Hardware Acceleration Based High-definition Video Realtime Stabilization Technique

Video Stabilization technology, also known as Image Stabilization(IS) is the technology to obtain stabilized images continuously from a certain acquisition equipment set on a platform. The current video stabilization technology either generates videos with obvious distortion, or has a high time processing complexity which leads to a low stabilization processing efficiency. In order to ensure the video stabilization effect, it is urgent to put forward an effective acceleration method to improve the efficiency of the existing video stabilization algorithm.In order to increase the speed of the video stabilization algorithm, this paper proposes a pipeline based tasks parallelization method and a CUDA based data parallelization method. Firstly, the paper analyzes every section of video stabilization algorithm by task parallelization method, then decomposes the whole stabilization process into several disparate tasks so that each pipeline segment only needs to handle one single task to complete the whole stabilization work. Therefore, the work that must be processed continuously can now be disposed simultaneously through the pipeline, whereby it will result in efficiency improvement of the stabilization system. Secondly, the paper analyzes the details of every pipeline segment of video stabilization algorithm by CUDA based data parallelization approach, and turns over the image data processing, which includes countless repeated calculations, to GPU. Thus the efficiency of each segment will increase since the GPU is far rapider than CPU while doing calculation works.Finally, the paper optimizes feature point trajectories based video stabilization algorithm by pipeline based tasks parallelization method and CUDA based data parallelization method, and implements a video stabilization system by programming languages such as CUDA and C, etc. The experimental results show that the processing efficiency of the video stabilization algorithm optimized by the two parallelization acceleration methods can reach a speed of 49 fps with HD video(1080P). In other words, the video stabilization system implemented in this paper meets the demand for highdefinition video real-time stabilization.