Zhuangdi "Andy" Xu
Ph.D. Student, Computer Science
School of Computer Science

Georgia Institute of Technology

Date: Thursday, Nov 4, 2021
Time: 9:00am-11:00am (ET)
Location (virtual):  https://bluejeans.com/661409279/0222

Proposal Committee:
Dr. Ramachandran, Umakishore (advisor, School of Computer Science, Georgia Institute of Technology)
Dr. Arulraj, Joy (School of Computer Science, Georgia Institute of Technology)

Dr. Tumanov, Alexey (School of Computer Science, Georgia Institute of Technology)
Dr. Rehg, James M (School of Interactive Computing, Georgia Institute of Technology)

Abstract:

The ubiquity of cameras in our environment and advances in computer vision have enabled novel applications combining sensing, processing, and actuation. These camera-based applications span a variety of domains, including safety, retail, and transportation. Since many of these applications are latency-sensitive and network bandwidth-hungry, in addition to being geo-distributed, edge computing has emerged as a new trend in catering to their computational needs. Meanwhile, low-cost processing resources such as Raspberry Pi and Google Coral TPU are enabling just-in-time processing of camera streams close to their sources. Thus, there is a perfect storm of technology enablers that could all benefit from a comprehensive geo-distributed architecture for smart camera systems at the edge of the network.

Using an exemplar application, namely, space-time vehicle tracking at video ingestion time, we build our dissertation research that presents a scalable system architecture for smart camera networks that contain the following four components: 

1. The on-device processing pipeline involves video analytics tasks such as object detection and classification.
2. Horizontal coordination is responsible for the communication protocol between cameras.
3. Camera topology management handles the camera failures and the addition of new cameras on the fly.
4. Visual data management stores the video and intermediate results that can serve other applications in the future.

This dissertation presents the design and implementation of these four system components for a geo-distributed smart camera system. We evaluate our prototype implementation using a real-world testbed built upon five roadside cameras on the Gatech campus and synthetic video analytics workloads built upon UA-DETRAC and Jackson.