bachelor thesis

SignatureBA 022
AuthorChen Liang
TitleInterpolation of dense motion vector fields from block-based motion vectors
TutorM.Sc. Marko Esche
ProfessorDr.-Ing. Thomas Sikora
AbstractIn hybrid video coding the block based motion vector is estimated and used the prediction of future frames. Together with the quantization of DCT-coefficients this leads to the introduction of block artefacts, which will impair the quality of decoded videos. Temporal trajectory filtering (TTF) is a way to reduce these artefacts. As a result, it is necessary to estimate individual motion trajectories of every pixel from block based motion vectors. Limitations of the method also exist because of the quantization of motion vectors and the resolution of images.
With the help of approaches from the field of Computer Graphics, a motion vector field will be constructed using an interpolated form of the transmitted motion vectors. At the same time, calculating time is also an issue to be considered. After the interpolation, the reconstructed motion vectors are evaluated by using a human-assisted motion annotation tool which basically compares our motion vectors with those of the same video sequence by calculating the mean-square-error. After getting satisfactory results, the reconstructed motion vectors are then further compared with original motion vectors and also motion vectors generated with the nearest-neighbours-method.
The reconstructed motion vectors derived through interpolation can better represent the motion trajectory compared to the original motion vectors, which of course will lead to an improvement of the temporal trajectory filter (TTF). This is validated within the HEVC-reference-encoder.
Key wordsHEVC, TTF, temporal trajectory filtering, motion estimation, motion compensation, implementation, motion vector evaluation