Research Of Monocular Vision Based Target Tracking And Positioning Techniques

Monocular vision tracking and positioning techniques, which have been widely applied in various industries belong to the field of navigation and positioning of moving targets with the advantages of simple structure and fewer calibration steps, collect image sequence of moving target continuously with a camera and then detect, track and locate the moving target from image sequences by using image processing methods and analyzing the nature or artificial features of the images. Monocular vision tracking and positioning is an important research field in computer vision, combining many other related fields including computer image processing, pattern recognition, artificial intelligence, and automatic control and so on. For the actual demands of the automatic alignment system developed for the refueling drogue of the tanker aircraft and the receiver’s refueling receptacle aligning, this paper studies the tracking and positioning problems of the moving target in monocular image sequences,using machine learning and pattern classification theories.Construction of the moving target tracking and positioning system based on monocular vision includes two modules: target tracking and target positioning, of which target recognition, classification and retrieval constitute the bases. But for object recognition, classification and retrieval algorithms, the recognition rate and retrieval accuracy are sensitive to light, easily affected by the environment, and so on. To solve this problem, this paper describes the sample images using image patches instead of pixels, discusses the feature abstraction method using both the small image patch and its contextual information to boost the recognition performance. In addition, for non cooperative target, because that the size of the image sample database increases exponentially with the accuracy requirements of parameter estimation, thus the training image database established for tracking and locating model has the problem of "dimension disaster". To solve this problem, this paper studies the image classification and recognition algorithm based on feature dimension reduction and Anchor graph based ranking, and discusses the image dissimilarity learning problem. Finally, this paper discusses the problem of moving target tracking and spatial locating and presents target tracking method based on image patches and target spatial locating method based on image classification and retrieval knowledges. At last, simulation experiments for aerial fighter automatic refueling alignment verify the conclusions presented by this paper.The main research contents of the thesis are listed as follows:Firstly, for the problem that the recognition rate of target is easily affected by target attitude rotation, occlusions and other environmental factors such as changes of light, a new local feature learning method called CLP is proposed. Using the Fisher discriminant criterion, this method can extract the local features effectively. Experimental results proved that classification algorithm based on CLP local feature has a higher recognition rate than other local feature method. In addition, image feature dimension reduction and similarity ranking paly an important role in image classification and retrieval. This paper presents a new framework for dimension reduing and classifying by combing with similarity leaning method after fusing NMF and PCA methods to reduce the high-dimension featuers of images, after considering the learning algorithm for Anchor hypergraph based similarity. This method improves the precision of the classification and retrieval performance. At last, experiments results verify this superiority.Secondly, for the feature similarity and dissimilarity problem during the target identification process, this paper presents a dissimilarity learning algorithm based on Adaboost algorithm. Combining the feature extraction and similarity rancking mehods, this new algorithm measures the different distribution of different feature points, constructs some candidate distance measures by using different distance measurement function and then combines different distance measures with AdaBoost algorithm. This new algorithm can match the optimal distance fuction to the selected feature to obtain the best similarity evaluation of two images.Thirdly, to inprove the real-time tracking performance of the moving target, this paper presents the expression method of target model based on statistical analysis of small image patches. Combining the k-means clustering algorithm and SVM classifier, this method uses small image patches instead of pixels and transforms the statistical analysis of color and gray information to statistical analysis of index values of quantified small image patches, which reduces the computation time of target tracking algorithm and improves the real-time performance of the system.Fourthly, this paper designs a system for online target tracking and positioning based on image classification and retrieval algorithm, combining the target tracking algorithem based on statistical analysis of small image patches and feature leaning menthods such as CLP method. Using the off-line well trained database that has been marked category and the distance symbols, this algorithm matches features between the real-time obtained images and the training database with image classification and retrieval algorithm, to obatain the the distance and the relative position between the target and the camera, which has advantages of high locationg precision and strong robustness.The research work of this paper can provide reference for our space flight target tracking and positioning research based on monocular vision.