University of Illinois Urbana-Champaign

DIMA system for real-time object detention

Chai, Yuji

This item is only available for download by members of the University of Illinois community. Students, faculty, and staff at the U of I may log in with your NetID and password to view the item. If you are trying to access an Illinois-restricted dissertation or thesis, you can request a copy through your library's Inter-Library Loan office or purchase a copy directly from ProQuest.

Permalink

https://hdl.handle.net/2142/103997

Description

Title
DIMA system for real-time object detention
Author(s)
Chai, Yuji
Contributor(s)
Shanbhag, Naresh R.
Issue Date
2019-05
Keyword(s)
  • Deep In-Memory Architecture
  • Machine Learning System
  • Hardware Acceleration
  • Object Detection
Abstract
In recent years, breakthroughs in machine learning and deep learning have shown their unlimited potential in autonomous driving, unmanned stores, etc. However, their superior capabilities come with high computational costs. While these techniques can achieve reasonable performance on servers or workstations equipped with multiple GPUs, they cannot be easily deployed on edge or IoT platforms. This challenge demands a computationally efficient solution to enable machine learning or deep learning for the edge. Towards this goal, this research focused on developing a system to accelerate video inference by utilizing deep in memory architecture (DIMA) ICs designed and prototyped recently in our research group. The DIMA IC embeds mixed-signal compute blocks within SRAM memory array to accelerate machine learning models for image classification with 3.1x higher power efficiency and 2.1x lower inference latency. While DIMA IC achieved fast inference on single cropped images, its overall test setup was not optimized for video inference. To address this issue, we replaced the original MCU + PC in the previous setup with a Raspberry Pi and enabled it to directly process image information and control the chip. As a result, the system performance improved from more than 10 seconds per image to around 13 ms inference time. This also enabled us to complete the system with real camera input. With frames streaming into the Raspberry Pi, it will preprocess the image and regional proposal for the chip. The chip will accelerate the image classification process and provide the system with real-time object recognition capability.
Type of Resource
text
Language
en
Permalink
http://hdl.handle.net/2142/103997

Owning Collections

Senior Theses - Electrical and Computer Engineering PRIMARY
The best of ECE undergraduate research