Georgia Institute of Technology
Scalable Hardware Accelerator Design for State-of-the-Art Computer Vision Algorithms
Abstract
dc:description.abstractThe objective of the proposed research is to develop an energy-efficient hardware accelerator targeted at tackling the scalability challenge presented by computationally intensive state-of-the-art deep learning models for multiple object tracking in High-Definition (HD) video streams. Using Field Programmable Gate Array (FPGA), the goal is to build an accelerator with high parallelism that can deliver real-time video processing performance. Different from existing works, the proposed research makes a solid step towards the synergized software/hardware co-optimization for realistic multi-object tracking model implementation. Compared to the state-of-the-art baseline, the proposed design achieves 12.5× latency reduction, 20.9× effective frame rate improvement, 5.83× lower power, and 9.78× better energy efficiency, without much accuracy drop. Furthermore, motivated by the lack of an effective framework for partitioning multi-modal multi-task deep learning models on multi-FPGA systems, the research explores the development of a novel mapping framework that can optimally achieve the desired performance objectives under the specified resource constraints while minimizing the overall system power consumption. For various computation-heavy models, experiment results demonstrate that the proposed framework can remarkably outperform existing mapping methods and lead to an average reduction of 35%, 62%, and 70% in the number of low-end, mid-end, and high-end FPGAs required to achieve the same throughput, respectively.
Degree
thesis:*- Level thesis:degree_level
- Masters
- Department dc:contributor.department
- Electrical and Computer Engineering
- Grantor dc:publisher
- Georgia Institute of Technology
- Year dc:date.issued
- 2023
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Kamath, Akshay Karkal
- Advisors dc:contributor.advisor
-
- Hao, Cong
- Yu, Shimeng
- Krishna, Tushar
Subjects
dc:subject × 7Rights
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/1853/72063