Real-time Panorama Stitching using a Single PTZ-Camera without using Image Feature Matching

Sammanfattning: In surveillance applications one thing to consider is how much of a scene one can cover with a camera. One way to augment this is to take images with overlap and blend them, creating a new image with bigger field of view and thereby increase the scene coverage. In this thesis work we have been looking at how one can create panorama images with a pan-tilt-camera and how fast it can be done. We chose a circular panorama representation for this. Our approach was that gathering enough metadata from the camera one can rectify the gathered images and blend them without matching feature-points or other computationally heavy operations. We show that this can be done. The images gathered was corrected for lens distortions and rolling shutter effects arising from rotating the camera. Attempts where made to find an optimal path for the camera to follow while capturing images. An algorithm to do intensity corrections of the images was also implemented. We find that one can rotate the camera at high speeds and still produce a good quality panorama image. The limiting factors are the precision of the meta data gathered, like motion data from the on-board gyro, and the lighting conditions, since a short shutter time is required to minimize motion blur. The quality varies depending on the time taken to capture the images needed to create the spherical projection. The fastest run was done in 1.6 seconds with some distortions. A run in around 4 seconds generally produce a good quality panorama image.