A Study On Target Recognition And Location Based On Feature Matching

Target recognition and location which recognice the interested objects in image and estimate the camera pose are the important part of Augmented Reality(AR).In the AR system, The computer-generated virtual objects or information was superimposed into the real scene to realize the reality "enhanced". In order to ensure registration accuracy and dynamic consistency of the virtual objects and real scene,real-time tracking of the camera position and orientation is needed.Owing to the complexity of the reality world, the first problem of the target recognition and pose estimation process to be solved is to extract efficient and stable matching points. Use the matches to identigy the same or simlar objects, estimate the transformation matrix between two adjacent views. Obviously,the accuracy and efficiency of feature matching affect the accuracy and real-time of target identification and location methods directly.It can be clearly seen that feature matching has become a key technology for identification and location.Currently,the methods of feature description and matching can be divided into two categories,one is feature matching based on local invariant,the other is feature matching based on statistical learning. However, there are some problems in feature matching and pose estimation following in three aspects.(1) Local invariant feature includes not only the own information of the keypoints,but also joints the pixels contributing to the keypoint around it to build a high-dimensional feature vector descriptor.The higher the dimension,the better the robustness,but the higher computational complexity. SIFT descriptor is128-dimensional feature vector. Obviously,the high dimension of such operators causes large computational burden,slow computing speed and take up a lot of memory space. In the case of tens of thousands of features in the environment,local invariant feature will not meet the requirements of real-time systems,as well as unfavorabnle for the expansion of real-time AR system on the mobile terminal.(2)The calculation of feature description is transferred to the offline training of the classifier in the feature matching technology based on statistical learning which has the advantage of low computational.However, these algorithms are with the strong randomness and poor stable because of the effection by the classifier performance. Random Ferns was proposed by Ⅴ.Lepetit etal in2009, replacing the trees by the non-hierarchical ferns and pooling their answers in a Naive Bayesian manner yields better results and scalability in terms of number of classes, we need to assume that the characteristic properties are independent,and divide them into several groups randomly. Because the pixels in the image patch surrounding a keypoint are not independent from each other,blindness grouping ignores the different significances of different combination of attributes contributing to the classifer.Inapproopriate division of attributes will affect the quality of the classifier,which eventually lead to low recognition rate.(3) When the camera internal parameters are known,estimating the camera relataive pose with the correspondence between the space features and its projection is the PnP problem discussed in this paper.The PnP algorithm can be divided into two types of iterative solution and non-iterative solution.The non-iterative algorithm directly uses algebraic method to calculate the pose of a camera.The non-iterative algorithm has little computation and fast speed,but low precise.So it is often used in the initial estimate of the iterative algorithm.The PnP problem is expressed as a constrained nonlinear optimization problem to obtain the numerical solution of the camera pose by solving the optimization problem.the algorithm has high precise. However,it is limited by the poorly initialized and a great deal of calculation is needed.Therefore, a compromise approach for the PnP problem is critical for the achievement of rapid and accurate target location.To solve the above problem,the research tasks can be summarized as follows: First,simplify the local invariant feature matrix extraction,as well as using efficient data structure and searh query method to improve the speed of matching; Second,develop feature matching technology based on statistical learning to reduce the time complexity, and provide meaningful information to guide the properties divided,continuously improve the rate of feature recognition and the accuracy of matching; Last but not least,design camera pose estimation algorithm with high precision and less computation, and minimize the impact of initial value.In this article, we discuss target recognition and location and research camera pose estimation methods based on feature matching, The main work of this paper are summarized as follows:(1) Fast feature matching based on local invariantAccording to the problem of high computation burden caused by the high strength description of the local invariant feature, Multi-dimensional binary index tree(Kd-tree) is used to establish multi-dimensional index,and nearest neighbor search algorithm is used to query point from the nearest data points,that is the match point.The introduction of Kd-tree greatly facilitate the search operation of high-dimensional feature matching.Experiments show that the query results are not necessarily matches. In this paper, the mismatching points are removed by the Random Sample Consensus(RANSAC) algorithm. The examples of image mosaic demonstrate the effectiveness of this feature matching methods.(2) Random Ferns feature matching based on conditional mutual informationIn this paper,conditional mutual information is introduced to guide the combination of characteristic properties, we proposed a Random Ferns feature matching algorithm based on conditional mutual information called MI-Fern. The main contents of MI-Ferns include:access to the training sample set, characteristic properties division, offline classifier training, feature recognition and matching. Offline training of the Navie Bayes Classifier is so crucial that the training result directly affects the oneline feature matching.The MI-Ferns is proudly ways to improve classifier performance to overcome the blindness of Random Ferns.In addition,the characteristic division based on mutual information has on effect in the time consumption of inline matching.(3) position and orientation estimation of calibrated cameraIn this article,we design a system which camera position and orientation is estimated by a fast and accurate PnP algorithm. The matches is extracted by MI-Ferns. The internal parameters of camera is calibrated by Zhang’s algorithm according to pinhole imaging model. A non-iterative PnP method is used to estimate the camera projection matrix. The central idea is to express the all3D points as a weighted sum of four virtual control points.The problem then reduces to estimating the coordinates of these control points in the camera referential.The solution with better accuracy and much lover computational complexity than non-iterative methods,much faster than iterative ones but with loss of accuracy. The Gauss-Newton scheme being introduced to increase the accuracy of the solution yields the same accuracy as the best iterative algorithm proposed by Lu etal in2000.The experimental result indicates that the3D-2D point correspondences obtaind by the improved Random Ferns algorithm can be used to determine the position and orientation of a camera.The affine transformation matrix obtained by RANSAC algorithm maintains reliability when locating the object position.In the process of camera movement,the image projection matrix can be derived from geometry view constraints to provide accurate registration location for the virtual objects superimposed exactly.