Graduate Exam Abstract

Rutuja Patil

M.S. Final
July 28, 2016, 9:30 am - 11:30 am
Computer Science Building Room 210
A Real Time Video Pipeline for Computer Vision Using Embedded GPUs

Abstract: Real time video processing is a requirement for applications like self driving cars, however this concept can be further explored for applications like video surveillance monitoring. Such applications can be made more efficient by applying optimizations common in scientific computing to vision pipelines. This work presents a case study of optimizing steps in a Computer Vision pipeline. The case- study is a real world background subtraction algorithm called ViBe on an NVIDIA developed JETSON TK1 GPU. This case study provides evidence that further stages of a vision pipeline are possible in real time. Data Movement is a major bottleneck in many applications. In this work we use a small, inexpensive GPU processor like the JETSON and place it close to the camera. Thus we attempt to solve the network traffic latency problem. The optimizations for this algorithm aim to reduce memory traffic by using data decomposition plans.The simultaneous use of GPU and CPU capability reduces overall execution time, and therefore latency. The optimization space for ViBe has been explored to achieve real time performance. With these optimizations we achieved a frame rate of 55.33 fps placing us well within the real time threshold of 30 fps for real time processing of video. Thus we laid the foundation to a starting stage of an optimized Computer Vision pipeline. We also achieved power efficient optimization with the use of the capabilities of the JESTON TK1.

Adviser: Prof. Ross Beveridge
Co-Adviser: Special Assistant Prof. Catherine Olschanowsky
Non-ECE Member: Prof. Stephen Guzik, Department of Mechanical Engineering
Member 3: Prof. Mahmood Azimi Sadjadi
Addional Members: N/A


Program of Study: