Research On Video Stabilization Technology By Feature Tracking

Electronic image stabilization is a new generation of image stabilization method, which employs the computer image processing techniques to stabilize the dithering video sequence. Because of its high precision, small size, light weight, low power consumption and low cost, it has found wide application in both the military and civilian areas. The key step is motion estimation and motion compensation. The current researches focuses include how to improve the efficiency of motion estimation as well as precision, how to judge the camera shooting scenes and random jitter, so that motion compensation can be done stably for real-time video stabilization.The fundamental principles and the main methods of electronic image stabilization algorithms are first discussed. Based on our researches on Scale-Invariant Feature Transform (SIFT), a new electronic image stabilization algorithm is proposed which employs the SIFT to estimate the parameters of the global motion. As SIFT feature points are invariant to image rotation, scaling and translation, the proposed scheme can handle the situations where there are both rotation and scaling in the initial video sequence. We also investigate the feature point extraction techniques based on wavelet transform and its application in electronic image stabilization systems. A new electronic image stabilization scheme is then presented based on wavelet feature points. In order to achieve better correspondences between the feature points, the direct information of the feature point neighboring regions are taken into consideration. The incorrect pairs of feature points are dismissed using the distance criteria. Finally, global motion estimation is done using the correctly matched feature point pairs, and the stabilized video sequence is obtained after motion compensation. As the direction information of the feature points are considered, the proposed scheme achieves better performance than traditional feature based schemes. Besides, the distance criteria also further dismiss the incorrect feature point pairs. Experimental results demonstrate that the proposed schemes can efficiently reduce the camera dithering.