Real-time Traffic Surveillance
With cities trying to reduce traffic accidents and improve traffic flow, hundreds of thousands of traffic cameras are currently deployed in major cities for traffic surveillance. However, processing such data in real-time is still challenging, due to the cost of bandwidth as well as hardware. In this project, we look at how to process such large data at a minimal cost by designing lightweight filtering algorithms, as well as utilization of low-powered edge devices.
Modeling of Edge Compute Infrastructure
While edge computing has become increasingly relevant to satisfy the demand for low-latency based compute services, the hardware used in edge computing has become extremely complex. This is due to the incorporation of processing units such as graphical processing units, neural processing units and field processing gate arrays. Our focus is on designing better abstraction models to identify which processing units should be used for such computation, and the tradeoffs involved in them. This would enable both lower latency and more energy-efficient computation.
With spectrum getting expensive, it is essential to monitor its usage and protect it from illegal users. Current state-of-the-art techniques deploy sensors using crowdsourcing. However, running each sensor costs energy and data. This project looks at various techniques of improving accuracy of spectrum monitoring while keeping its cost low.
With a large number of small devices being connected to the Internet, it is essential for applications running on them to utilize the capability of more powerful devices located nearby or the cloud. A key issue is to determine the location of execution of each component of the applications, while ensuring that the latency and energy requirements of execution are satisfied. Current studies depend on heuristics which may provide poor results for different applications, or some optimization solver which may take a long time to compute. In contrast, we propose algorithms that both improve the performance of offloaded applications in actual applications, as well as provide some performance guarantees. We validate our approach using trace-driven simulations.
360 Video Streaming
Although 360 videos are becoming popular, streaming them over wireless networks is still challenging. A key reason for this problem is that a lot of the data downloaded is not seen by the viewer and is, therefore, wasted. In this project, we combine state-of-the-art computer vision techniques and bitrate adaptation algorithms to make streaming over wireless networks feasible. By integrating our technique with an off-the-shelf 360 video player, we show that significant data savings can be made without much loss of user Quality of Experience.