Annular Panoramic Perception Of Multiple Cameras For Multimedia Sensor Networks

The sensor node used in traditional wireless multimedia sensor network(WMSN)has only one camera,thus it has a limited perception area and cannot achieve omnidirectional observation at the same time.The observation network composed of multiple visual sensor nodes has disadvantages of covering imbalance and data transmission redundancy.So it is necessary to introduce panoramic technology to enable the sensor node to have the ability of omnidirectional perception.This article makes a research on quick panoramic image synthesis with the need of WMSN in order to make the omnidirectional perception of the sensor node of WMSN come true.This thesis firstly describes the image registration algorithms and image blending algorithms based on multiple cameras.Detailed analyses are made on the application scopes of image transformation models and image projection models.After comparing the area-based registration method and feature-based registration method,the image registration algorithm based on feature points is used in this thesis.With the pros and cons analyses of blending algorithms in time domain and wavelet domain,the linear fusion method is chose because of its better performance in time consuming.After choosing the registration method,this thesis verifies two representative feature point detection algorithms,Harris algorithm and SIFT algorithm,on detecting number and matching effect of the detected feature points.The result shows that SIFT is better and as a result it is used in the thesis.The coordinates of the feature points are normalized in order to avoid wrong identification when applying RANSAC.This thesis introduces Moving DLT algorithm to decrease registration error by parallax when analyzing the deficiencies of the registration method based on the global homography.Meanwhile,on purpose of maintaining the perspective projection relationship between the image and object scene,SPHP algorithm is also introduced.In order to combine the advantages of Moving DLT and SPHP,the paper modifies SPHP,and its results manifest that the modified algorithm can both decrease the registration error and maintain the viewpoint of the image.After the analysis of the influence factors of real-time in video stitching,dual-liner blending algorithm is chose because of its better blending effect and low complexity.A new weight function is invented on purpose of being used in the images transformed by the modified SPHP.The paper designs a new look-up table method called" Inverse Look-up Table Way Based on Forward interpolation",which converts registration to pixel carry and blending to weight finding.Experiment results show that the panoramic image stitching algorithm proposed in this paper has favorable performances on registration accuracy and blending effect.In C++ platform,it cost only 26.49ms to produce a panoramic image with a resolution of 1544x372 from five images with a resolution of 648×364,which can meet the requirements of real-time panoramic video output.