Key Frame Extraction Based on Binocular Cameras

Abstract

Key frames provide the effective summary information for video retrieval and browsing, which mostly represents the main content for video sequences. The statistical results prove that color, as a feature that represents image content, proved widely used in the key frame extraction. Stereo correspondence has been an indispensable research direction in the field of computational stereo vision. Stereo vision is introduced for the purpose of improving previous key frame extraction algorithms, which is the key contribution of our work presented in this paper. A novel image similarity measurement is constructed by combining the color moment with SAD. Experimental results demonstrate that the average evaluation coefficient obtained by the improved algorithm in this paper reach up to 95% in the situation of camera stillness and motion. The novel algorithm solves the issue of redundancy which stems from the large changes of content among consecutive frames in the process of extracting key frames using monocular cameras, and performs more robust compared with traditional algorithms.