Recognition of Disordered Targets for Robot Based on 3D Visual Clustering and Matching

For the problem of the classification and location of flat workpieces with weak texture, high similarity and no significant normal angle feature in the sorting task of industrial robot, an algorithm about recognition of disordered targets for robot based on 3D convexity clustering and E-ICP matching is proposed, which provides the strategy of motion planning for the manipulator to grasp. Firstly, the preprocessed point cloud is segmented into super voxel blocks and clustered by locally convex connection. Then the multiple point cloud clusters are smoothed by MLS respectively to solve the fluctuation of sensor data. After that, in order to reduce the mismatches and improve the running speed, the edges are extracted by the angles among the direction vectors formed by the three-dimensional points of the point cloud cluster and their neighboring points in the local coordinate system. Next, rough matching and fine matching through E-ICP (Edge-ICP) algorithm are carried out between the clusters and the target model in the database. Finally, based on the nearest average distances in the matching, the sort of scores is carried out to complete the tasks of target recognition and pose estimation. Experimental results demonstrate the feasibility and effectiveness of the proposed method.