There is an urgent need to extract key information automatically from video for the purposes of indexing, fast retrieval, and scene analysis. To support this vision, reliable scene change detection algorithms must be developed. This paper describes a novel algorithm for wipe scene change detection in uncompressed and MPEG-1 compressed video sequences using statistical and geometric properties of each image. An efficient algorithm is also proposed to estimate statistical features in compressed video without full frame decompression. From uncompressed and MPEG compressed frames, thumbnails are obtained where pixels are the averages and variances of luminance values of macroblocks. Differences between thumbnails are computed and thresholded, and straight lines are detected in the resulting binary images. Persistence and motion of these lines indicate the presence of a shot transition using wipes. Results on video of various content types are reported and validated with the proposed schemes. Furthermore, results show that the accuracy of the detection is above 95% for uncompressed and above 90% for MPEG-1 compressed video.
David R. Bull,
"Wipe scene change detection and classification in video sequences," Journal of Electronic Imaging 13(2), (1 April 2004). https://doi.org/10.1117/1.1666874