The Research Of Real-time Electronic Image Stabilization Using CUDA

The purpose of electronic image stabilization(EIS) is to extract the unstable jitter components from the video, and then obtain the stable video output by using motion compensation to reduce or eliminate the abnormal jitter. EIS system is mainly composed of three parts: image preprocessing, motion estimation and motion compensation. Among them, the estimation of motion vector between frames is most importtant and time consuming. Whether we can achieve accurate and rapid motion vector estimation largely determines the performance of the EIS.With the advent of the graphics processing unit(GPU), it becomes possible to accelerate algorithm by combining the hardware technology and software together. In view of the strong data processing ability and good programmability of the compute unified device architecture(CUDA) technology proposed by NVIDIA company, this paper focuses on the C UDA parallel programing approach and discusses its working principle, system structure and parallel processing mechanism in detail.In this thesis, we first introduce the application prospect and development status of the EIS technology, and describe its fundamental theory, the basic methods and processing procedure in detail. This paper proposed a novel EIS algorithm by combining the SURF feature detection and Kalman predictor together. The data with higher calculation intensity is accelerated with the CUDA parallel programming. Moreover, based on the features of program module itself and the pending data, the specific optimization is applized in the program. The optimazed CUDA program' time consuming is greatly reduced and successfully achieves real-time processing.Due to the high-precision of motion vector estimation, the optical flow method has higher study value in both motion estimation and object detection fields. However, it is difficult to achieve real-time processing due to the large amount of data and worse realtime character, which greatly reduces the practical application of the method. This paper realized the optical flow approach with the CUDA parallel programming technologh, largely improved the the processing speed, and layed the foundations for its application in the future.Simulation results show that the proposed EIS can achieve obvious stabilization results for actual videos. Comparing the algorithm's running time in GPU and CPU, it can be seen that the program after optimization with CUDA has a better time accelibration and can be realized in real time. For the case of optical flow method, its CUDA parallel program runs 15-25 times faster than its C program.