|
Abstract:
|
Video conferencing and digital television applications use compressed video streams. In order to achieve required compression ratios for real-time operation, not only the performance of the data processing systems must be increased, but also the compression methods must be developed. Motion estimation is the most computationally demanding task in video compression and, therefore, several implementation alternatives are given in this thesis. Current video compression standards define block-based motion estimation methods to be used. This thesis describes three different approaches to achieve standards compatible motion information. Block-matching approach is the most well known and commonly used. Differential method and Fourier method are often overlooked due to more complex implementation in either hardware or software. However, these methods offer more sophisticated approach for motion estimation and they are very promising especially in the future when the technology develops.Four different block-matching algorithms, spiral and raster-scan full search, three-step-search and its improved version, new-three-step-search, were implemented on Texas Instruments TMS320C62x VLIW processor. Implementations are explained in detail and the bencmarking results are given in three different external memory configurations. With fastest memory type and QCIF frame size a maximum of 21 encoded frames per second can be achieved. In practise, this means that real-time operation cannot be achieved. Performance analysis shows that the performance is mostly limited by the small size of internal program memory. In addition, the incapability of processing 8-bit data and too fixed functional units limits the performance of the processor. /Kir10 |