Graduation Year
2020
Document Type
Thesis
Degree
M.S.Cp.
Degree Name
MS in Computer Engineering (M.S.C.P.)
Degree Granting Department
Engineering Computer Science
Major Professor
Srinivas Katkoori, Ph.D.
Co-Major Professor
Hao Zheng, Ph.D.
Committee Member
Mehran Mozaffari Kermani, Ph.D.
Keywords
CNN, Distributed System, FPGA, PYNQ, YOLO
Abstract
In the era of IoT (Internet of Things) and edge computing, there is a rising need for real-time applications in the domain of computer vision. The increase in hardware computing capabilities gave rise to applications of neural networks in various fields. Implementing IoT with neural networks in domains such as image and video recognition has shown promising performance when deployed in complex environments. There is an emerging demand for applications that require data computation in real-time with low latency. In an effort to address these issues, while keeping in mind the computing capabilities of IoT devices, we seek to develop a framework for efficient object detection on a distributed constrained platform system. In this thesis, a scalable and adaptable network for fast and easy convolutional neural network prototyping on the Xilinx PYNQ cluster has been proposed.
We employed PYNQ Z1 AP-SoC (All Programmable System-on-Chip) as the IoT edge node platform and integrated state-of-the-art algorithms for object detection. The distributed architecture is robust and exploits the heterogeneous computing capability of the PYNQ platform. The proposed work is on a wireless distributed network with minimal communication latency. We demonstrate the framework for low frame rate applications where the scenery is not changing rapidly. We were able to achieve 19.23 frames per second with three IoT nodes with Binarized Neural Network (BNN) image classification algorithm. The frames per second rate is directly proportional to the number of nodes in the network. The communication latency can be offset by the scalability offered by the distributed framework.
Scholar Commons Citation
Kalyanam, Lakshmikavya, "Edge Computing for Deep Learning-Based Distributed Real-time Object Detection on IoT Constrained Platforms at Low Frame Rate" (2020). USF Tampa Graduate Theses and Dissertations.
https://digitalcommons.usf.edu/etd/9541
