A Parallel Image Stabilization Algorithm Based On CUDA

Nowadays more and more people enjoy using digital camera or smart phone to take pictures and videos. When using these devices for image recording, a shake or jitter may happened and this may cause the video becomes wobble and blurred. These unstable videos have many negative effects for user observation, image processing, and data storage. So it’s necessary to stabilize the video. At the other hand, the performance of the GPU has a great improvement with the development of integrated circuit. GPU has a great advantage than CPU when processing image data or doing data-intensive computing. According to above background, this paper proposes a parallel electronic image stabilization algorithm which makes full use of the ability of GPU in parallel computing to achieve real-time image stabilization.In order to stabilize videos with using GPU, it’s necessary to consider the performance, complexity and parallelizability of the algorithm when deciding the methods of each step in the process of image stabilization. Image stabilization contains two parts:motion estimation and motion compensation. In this paper, corner as a feature of image is used for motion detection. Harris corner detection algorithm is used in the process. Modified Hu-moment is used to describe the feature of each corner and corner’s match between two images will use it to implement the match process rapidly and effectively. Choosing bi-direction matching algorithm to match the corners of two adjacent images and using RANSAC algorithm to remove the mismatching corner pairs. Use similarity transformation model to describe the motion between two images and then calculate the local motion parameters. When estimating the global motion, Gaussian filter is selected to filter the motion parameters with appropriate length and center point of the filter window and thus high frequency components of the movement is filtered out and effective movement is retained. In the motion compensation part, bi-linear interpolation algorithm is select in this paper to get the image after image stabilization.All these steps of image stabilization are fully parallelized when implement these algorithms in the NVIDIA’S CUDA platform. A thread will process a pixel when detecting the Harris corner or compensating the motion; during the process of corner match and mismatching corner pairs removing, a corner will use a thread. After accomplish the design, Test samples in different resolution is used to compare the speed of process for image stabilization between CPU and GPU.Experimental results show that the algorithm for image stabilization in this paper has a good performance when deal with the video which exists rotation, translation and slightly zoomed, and the parallelization electronic image stabilization algorithm in the paper has good acceleration effect.