Study Of Electronic Image Stabilization Based On Feature Detection And Description For Vehicles

The bumping and shaking of a running vehicle will seriously affect the quality of the videos collected by onboard camera system and thus go against the observation and interpretation of decision information in the images.There is an urgent theoretical and practical need for the platform to remove the video jitter,improve the video quality,and process the blurry images through fast and effective jitter compensation and smoothing & stabilization.In order to mainly solve the problem of video jitter faced by the onboard platform,this paper analyzes the origin of fuzzy video sequence according to the imaging principle and motion mode of onboard camera system,and studies in depth the basic theory and key technology of electronic image stabilization method for vehicle.In addition,in view of the features of onboard image stabilization method and the problems in current electronic image stabilization technology,this paper designs an effective real-time image stabilization method to stabilize the jittering videos.The main research efforts and results of this paper are as follows:1.The mechanism and characteristics of onboard electronic image stabilization system have been analyzed in detail.Firstly,the influence of camera shaking on image plane was analyzed.Then the coordinate transformation models of different motion modes in 2D images were introduced.The homography matrix for global motion vector algorithm in the affine transformation was deduced.The evaluation criteria of image stabilization result were presented.In addition,the key technology of electronic image stabilization method was introduced,and the main problems in current image stabilization technology were analyzed.2.In terms of the image stabilization technology of onboard platform,the dejitter method based on feature extraction and matching was examined in depth,and the contradictory relation between real-time performance and effectiveness was analyzed for the image stabilization algorithm.As the real-time requirement of onboard platform is more stringent,the image stabilization method we selected is based on binary feature extraction and matching.In order to deal with the problems of low robustness and high mismatch rate that were commonly seen in the binary feature algorithms,this paper proposed a feature extraction strategy combining brightness information with extremums,and adopted the nonlinear diffusion filtering for preprocessing to sharpen the edge information in the image.This method can improve the efficiency of feature extraction by quickly segmenting potential feature corners by brightness,calculating the gradient variation in their surrounding pixel regions,and then rapidly identifying those feature corners with strong significance.The experimental result shows that,the feature points extracted by this method are more likely to appear on the boundary between two scenes so that they are unlikely to deform or disappear and easy to identify.This can effectively improve the algorithmic robustness and adapt the algorithm to image stabilization design in the videoing process of diversified all-weather scenes.3.The onboard platform had only two degrees of freedom(pitch and azimuth)but was well-conditioned for real-time video frame processing.Therefore a "breadth-to-depth" binary feature characterization method is proposed to intensify the differences between descriptors.Besides,the brightness contrast threshold of descriptors was also added to further improve the descriptor self-significance.Furthermore,the proposed method removed the rotation invariance of the descriptors based on a binary feature extraction and matching algorithm.The experimental result shows that,while meeting the real-time requirement,the proposed method establishes the relation between features and their surrounding pixel regions layer by layer and thus dramatically enhances the effectiveness of binary feature algorithms.4.By considering the need for long-time target monitoring in a designated area and the two disadvantages of onboard mobile camera,namely small lens aperture and limited focal length,this paper designed an image stabilization method based on target trajectory,which could adapt to main motion states featuring small movements and minor scope variation.Firstly,at the feature characterization stage,brightness information was used to quickly segment the features and create a binary feature descriptor with strong significance.Then,a sparse search template was proposed at the feature selection stage to further improve the efficiency of algorithm execution.Finally,a static library for storing initial information on the target and a dynamic library for continuously updating the information on the target's appearance or motion variation were built.Through the comparison of the information in static and dynamic libraries and in the search template region,the position of the target of interest was identified,and a reference image-stabilization method for monitoring the trajectory of changing target position was established.The experimental result shows that,the proposed image stabilization method can effectively improve the accuracy of target information extraction,monitor the target position for a long time,and output a stable video sequence.5.An image stabilization system was built for smart mobile device on IOS and Android embedded systems.To meet the real-time requirement of smart mobile devices upon the analysis on its characteristics,the features were quickly segmented through the brightness contrast of grey-scale images,and the binary feature descriptor enabling fast depiction and matching was utilized.Then,to enhance the execution efficiency of image stabilization algorithm,the method of feature position re-inspection was used,through the limitation on the positions of feature-point matching pairs,to weed out wrongly matched feature point pairs afar and improve the estimation accuracy of global motion vectors influenced by the local motion of a moving target.The experimental result shows that,under the condition that the device processor has a limited computing capacity and a high real-time requirement,the proposed method is robust enough to apply to smart mobile devices with a limited computing capacity.