Fast Global Motion Estimation for Compressed domain videos

Abstract

Global motion estimation, one of the key techniques in video processing applications, involves the detection of the motion experienced by the camera. Given a set of motion vectors, the main focus was on finding the motion undergone by the source at the time of recoding. The captured data is available in their compressed form, where the video sequences in the form of motion vectors. Using this data, an method was proposed which estimated the motion and presented the results in the form of a matrix where each of the entries in the matrix denoted a particular motion. Motion estimation is one of the main task when it comes to applications like, video stabilization, object segmentation, etc,.
The main work of thesis involves the estimation of the global motion parameters which exhibit the amount of motion undergone by the camera. The works carried out consists of global motion estimation using the concept of statistical mode. The main aim of this work is to tackle the issues addressing computational complexity in terms of PSNR and running time. The method uses the concept of statistical mode where the running time is reduced as it utilizes just one iteration as compared to some state of the art methods. A parametric motion model was taken into consideration for the estimation of the motion parameters, where the motion vectors were dealt with, in their polar form. The histograms of the magnitude and the orientation of the motion vectors were used and the inlier detection or the outlier removal, was carried out by finding the modal values of the two histograms. In order to justify the fact that the an estimator is not an ideal one, the deviation values are also considered where the deviation was calculated using the directional self information value and outlier tolerance percentage. This provided with a threshold value and range depending in which any motion vector was classified as an inlier or an outlier. The final inlier map was later used for the calculation of the motion parameters’ matrix. To tackle the problems related to zooming sequences, the concept of clustering was used where the motion vectors was clusters into different clusters and the method was separately applied to different clusters and the final results were averaged. The K-means clustering was used for the clustering of the data. The method thus helped in reducing the computational time as in made use of one iteration per motion vector as opposed to state of the art techniques which carried out the estimation process in an iterative form. The method was also flexible when the question of thresholding arises as the value of outlier tolerance is dependant on the user and can be of any value. The proposed method is compared with other state of the art methods and it gave a better results than these methods both in terms of computational time and accuracy. This method outperforms the other methods in terms of number of iterations, thus significantly decreasing the total running time. The method reduced the computational time by a considerable amount of about 90% while still maintaining the same accuracy as the other state of art methods.