Motion Estimation And Video Compression Algorithms Based On Regional Features

With the rapid development of information technology and instant messaging in recent years,video traffic now accounts for more than 75%of all Internet traffic.The newest and most efficient video coding standard HEVC(High Efficiency Video Coding)which was raised by JCT-VC and finalized in 2013 has shown its great advantages over H.264/AVC that is the most popular video coding standard around the world now.There are some disadvantages still exist in HEVC as it follows the traditional block based hybrid coding framework.It gives a brief introduction on the history of video coding standards and some other video and image coding algorithms based on features which were raised by others and showed their good points.Some new algorithms were proposed in this paper based on previous work.Firstly,a coarse motion estimation method was adopted for the difficulty in balancing the search range and the computation complexity when compressing HD,UHD or higher resolution videos.The redundancy when applying motion estimation method to blocks in the same area can also be greatly reduced by this algorithm.Instead of using difference between all the pixels in the areas as matching criteria,the match scores between scale and rotation invariant features which show greater robustness than original SURF features are adopted.The percentage of matched features in all features can be very useful as it takes a tumble when scene change happens.Then,an improved RANSAC regression algorithm with error correction method was raised to segment the areas with different type of motion.The experiment results show that the algorithm proposed in this paper is better than RANSAC when applying to video sequences which have several moving objects.The insufficient of applying HEVC intra-frame estimation method to flat area with gradual change pattern were observed.A novel algorithm which describes flat area well based on the observation of the distribution patterns of pixels and gradients were explained in this paper.The experiments showed that this algorithm overcomes the intra-frame estimation method adopted by HEVC when used as predicted image generator in flat areas with gradual change pattern.On the basis of the above works,a video sequence compression framework based on regional features of video sequences and a new quantization matrix are proposed in this paper.The experiment results show that the motion estimaton method adopted in the framework can accelerate HEVC encoder without reducing quality.The compressibility of residual images of algorithm proposed in this paper was tested.The experiments show that the algorithm is competitive with HEVC.Due to the complexity of extracting features from frames,the coarse motion estimation method cannot be done fast enough.Therefore,an accelerating algorithm running on CUDA platform was proposed in the paper.Thanks to the shared memory which has extreme low latency when communicating with cores within the same SMX,the algorithm accelerates the feature extracting process for at least forty times and keeps the same accuracy as on the CPU platform which is not the normal way when using CUDA platform due to the poor double-precision floating point performance of NVIDIA GPUs.The experiment shows that extracting all the features of one frame which resolution is 1080p costs less then 100ms under normal circumstances.