Optimizing large-scale camera analytic systems with wireless insights
Abstract/Contents
- Abstract
- Cameras are key enablers for a wide range of IoT use cases including smart cities, intelligent transportation, AI-enabled farms, and more. Many of them are connected to the Internet via wireless technologies like LPWAN(LoRaWAN, SigFox) and 5G. Conventional camera analytics systems treat the underlying network as a pipe, ignoring the fact that wireless links are lossy and wireless is a broadcast medium. First, I present Starfish that is inspired by the lossy link insight. Starfish is a resilient image compression framework design for LPWAN that processes all the information loss in the application layer, thus simplifying the overall system design. To our best knowledge, it is the first DNN-based image compression framework that runs efficiently on low-cost AIoT hardware. It could be adapted to different task objectives and datasets, and exhibits graceful degradation in lossy scenarios without the need to design and tune source/channel coding. It uses neural network architecture search to generate DNN configuration thus reducing human labor and enable generalization to diverse AIoT hardware. Such end-to-end design reuses the DNN inference pipeline as in object detection/image classification tasks thus it could be easily adopted on various DNN accelerators. Next, I present YOUO (You Only Upload Once) that leverages the broadcast insight. YOUO is an object re-identification system that shifts computation from cloud to devices for better privacy and lower network/computation cost. Each target object may have been captured multiple times by multiple cameras in large-scale camera networks. With the help of multicasting service provided in 5G eMBMS, YOUO enables efficient collaboration across cameras, thus significantly reduced network traffic by more than two orders of magnitude based on experiment results on a large-scale vehicle re-identification dataset. YOUO is 3x as efficient in bandwidth when compared with optimal single-camera filtering with minimal accuracy loss while achieving 70%+ accuracy when uploading only once per object (the absolute minimum) for the entire camera network.
Description
Type of resource | text |
---|---|
Form | electronic resource; remote; computer; online resource |
Extent | 1 online resource. |
Place | California |
Place | [Stanford, California] |
Publisher | [Stanford University] |
Copyright date | 2021; ©2021 |
Publication date | 2021; 2021 |
Issuance | monographic |
Language | English |
Creators/Contributors
Author | Hu, Pan |
---|---|
Degree supervisor | Katti, Sachin |
Thesis advisor | Katti, Sachin |
Thesis advisor | Asgar, Zain |
Thesis advisor | Prabhakar, Balaji, 1967- |
Degree committee member | Asgar, Zain |
Degree committee member | Prabhakar, Balaji, 1967- |
Associated with | Stanford University, Department of Electrical Engineering |
Subjects
Genre | Theses |
---|---|
Genre | Text |
Bibliographic information
Statement of responsibility | Pan Hu. |
---|---|
Note | Submitted to the Department of Electrical Engineering. |
Thesis | Thesis Ph.D. Stanford University 2021. |
Location | https://purl.stanford.edu/tq397gn7659 |
Access conditions
- Copyright
- © 2021 by Pan Hu
- License
- This work is licensed under a Creative Commons Attribution Non Commercial 3.0 Unported license (CC BY-NC).
Also listed in
Loading usage metrics...