A Fast Dense Feature-Matching Model for Cross-Track Pushbroom Satellite Imagery

Feature-based matching can provide high robust correspondences and it is usually invariant to image scale and rotation. Nevertheless, in remote sensing, the robust feature-matching algorithms often require costly computations for matching dense features extracted from high-resolution satellite images due to that the computational complexity of conventional feature-matching model is [Formula: see text]. For replacing the conventional feature-matching model, a fast dense (FD) feature-matching model is proposed in this paper. The FD model reduces the computational complexity to linear by splitting the global one-to-one matching into a set of local matchings based on a classic frame-based rectification method. To investigate the possibility of applying the classic frame-based method on cross-track pushbroom images, a feasibility study is given by testing the frame-based method on 2.1 million independent experiments provided by a pushbroom based feature-correspondences simulation platform. Moreover, to improve the stability of the frame-based method, a correspondence-direction-constraint algorithm is proposed for providing the most favourable seed-matches/control-points. The performances of the FD and the conventional models are evaluated on both an automatic feature-matching evaluation platform and real satellite images. The evaluation results show that, for the feature-matching algorithms which have high computational complexity, their running time for matching dense features reduces from hours level to minutes level when they are operated on the FD model. Meanwhile, based the FD method, feature-matching algorithms can achieve comparable matching results as they achieved based on the conventional model.