An Empirical Study on Estimating Motions in Video Stabilization

摘要

The objective of video stabilization is to remove undesirable motion effects so that only intentional motion effects are retained. The primary benefit of video stabilization is to improve video quality. We present an empirical study that addresses some important practical issues on estimating motions in the development of video stabilization applications. Specifically, we use synthetic test data to investigate the performance of the Kanade-Lucas-Tomasi algorithm. Our contributions include the following. First, We propose a measure, "average pixel deviation" (APD), to directly assess the accuracy of estimated motion parameters in comparison to true motion parameters, which is capable of overcoming some shortcomings of measures used in previous performance studies. Second, a practical issue of error accumulation often arises during the estimation of motion between frames, which has not been addressed in the previous studies to the best of our knowledge. We propose a novel periodic correction strategy, which is capable of effectively reducing error accumulation.