Research On Target Recognition And Location Based On Local Invariant Features And Point Cloud Registration

With the continuous development of computer vision technology,the types of visual sensors are more abundant and the range of use is wider and wider.In recent years,the world's robot market has continued to expand,and the market share of intelligent service robots has grown significantly.The ability to identify and locate target objects is the basis for intelligent service robots to better serve humans and perform certain tasks independently.Binocular vision sensors are increasingly used in intelligent service robot systems because of their strong ability to acquire and adapt to external environmental information.Different from industrial robots operating in a structured environment,the working environment of intelligent service robots is more complex and faces challenges such as illumination,occlusion,and weak textures.Therefore,the robustness and real-time performance of target object recognition and location algorithms are improved.Sexuality is a hot topic in the current research field.The main research work of this paper is as follows:Firstly,we study the currently used local invariant feature matching algorithms,analyze and compare the rotation invariance,scale invariance,positioning accuracy and matching efficiency of each algorithm.For the disadvantages that the feature matching algorithm can not take into account the invariance and efficiency of matching scales,this paper analyzes the advantages and disadvantages of each algorithm,selects SURF algorithm and ORB algorithm respectively to improve,and studies the algorithm fusion,and the results of feature matching are wrong.Match points are removed to improve the matching accuracy.Using Grab Cut algorithm to segment the image of the identified target object,aiming at the problem of human-computer interaction in the image segmentation process,the feature matching algorithm is connected with the image segmentation algorithm to improve the automation degree of the algorithm.Secondly,based on the image segmentation of target object,the centroid of the target is calculated based on the image moment,the efficiency of centroid matching is improved by using the epipolar constraint,and the spatial three-dimensional coordinates of the target centroid is obtained by combining the internal and external parameters of the binocular camera.Based on the point cloud registration to obtain the target pose,the rigid point cloud registration model and the registration index are studied.For the common point cloud registration algorithm,the registration process is divided into the higher initial value requirements.Rough registration and fine registration are two phases.Based on the rough point registration of the four-point base,the selection conditions of the four-point base are constrained,and the number of equal bases is reduced to improve the efficiency of the coarse-fitting phase.In the fine registration phase,the points in feature-rich regions are extracted and the search process in the nearest point is accelerated to improve the accuracy and efficiency of registration.The two stages of registration were verified experimentally using standard data sets.Finally,calibration experiments are performed on binocular cameras and scene depth information is acquired by reading the left and right video streams.The effectiveness and versatility of the recognition algorithm are verified for different target objects in different environments and different pose states.In order to further verify the algorithm's ability of recognition and location under the actual robot operating environment,the UR10 robotic arm platform was set up,changing the pose of the target object.Robotic grab simulation experiments and physical experiments were performed to verify the practicability of the algorithm.