Multi-camera Video Stitching Based on Foreground Extraction

This paper presents an automatic methodology to stitch multiple videos from different views together based on foreground extraction.Although video stitching technique has been studied by many researchers,the stitching results are still limited by the challenges caused by the moving objects,especially the foregrounds that are notably close to the cameras.The method was mainly concerned with the inaccurate registration and the stitching ghost caused by the moving foreground in the videos.The background frames of the videos were extracted and registered to find out the correlation between them.The moving objects that were separated from the background frames were sent to a different function to find out whether the objects from different videos represented the same one caught in matching views.The objects were then chosen as block to be mosaiced back or dissolved to avoid ghost caused by directly stitching.Experimental results show that it is a highly-matched and robust video stitch algorithm,and can basically achieve real-time stitch.