Motion Estimation in h.264 encoder
Download as PPTX, PDF2 likes324 views
This document discusses video compression techniques. It begins by explaining the large file sizes required for uncompressed video frames and motivates the need for compression. It then describes some key techniques for video compression including spatial and temporal correlation, temporal modeling using predicted and residual frames, block-based motion estimation, and motion compensation. Specific algorithms discussed are mean absolute difference, mean squared error for cost functions, and adaptive rood pattern search for motion estimation. Examples of motion between video frames are provided. The techniques aim to reduce file sizes by removing spatial and temporal redundancy between frames.
Motion Estimation in h.264 encoder
- 1. NS Talal Khaliq Project Supervisor: Dr. Shoiab A Khan
- 2. Outline
- 3. Motivation ‘A picture is worth a thousands words.’ If this holds true, how a moving picture (video) which contains so much information is transmitted so efficiently?
- 4. Problem background Example an single video frame with 720x576 pixels with color depth of 24 bits per pixel with 29.97 frames per second uses approximately 200Mbs, thus for a two hour program at this rate takes over 200 GB which is practically impossible to store.
- 5. What is Video Compression? It refers to reducing the quantity of data, and is a combination of spatial image compression and temporal motion compensation.
- 7. Temporal Model It reduces redundancy between transmitted frames by forming a predicted frame and subtracting this from the current frame. The resulting residual (difference) frame contains less energy. The residual frame is then encoded.
- 8. Block-based Motion Estimation This method is used to ‘compensate’ for motion of rectangular frames or ‘blocks’ in current frame. It involves finding a 4x4 sample region in a reference frame that closely matches the current macroblock. Macroblock with minimum energy is chosen as ‘best match.’
- 9. Cost Function Mean Absolute Difference(MAD), Mean Squared Error(MSE), where N is the side of macroblock, Cij and Rij are the pixels being compared. 1 0 1 0 2 || 1 N i N j ijij RC N MAD 1 0 1 0 2 2 )( 1 N i N j ijij RC N MSE
- 10. Motion Compensation The selected best matching region in the reference frame is subtracted from the current macroblock to produce a residual macroblock. This residual macroblock is encoded and transmitted together with a motion vector describing the position of the best matching macroblock. Motion vector is the offset between the current block and the position of the candidate region.
- 11. Past Frame Current Frame Frame Segmentation Blocks Search Threshold Block Matching Motion Vectors Motion vector Correction Blocks Prediction Error Transmission
- 12. Example Video Frame 10 Frame 11
- 14. Adaptive Rood Pattern Search Algorithm General motion in the frame is usually coherent. It uses the motion vector of macro block to its immediate left to predict its own motion vector. It directly puts the search in an area where there is a high probability of finding a good matching block.
- 15. Predicted motion vector is (3,-2) and step size S, S=max(3,-2)=> 3.
- 16. Frames Macro block area defined Frame Scan S=max(|X|,|Y|) SDSP Calculate min cost LDSP Start loop again Motion vectors
- 17. Advantages We do not have to compute whole frame like in Exhaustive Search. It does not waste time doing LDSP. It starts with SDSP unlike in Diamond Search. It does not always start from centre or extreme left and thus saves computation time.
- 18. Video 2 Frame 110 Frame 113
- 20. Video 3 Frame 220 Frame 222