than 95% accuracy up to 5 meters away.

Passive radio frequency identification (RFID) tags are ubiquitous today due to their low cost (a few cents), relatively long communication range ($\sim$7-11~m), ease of deployment, lack of battery, and small form factor. Hence, they are an attractive foundation for environmental sensing. Although RFID-based sensors have been studied in the research literature and are also available commercially, manufacturing them has been a technically-challenging task that is typically undertaken only by experienced researchers. In this paper, we show how even hobbyists can transform commodity RFID tags into sensors by physically altering (`hacking') them using COTS sensors, a pair of scissors, and clear adhesive tape. Importantly, this requires no change to commercial RFID readers. We also propose a new legacy-compatible tag reading protocol called Differential Minimum Response Threshold (DMRT) that is robust to the changes in an RF environment. To validate our vision, we develop RFID-based sensors for illuminance, temperature, touch, and gestures. We believe that our approach has the potential to open up the field of batteryless backscatter-based RFID sensing to the research community, making it an exciting area for future work.

5G New Radio (NR) is designed to operate under a broad range of frequency bands and support new applications with ultra-low latency. To support its diverse operating conditions, a set of different OFDM numerologies has been defined in the standards body. Under this numerology, it is necessary to perform scheduling with a time resolution of ∼100 μs. This requirement poses a new challenge that does not exist in LTE and cannot be supported by any existing LTE schedulers. In this paper, we present the design of GPF -- a GPU-based proportional fair (PF) scheduler that can meet the ∼100 μs time requirement. The key ideas include decomposing the scheduling problem into a large number of small and independent sub-problems and selecting a subset of sub-problems from the most promising search space to fit into a GPU. By implementing GPF on an off-the-shelf Nvidia Quadro P6000 GPU, we show that GPF is able to achieve near-optimal performance while meeting the ∼100 $\mathrmμs time requirement. GPF represents the first successful design of a GPU-based PF scheduler that can meet the new time requirement in NR.

This paper presents a battery-free wearable eye tracker that tracks both the 2D position and diameter of a pupil based on its light absorption property. With a few near-infrared (NIR) lights and photodiodes around the eye, NIR lights sequentially illuminate the eye from various directions while photodiodes sense spatial patterns of reflected light, which are used to infer pupil's position and diameter on the fly via a lightweight inference algorithm. The system also exploits characteristics of different eye movement stages and adjusts its sensing and computation accordingly for further energy savings. A prototype is built with off-the-shelf hardware components and integrated into a regular pair of glasses. Experiments with 22 participants show that the system achieves 0.8-mm mean error in tracking pupil position (2.3 mm at the 95th percentile) and 0.3-mm mean error in tracking pupil diameter (0.9 mm at the 95th percentile) at 120-Hz output frame rate, consuming 395 µW mean power supplied by two small, thin solar cells on glasses side arms.

This paper addresses the performance of systems which use commercial wireless devices to make bistatic RF channel measurements for non-contact respiration sensing. Published research has typically presented results from short controlled experiments on one system. In this paper, we deploy an extensive real-world comparative human subject study. We observe twenty patients during their overnight sleep (a total of 160 hours), during which contact sensors record ground-truth breathing data, patient position is recorded, and four different RF breathing monitoring systems simultaneously record measurements. We evaluate published methods and algorithms. We find that WiFi channel state information measurements provide the most robust respiratory rate estimates of the four RF systems tested. However, all four RF systems have periods during which RF-based breathing estimates are not reliable.

Parallel backscatter is a promising technique for high throughput, low power communications. The existing approaches of parallel backscatter are based on a common assumption, i.e. the states of the collided signals are distinguishable from each other in either the time domain or the IQ (the In-phase and Quadrature) domain. We in this paper disclose the superclustering phenomenon, which invalidates that assumption and seriously affects the decoding performance. Then we propose an interstellar travelling model to capture the bursty Gaussian process of a collided signal. Based on this model, we design Hubble, a reliable signal processing approach to support parallel backscatter in the wild. Hubble addresses several technical challenges: (i) a novel scheme based on Pearson's Chi-Square test to extract the collided signals' combined states, (ii) a Markov driven method to capture the law of signal state transitions, and (iii) error correction schemes to guarantee the reliability of parallel decoding. Theoretically, Hubble is able to decode all the backscattered data, as long as the signals are detectable by the receiver. The experiment results demonstrate that the median throughput of Hubble is $11.7\times$ higher than that of the state-of-the-art approach.

