Research On Advanced Video Stabilization Technology

With the popularity of mobile devices such as smartphones,people can shoot and share videos anytime,anywhere.However,the videos captured by the handheld camera device suffer the problem of jitter,which will cause discomfort to the video viewer.Therefore,it is necessary to perform stabilization processing on the jittered video,that is,video image stabilization.In addition,video stabilization technology can be as preprocessing of other video processing processes,such as object detection,video compression,etc.,to improve the accuracy and robustness of these tasks.The current video image stabilization algorithms can handle jittery videos to some extent,but there are still some problems.In addition,with the development of virtual reality technology and the popularity of panoramic cameras,panoramic video also faces the problem of jitter,and due to the characteristics of the video,it is more difficult to handle.The main task of this paper is to study sensor-based video stabilization algorithms and video stabilization algorithms for panoramic video characteristics.The research work includes:This article introduces the image-based video stabilization algorithm and sensor-based video stabilization algorithm and their advantages and disadvantages.Based on the sensor-based video stabilization algorithm,we apply the deep learning method to the field of video stabilization.We propose a video stabilization algorithm based on multi-modal neural network,using sensor's information to calculate the camera's motion path.During the path optimization process,we train the multi-modal neural network to predict the smoothing result of the corresponding frame.This paper mainly introduces the data set processing,label generation,loss function definition,and model construction of the neural network model,and compares performance from path smoothness and video stability through experimental data.This algorithm can remove the jitter in the video better,and can achieve the effect of real-time processing.The calculation is simple and robust,and it is easy to integrate into mobile terminals.Focusing on the characteristics of panoramic video,we propose a viewpoint-adaptive panoramic video stabilization algorithm,using 3D models to describe the motion path.Then,we find the optimized path by solving the optimining problem,adding the directional constraints.This way,we can make the saliency area presented at the positive direction as much as possible,that is,in the user's field of vision.Experiments prove that the image stabilization algorithm in this paper can eliminate the jitter in the video,and can present significant areas to video viewers,further improving the viewing experience.