A fast feature-based video mosaic
Lin Wu, Jingxin Hong, Hao Zhang, Lin Li · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 2009
This paper presents a fast method to construct an entire panorama from video sequence. Firstly, SURF, which has good performance in computation time and accuracy, is used to detect features for frames from video streams. Secondly, a novel matching scheme based on hash mapping and the ratio between the Euclidean distance to the nearest and the second nearest neighbors is proposed to match SURF features. As there may be some error matchings generated above, especially in the presence of objects moving, a RANSAC technique is applied to eliminate outliers. Besides, in order to avoid mosaicing all video frames which typically contains significant redundancy, we adaptively identify key frames based on the number of tracked feature points. Finally, the wide field of video image is got by stitching the key frames mapped to a reference coordinate system. Experimental results demonstrate the fast-speed and superior quality of this proposed method.