An Automatic Image Registration Algorithm for Tracking Moving Objects in Low-Resolution Video

We propose an automatic image registration algorithm for tracking moving objects in low-resolution videos. The algorithm uses SIFT keypoints to identify matching stationary points in the input frame and the base frame. The best set of matching stationary points is used to create a spatial transform to register the points on the moving objects in the input frame to the base frame. We examined two probabilistic methods and one deterministic method of identifying the stationary points with two different fitness measures (Euclidean and Mahalanobis) and two spatial transforms (affine and projective). Our experiments on two low-resolution videos indicate that our algorithm performs better using the affine transform than the projective transform. However, the differences in average pixel error between the methods of determining stationary points and fitness measures are statistically insignificant. Therefore, which one to use depends on the execution speed and confidence interval required for the application.