Head Movement Prediction using FCNN

Viewport adaptive streaming of 360-dgree videos relies on accurate prediction of the viewport, while the user generally suffers from significant quality degradation under long delay settings. To deal with this issue, advanced methods for long-term viewport prediction are highly desired to improve viewport prediction accuracy. To more accurately capture the non-linear relationship between the future and past viewpoints, this paper proposes a Fully Connected Neural Network (FCNN) model to make future predictions, which is light in computation. The input data such as yaw values, pitch values, Estimated Weighted Moving Average (EWMA) of yaw values, and EWMA of pitch values, are transformed into sine and cosine angles before feeding into the encoding layer of the FCNN model by considering the roll angle to zero. After transforming the data input into the proposed FCNN model, a long-term prediction length of up to 4 seconds has been explored, to capture the non-linear and long-term dependent relation between past and future viewport positions more accurately. Experimental results show that the proposed scheme performs well for the large size prediction window.