Automatic and robust infrared-visible image sequence registration via spatio-temporal association

To solve the problems of the large differences in gray value and inaccurate positioning of feature information during infrared-visible image registration, we propose an automatic and robust algorithm for registering planar infrared-visible image sequences through spatio-temporal association. In particular, we first create motion vector distribution descriptors which represent the temporal motion information of foreground contours in adjacent frames to complete coarse registration without feature extraction. Then, for precise registration, we extracted FAST corners of the foreground, which are described by the spatial location distribution of contour points based on connected blob detection, and match these corners using bidirectional optimal maximum strategy. Finally, a reservoir updated by Better-In, Worse-Out (BIWO) strategy is established to save matched point pairs and obtain the optimal global transformation matrix. Extensive evaluations on the LITIV dataset well demonstrate the effectiveness of the proposed algorithm. Particularly, our algorithm achieves lower registration overlapping errors than the other two state-of-the-arts.