Study On Technologies Of Moving Targets Localization Using Binocular Stereo Vision In Haze Environment

With the raidly development and widely application of positioning technology,the society’s demands for high-precision localization are increasing.However,when there is no satellite signal,such as complex environments and indoor environments,GNSS technology cannot play a role.Due to the simple convenient uses,low costs,and highaccuracy of localization,vision sensors are widely used.Visual localization has become a research hotspot.Presently,the visual localization technology is widely used in the motion platform in active self-positioning,but the studies of passive localization are quite rare.The realization of highly robust and high-precision passive localization algorithm of the moving targets has important implications to the situation where the locating targets can not carry a sensor and the acquisition of high-precision localization information.This dissertation aims at realizing a high accuracy binocular visual passive localization of moving targets in haze environment.The key technical problems and theoretically explored have solved.A number of research results have been achieved in image defogging,moving object detection,stereo image matching,and targets localization:（1）To solve the problem of many complex disaster environments such as fires,earthquakes,and underground accidents,which are accompanied by smoke and dust that greatly affect the image quality and cause great interference to subsequent targets positioning.An end-to-end learning fully convolutional network is developed.We use residual learning to simplify the learning process.By constructing a network architecture of a fully convolutional and two fully connected layers network,then combining （?）₂ loss and perceptual loss as the linear loss function of the residual mapping,the performance of network training is improved and the "Artifact" problem is reduced.The algorithm directly learns the residuals between the haze image and the non-haze image,without estimating the atmospheric light value or medium transmission,which overcomes the constraints of the traditional dehazing method by the atmospheric scattering model,realizes the end-to-end image dehazing.（2）To solve the problem of moving targets detection by light changes,scale changes,shape changes and angle changes of the targets,"ghost" problem,and fixing or moving camera interferences.This dissertation studies the moving targets detection and extraction algorithm based on optical flow estimation.First,we start with the estimation of optical flow field of current images,analyzes the data term and smoothing term of the optical flow energy equation.The anti-interference LK optical flow method is combined with the HS optical flow method to form the data items of the optical flow equation according to reduce optical flow residual effects,then a high-precision optical flow estimation equation is constructed.We build a image pyramid to estimation optical flow with large displacement.The constructed optical flow energy function is solved by internal and external nesting loop.According to the Delaunay triangulation occlusion determination,the "ghost" in the detection of moving objects is processed,and a high-precision image dense optical flow vector field is obtained.Second,we perform threshold segmentation on the obtained optical flow vector field to obtain the foreground area of the moving targets.In order to solve the problem that the traditional OTSU optical flow thresholding algorithm cannot extract the foreground area of the moving targets when the camera is not fixed,we find that the background area of the image sequence acquired when the camera is moving conforms to the law of image affine transformation.Based on this,we propose to extract the feature points in the sequence of images before and after the frame and calculate the affine transformation matrix of the background model between the two images,and then delete the feature points in the foreground area because of the moving targets do not conform to the background model,then obtain feature points of the background area satisfying the affine transformation matrix,which are called robust feature points.We use the optical flow value at the robust point as the optical flow threshold to calculate the binarized image of the foreground area of the moving targets.The algorithm we proposed can be applied to both the fixing and moving background scence,which have high-accuracy detection and extraction results of rigid moving targets or non-rigid moving targets.（3）In multi-moving targets localization detection,most of the existing moving targets detection and extraction algorithms based on optical flow method only preliminarily detect and segment the foreground regions of all moving targets together,and do not further refine and segment different moving targets.As a result,multiple targets will appear in the foreground regions of preliminarily extracted moving targets,which cannot realize the recognition of single targets or the subsequent multi-targets localization.The further research is carried out on the basis of the realization of the moving targets region detection and extraction algorithm based on high-precision optical flow estimation and robust feature point optical flow thresholding processing,in order to realize the refined recognition and region segmentation between different targets.Through analysis,we find that the magnitude and direction of the optical flow of different moving targets are quite different.Based on this,we analysis the optical flow histogram and K-means clustering analysis on the optical flow vector field in the foreground area of the moving targets extracted based on the high-precision optical flow estimation and robust feature point optical flow thresholding algorithm.Stream clustering analysis to segment and extract different moving target foreground regions.Experiment results show that the multi-targets segmentation algorithm we proposed has high-precision recognition and extraction results for rigid moving targets,and it has a better clustering and segmentation effect for non-rigid moving targets with small differences in optical flow within the same moving targets.（4）To solve the problem of stereo matching between the deformation of the targets area and the repeated texture area in the stereo matching,and to improve the accuracy and efficiency of the stereo matching.This dissertation combines Deep Matching with the stereo matching constraint criterion.First,we calculate the matching bottom-level association graph,which similarity measure is expressed by feature descriptors.Second,through down-sampling,pooling operation,translation transformation and image regularization to build a pyramid on the degree of image relevance from the bottom up。Then trace back from the top of the pyramid to the bottom to obtain a high-precision dense stereo matching point pair.Finally,performing post-matching processing to obtain the final stereo matching result.（5）The above algorithms are integrated and applied to the actual moving targets detection and localization experiments in simulated smoke environments.The results show that the series of algorithms proposed in this dissertation have good practical effects.