Research Overview

Edge computing for large-scale video analytics in mobile IoT platforms (2013-Present)

Camera deployments are ubiquitous in video surveillance and traffic monitoring applications for smart cities, generating a huge amount of data. However, today most city-wide solutions analyze it offline after an incident. A large part of the problem arises from extremely slow download times caused by bandwidth and connectivity limitations to these cameras. Toward this end, I have worked on a novel real-time distributed wireless surveillance system called ‘‘Vigil’’ that enables real-time tracking and surveillance across smart cities. The key innovation in the system comes content-aware compression that uses lightweight feature extraction on the video frames at/near camera nodes to rank their importance and then transmits prioritized content to be analyzed further at the cloud based on available bandwidth. The proposed content-aware compression approaches combined with intelligent traffic scheduling results in a 10x wireless capacity savings over systems that upload all videos to the cloud, while maintaining the same accuracy in detecting objects of interest. I ran a pilot deployment of the system both at the Microsoft Research campus in a white-space network and at University College London in a Wi-Fi network. I am extending this work to further analyze the trade-off between energy and bandwidth in such video analytics IoT platforms.

Queries over camera streams across a city-wide deployment are extremely challenging. While hauling all streams to a single datacenter requires storage and parsing of petabytes of data, most of the data is irrelevant to the queries. To enable efficient queries for traffic monitoring over large-scale city-wide camera installations, I built a system ‘‘Optasia’’ that brings together advances from two areas—machine vision and big data analytics systems. The key innovation comes from modularizing the vision modules, such as classifying vehicles by color and type, so that we can apply SQL-type relational dataflow to process the video data efficiently by discarding irrelevant columns early, de-duplicating common modules, and parallelizing the processing. It automatically parallelizes computation as video input size grows or number of cameras increase. I also propose novel chunk-level parallelism that queries video over time that study the traffic flow of vehicles. Evaluation on traffic videos from a large city on complex vision queries shows high accuracy with many fold improvements in query completion time and resource usage.

I am extending this work to vehicular platforms in Celestini Project launched in early Jan 2017 in collaboration with IIT Delhi.

  • A. Chowdhery, M. Levorato, I. Burago and S. Baidya, ‘‘Urban IoT Edge Analytics,’’ book chapter accepted to Fog Computing in the Internet of Things (Intelligence at the Edge), Springer, 2017.

  • I. Burago, M. Levorato, and A. Chowdhery, ‘‘Energy-Delay tradeoff for an Edge-Assisted Mobile Video Acquisition and Processing System,’’ accepted to IEEE International Conference on Sensing, Communication and Networking (SECON), 2017.

  • Y. Lu, A. Chowdhery, and S. Kandula, ‘‘Optasia: A Relational Platform for Efficient Large-Scale Video Analytics,’’ ACM Symposium on Cloud Computing (SoCC), Santa Clara, CA, 2016. [PDF] [MSR Tech Report] [Slides][Bibtex entry]

  • T. Zhang, A. Chowdhery, V. Bahl, K. Jamieson, and S. Banerjee, ‘‘The Design and Implementation of a Wireless Video Surveillance System,’’ ACM MobiCom, Paris, France, 2015. [PDF] [Slides] [1-min Video] [Video Presentation] [Bibtex entry]

Fog networking for Networked drone cameras (2015-Present)

Drones equipped with cameras have become increasingly popular with emerging applications in disaster response, live event streaming and industrial inspection. The video data link is critical to the mission success when drone cameras stream the captured video to one or more ground stations over a wireless channel. The quality of the wireless link varies based on the drone speed and distance of the ground-station. The sheer volume of collected data demands local video processing because transmitting all the data to a ground station or the cloud will tax wireless networks. Yet many applications, especially surveillance and disaster response, need inputs in real-time from a human operator. Today, as little as 5 percent of the data collected by military drones reaches analysts who need to see them. I built a system SkyEyes that adapts video bitrates to rapid wireless channel fluctuations as the drone moves by predicting future throughputs using spatial throughput maps and location information. The proposed predictive approach allows the system to meet latency requirements of tens of milliseconds. I am extending this approach to a fleet of drones that capture and stream video of live events, such as sports game, where we optimize the tradeoff between maximizing video-quality and area of coverage.

  • X. Wang, A. Chowdhery, and M. Chiang, ‘‘Networked Drone Cameras for Sports Streaming,’’ accepted to IEEE International Conference on Distributed Computing Systems (ICDCS) 2017. [3-min Video]

  • X. Wang, A. Chowdhery, and M. Chiang, ‘‘SkyEyes: adaptive video streaming from UAVs,’’ Third Workshop on Hot Topics in Wireless (HotWireless’16) (Invited Paper), New York, USA, 2016. [PDF] [Bibtex entry]

Data driven approach to wireless spectrum crunch (2012-2016)

The combination of exclusive-use spectrum licensing and growing demand for voice, data, and video applications is leading to artificial spectrum scarcity. Knowledge about active radio transmitters is critical to resolve spectrum scarcity: spectrum regulators can use this information to assign spectrum, licensees can identify spectrum usage patterns and better provision their future needs, and dynamic spectrum access applications can innovatively re-use unused portions of the spectrum.

