Improving motion vector prediction using linear regression

Abstract

The motion vectors take a large portion of the H.264/AVC encoded bitstream. This video coding standard employs predictive coding to minimize the amount of motion vector information to be transmitted. However, the motion vectors still accounts for around 40% of the transmitted bitstream, which suggests further research in this area. This paper presents an algorithm which employs a feature selection process to select the neighboring motion vectors which are most suitable to predict the motion vectors mv being encoded. The selected motion vectors are then used to approximate mv using Linear Regression. Simulation results have indicated a reduction in Mean Squared Error (MSE) of around 22% which results in reducing the residual error of the predictive coded motion vectors. This suggests that higher compression efficiencies can be achieved using the proposed Linear Regression based motion vector predictor.