Robust Model Fitting Method And Application For Camera Localization Via Parameter Estimation

Model fitting is a basic research in the field of artificial intelligence.It involves the knowledge of image processing and pattern recognition and has been widely used in robot localization and automatic navigation,panoramic photography etc.The purpose of model fitting is to estimate the inliers and the parameters of the model from the input datasets with a large number of outliers.It is still a research hotspot for scholars to remove outliers and estimate the inliers of the model structures correctly.At present,many model fitting methods have been proposed by researchers at home and abroad,but there are still some problems when handling the complex multi-structure model with a large proportion of outliers.And the running time and accuracy can't meet the needs of practical application.In order to solve the problems above,a robust model fitting method based on spectral clustering to remove outliers is proposed in this paper.Combined with the application of model fitting in the parameter estimation for camera localization,a method based on deep learning for parameter estimation of camera localization is proposed.Outliers Removed via Spectral Clustering for Robust Model Fitting(ORSC)is an improvement method based on traditional preference analysis model fitting method.Most of the traditional methods require a separate step to discriminate outliers,which will result in outlier being misestimated as inliers.On the contrary,the inliers are mistaken for outliers to remove,which affects the accuracy of model fitting.In the proposed method of this paper,firstly,we give weights for the model hypothesis obtained by multiple sampling,and the invalid hypothesis is removed.Then,the concept space of similar matrix is established in the valid model hypothesis.In this space,the distance distribution between outliers and inliers is different.The inliers is farther from the origin of space than outliers.According to the distance distribution,the subspace classes can be automatically determined based spectral clustering,which saves computing time.In addition,we can use the clustering result to guide the follow-up sampling to get more clean data points for hypotheses generation,which improves the accuracy of fitting.The experimental results in line fitting,circle fitting and real image show that the proposed method is more robust than the traditional methods.Camera localization based on deep learning for parameter estimation is an important application of model fitting.Traditional geometric methods estimate inliers of camera pose based on SFIT features,by matching features between images and removing outliers.However,these methods are highly dependent on image texture information and have complex computation in feature processing.In this paper,camera localization based on deep learning for parameter estimation is proposed.The pose of the camera is regressed by a convolution neural network when input a RGB image.The 6-DOF pose vector of camera localization is output.This convolution neural network is an improvement method based on PoseNet.By Batch Normalization,the problem of gradient disappearance is solved,and the training time is faster.In addition,break the large convolutions into small convolutions,reducing the computational complexity.The validity of this method is proved by the experiment on the public datasets.