Real time video pipeline for computer vision using embedded GPUs, A

2016 Fall.Includes bibliographical references.This thesis presents case study confirming the feasibility of real time Computer Vision applications on embedded GPUs. Applications that depend on video processing, such as security surveillance, can benefit from applying optimizations common in scientific computing. This thesis demonstrates the benefit of applying such optimizations to real time Computer Vision applications on embedded GPUs. The primary contribution of this thesis is an optimized implementation of ViBe targeting NVIDIA's Jetson TK1. ViBe is a commonly used background subtraction algorithm. Optimizing a background subtraction algorithm accelerates the task of reducing the field of view to only interesting patches of the frames of the video. Placing portable hardware close to capturing devices in the surveillance system reduces bandwidth requirements and cost. The goals of the optimizations proposed for this algorithm are to 1) reduce memory traffic 2) overlap CPU and GPU usage 3) reduce kernel overhead. The optimized implementation of ViBe achieves a frame rate of almost 55 FPS beating the real time goal standard of 30 FPS for real time video. This is a small portion of the real-time window leaving processing time for additional algorithms like object recognition