Efficient estimation of 3D camera motion parameters with applications to video coding

Abstract

In image sequence coding, motion estimation (ME) and motion compensation (MC) techniques play an important role to reduce the information transmission rate while preserving the acceptable image quality. Most video coding standards use local motion estimation only to form a prediction of the current frame. However, in general, both global motion caused by the movement of the camera and local motion caused by the movement of the object exist in the image sequences. Therefore, the global ME and MC can improve the motion prediction and reduce the side information greatly. In this dissertation, we present several methods for estimating and compensating 3D camera motion in image sequences for the applications to the video coding system. First, we present a new linear motion parameter model which describe the camera motion parameters : zoom, pan, tilt, swing, and focal length. To estimate camera motion parameters based on the proposed model, the mixed least squares - total least squares (MLS-TLS) method is proposed. The proposed estimation procedure consists of feature correspondence establishment, finding the parameters by fitting the correspondence data to the proposed model equation, and outliers rejection to reduce feature matching error. Unlike the existing linear least squares techniques, the proposed method accurately estimates even the large rotation angles and the focal length. Experimental results show that the proposed method outperforms the conventional methods under identical conditions, especially for a large rotation images. Second, in order to improve the estimation accuracy of the global motion parameter, we propose an efficient outlier rejection method which selects the reliable observation points adaptively and eliminates the bad observation points effectively. Unlike the existing linear recursive algorithms with an arbitrary thresholding, the proposed algorithm converges fast and is robust to feature matching error. Experimental results show that...