Fast and robust image registration is an important research problem in motion video processing. In this paper

we present a novel invariant feature based automatic image registration method to deal with the large image transformation

rotation

scale and illumination changes. The algorithm includes three parts: invariant feature detection

description and matching. First

the multi-scale space of the image is created via a multi-level box filter

and then the feature points are detected in scale space by considering the quality and special distribution simultaneously. Second

we use the Principle Component Analysis to descend dimension of SIFT(Scale invariant feature transform) for feature description. Finally

based on the principle component of the descriptors

a cascade filtering is designed to speed up the feature matching. Experiments with motion videos captured by the dome camera and the Unmanned Aerial Vehicle demonstrate that the proposed algorithm has satisfied performance

and it can provide a solid foundation for further processing such as moving object detection

tracking and classification from moving platform.