Multiview Video Coding Accelerated on Multicore Architectures

This thesis deals with the design and implementation of extremely parallel fast motion / disparity estimation algorithm for multicore architectures. Currently, H.264/AVC is the most widely used commercial video compression standard and is based on single view. Recently, Multi-view Video Coding (MVC) has also been standardized as an extension to H.264/AVC for supporting 3D and Free Viewpoint video. As MVC is an extension to H.264/AVC, so it achieves compression not only by exploiting temporal and spatial prediction but also exploits inter-view redundancies using motion estimation tool. In H.264/AVC, motion estimation is the most important tool employed by the video encoder to mitigate temporal redundancies but it is also the most time consuming. Consequently, in MVC, the time consumed for efficient encoding is even higher as the encoder has to perform temporal as well as inter view predictions. This thesis proposes a parallel low-complexity rate-distortion optimized motion/disparity estimation algorithm that can be implemented on multicore architectures such as Graphical Processing Unit (GPU). Recently, GPU has emerged as a commercially viable multicore platform for accel- erating computationally extensive applications and has also been applied for improving video encoder performance. Generally, the bit rate cost during motion vector calculation is ignored while implementing parallel motion estimation algorithms on GPU, due to the unavailability of the spatially predicted motion vectors, which leads to rate-distortion performance degradation. The proposed approach is able to perform the complex prediction task by means of an efficient distribution of all the computations over the GPU by mitigating the spatial dependencies. The experimental results show that the proposed scheme achieves significant speedup and has comparable rate-distortion performance with respect to sequential fast motion estimation algorithm. The proposed algorithm is also used for exploiting inter-view prediction in MVC and is implemented on the GPU exploiting view and block level parallelism simultaneously. The results for MVC suggest a significant speedup with negligible loss in coding efficiency.