I built a novel system to achieve these goals that continuously senses and characterizes the radio spectrum. Current system designs struggle to achieve large-scale spectrum measurements over time, frequency and space. They can't identify transmitters occupying the spectrum without prior knowledge. The proposed system design includes an end-to-end system to gather RF measurements, called Spectrum Observatory, and a new algorithm, called TxMiner, to characterize the transmitters using unsupervised machine learning. The evaluation of the system on real-world spectrum measurements between 30MHz and 6GHz show that it can detect active transmitters and map their frequency and temporal characteristics. Further, the proposed system can detect rogue transmitters and identify opportunities for dynamic spectrum access with a novel goodness metric.

  • M. Zheleva, R. Chandra, A. Chowdhery, M. Valerio, P. Garnett, A. Kapoor, and A. Gupta, “Enabling a Nationwide Radio Frequency Inventory Using the Spectrum Observatory,” accepted to IEEE Transactions on Mobile Computing, 2017.

  • M. Zheleva, R. Chandra, A. Chowdhery, A. Kapoor, and P. Garnett, ‘‘Txminer: Identifying transmitters in real-world spectrum measurements,’’ IEEE DySpan, Stockholm, Sweden, 2015. [PDF] [Slides] [Bibtex entry]

  • A. Chowdhery, R. Chandra, P. Garnett, and P. Mitchell, ‘‘Characterizing Spectrum Goodness for Dynamic Spectrum Access,’’ IEEE Allerton (Invited paper), Oct. 2012. [PDF]

Network MIMO: Balancing interference cancellation and mitigation (2010-2012)

While dynamic spectrum access techniques strive to increase the amount of available spectrum, interference continues to limit the data-rates of today's wireless and wire-line networks in any usable spectrum band. Cooperative communications promise significant data-rate gains in the next-generation networks by allowing multiple transmitters (or receivers) to cooperate and cancel interference. While information-theoretic concepts of cooperative communications (broadcast/multiple-access channel) have been well-understood in the last decade, the challenges from real-world deployments have just begun to emerge. A key challenge in practical deployments is that only a limited number of transmitters (or receivers) may be allowed to cooperate owing to limited resource or deployment constraints. As a result, the cooperation benefits in terms of data-rate gains can diminish significantly in such limited cooperation scenarios.

Network multi-input multi-output (MIMO) techniques, such as base-station (BS) cooperation, mitigate inter-cell interference in wireless cellular networks at the high cost of sharing the user data across cooperation links between BSs. My research has proposed algorithms that retain more than 65% of the network MIMO interference-cancellation benefits even when the cooperation capacity on links on BSs is limited to 50% of the full-capacity. Apart from power-spectrum optimization, the proposed solutions try to exploit the additional degrees of freedom in wireless networks, i.e. which users are scheduled in any time slot and which BSs cooperate on any frequency tone.

  • S. Mehryar, A. Chowdhery, and W. Yu, ‘‘Dynamic Cooperation Link Selection for Network MIMO Systems with Limited Backhaul Capacity,’’ IEEE International Conference on Communications (ICC), Jun. 2012. [PDF] [Poster]

  • A. Chowdhery, W. Yu, and J. M. Cioffi , ‘‘Cooperative Wireless Multicell OFDMA Network with Backhaul Capacity Constraints,’’ IEEE International Conference on Communications (ICC), Jun. 2011. [PDF] [Poster]

  • H. Dahrouj, W. Yu, and A. Chowdhery, ‘‘Achievable Rate Improvement Using Common Message Decoding for Multicell Networks,’’ Asilomar Conference on Signals, Systems and Computers, Nov. 2010. [PDF]

Enabling Gbps for last-mile Internet access (2009-2013)

My research proposes dynamic spectrum management (DSM) techniques that allow the limited cooperation scenarios to retain most of the cooperation benefits in next-generation systems while maintaining the performance of legacy systems. The theoretical bounds are obtained in terms of achievable rate-regions for the mixture of interference and broadcast/multiple-access channels, where power-spectrum optimization mitigates interference and non-linear decision-feedback equalization cancels interference. Further, I use the insights from theoretically optimal solutions to derive low-complexity practically-implementable algorithms for wire-line digital-subscriber-line (DSL) systems and wireless cellular networks. Among notable results, the proposed algorithms for DSL networks promise a successful co-existence of multi-100 Mbps to 1 Gbps copper-access networks with legacy networks in the next generation, thus enabling gradual incremental upgrade of DSLs to Gbps backhaul for wireless connectivity everywhere. I have also led the standardization of the proposed practically-implementable solutions in DSL standards in USA and UK.

  • K. Kerpez, J. M. Cioffi, S. Galli, G. Ginis, M. Goldburg, M. Mohseni, and A. Chowdhery, ‘‘Compatibility of Vectored and Non-Vectored VDSL2,’’ IEEE Conference on Information Sciences and Systems (CISS), Mar. 2012. [PDF] [Poster]

  • A. Chowdhery, and J. M. Cioffi, ‘‘Dynamic Spectrum Management for Upstream Mixtures of Vectored & Non-vectored DSL,’’ IEEE Globecom, Dec. 2010. [PDF] [Poster]

  • H. Zou, A. Chowdhery, and J. M. Cioffi, ‘‘A Centralized Multi-Level Water-Filling Algorithm for Dynamic Spectrum Management,’’ Asilomar Conference on Signals, Systems & Computers (Invited paper), Nov. 2009. [PDF]

  • J. M. Cioffi, H. Zou, A. Chowdhery, S. Jagannathan, and W. Lee, ‘‘Greening the Copper Access Network with Dynamic Spectrum Management,’’ International Journal of Autonomous and Adaptive Communications Systems, Vol. 3, No. 4, pp. 369-395, 2010. [PDF]