Mobile vision systems such as smartphones, drones, and augmented-reality headsets are revolutionizing our lives. These systems usually run multiple applications concurrently and their available resources at runtime are dynamic due to events such as starting new applications, closing existing applications, and application priority changes. In this paper, we present NestDNN, a framework that takes the dynamics of runtime resources into account to enable resource-aware multi-tenant on-device deep learning for mobile vision systems. NestDNN enables each deep learning model to offer flexible resource-accuracy trade-offs. At runtime, it dynamically selects the optimal resource-accuracy trade-off for each deep learning model to fit the model's resource demand to the system's available runtime resources. In doing so, NestDNN efficiently utilizes the limited resources in mobile vision systems to jointly maximize the performance of all the concurrently running applications. Our experiments show that compared to the resource-agnostic status quo approach, NestDNN achieves as much as 4.2% increase in inference accuracy, 2.0× increase in video frame processing rate and 1.7× reduction on energy consumption.

Various sensing modalities have been exploited for indoor location sensing, each of which has well understood limitations, however. This paper presents a first systematic study on using the electromagnetic field (EMF) induced by a building's electric power network for simultaneous localization and mapping (SLAM). A basis of this work is a measurement study showing that the power network EMF sensed by either a customized sensor or smartphone's microphone as a side-channel sensor is spatially distinct and temporally stable. Based on this, we design a SLAM approach that can reliably detect loop closures based on EMF sensing results. With the EMF feature map constructed by SLAM, we also design an efficient online localization scheme for resource-constrained mobiles. Evaluation in three indoor spaces shows that the power network EMF is a promising modality for location sensing on mobile devices, which is able to run in real time and achieve sub-meter accuracy.

Mobile and IoT scenarios increasingly involve interactive and computation intensive contextual recognition. Existing optimizations typically resort to computation offloading or simplified on-device processing. Instead, we observe that the same application is often invoked on multiple devices in close proximity. Moreover, the application instances often processsimilar contextual data that map to thesame outcome. In this paper, we proposecross-device approximate computation reuse, which minimizes redundant computation by harnessing the "equivalence'' between different input values and reusing previously computed outputs with high confidence. We devise adaptive locality sensitive hashing (A-LSH) and homogenized k nearest neighbors (H-kNN). The former achieves scalable and constant lookup, while the latter provides high-quality reuse and tunable accuracy guarantee. We further incorporate approximate reuse as a service, called \name, in the computation offloading runtime. Extensive evaluation shows that, when given 95% accuracy target, \name\ consistently harnesses over 90% of reuse opportunities, which translates to reduced computation latency and energy consumption by a factor of 3 to 10.

Increasingly, vehicles sold today are connected cars: they offer vehicle-to-infrastructure connectivity through built-in WiFi and cellular interfaces, and they act as mobile hotspots for devices in the vehicle. We study the connection quality available to connected cars today, focusing on user-facing, latency-sensitive applications. We find that network latency varies significantly and unpredictably at short time scales and that high tail latency substantially degrades user experience. We also find an increase in coverage options available due to commercial WiFi offerings and that variations in latency across network options are not well-correlated. Based on these findings, we develop RAVEN, an in-kernel MPTCP scheduler that mitigates tail latency and network unpredictability by using redundant transmission when confidence about network latency predictions is low. RAVEN has several novel design features. It operates transparently, without application modification or hints, to improve interactive latency. It seamlessly supports three or more wireless networks. Its in-kernel implementation allows proactive cancellation of transmissions made unnecessary through redundancy. Finally, it explicitly considers how the age of measurements affects confidence in predictions, allowing better handling of interactive applications that transmit infrequently and networks that exhibit periods of temporary poor performance. Results from speech, music, and recommender applications in both emulated and live vehicle experiments show substantial improvement in application response time.