Panoramic Video Stitching

Abstract: In this thesis a system for creating panoramic video has been developed. The panoramic video is formed by stitching several camera streams together. The system is designed as a vehicle mounted system, but can be applied to several other areas, such as surveillance. The system creates the video by finding features that correspond in the overlapping frames. By using cylinder projection the problem is reduced to finding a translation between the images and using algorithms such as ORB matching features can be detected and described. The camera frames are stitched together by calculating the average translation of the matching features. To reduce artifacts such as ghosting, a simple but effective alpha blending technique has been used. The system has been implemented using C++ and the OpenCV library and the algorithm is capable of processing about 15 frames per second making it close to real-time. With future improvements, such as parallel processing of the cameras, the system may be speeded up even further and possibly include other types of image processing, e.g. object recognition and tracking.