The Bagadus system has been introduced as an automated soccer analysis tool, and consists of an analysis subsystem, tracking subsystem and video subsystem. By automating the integration of these subsystems, Bagadus allows for simplified soccer analysis, with the goal of improving athletes' performance. The system is currently installed at Alfheim stadium in Tromsø, Norway. A part of the video subsystem is the generation of panorama videos from four HD cameras. However, the pipeline for panorama video generation in the first version of the system did not manage to do this online and in real-time.
In this thesis, we present how to build an improved panorama stitcher pipeline that is able to stitch video from four HD cameras into a panorama video online and in real-time. We describe in detail the architecture and modules of this pipeline, and analyze the performance, where we demonstrate real-time, live capture, processing and storage of four individual camera feeds and generation of a panorama video on a single machine. In addition, we focus on how background subtraction can be used to improve the pipeline. As part of this, we discuss how we can utilize player position data to improve the background subtraction process, and also discuss in detail how to optimize the background subtraction process on CPU and GPU.