Lock Pin Recognition Algorithm Based On Self-Built Point Cloud Library

Automatic disassembly and installation of container lock pin is one of the key links to realize automation of large terminals.In the installation process,different types of lock pins are transported to the designated identification area by conveyor belt.According to the identification results,the mechanical arm selects the matching fixture of lock pin for automatic lock pin installation.At present,there are various types of lock pins,but there is a shortage of point cloud models that can be used for identification,and even unknown lock pin may exist in dock operations.Therefore,it is particularly important to quickly establish a complete lock pin point cloud model for identification.This paper firstly selects the point cloud modeling algorithm based on turntable to reconstruct the lock pin point cloud model.Then extracts multi-view point clouds from the reconstructed point cloud model to establish the multi-view point cloud database of lock pin.Finally,based on the self-built point cloud library,multiple descriptors are applied to carry out lock pin identification and verify the availability of the self-built point cloud library.The main research contents are as follows:(ⅰ)The 3D point cloud model reconstruction of lock pin is studied.According to the complex structure and irregular shape of the lock pin,this paper chooses the reconstruction algorithm based on parameter calibration of turntable.The sensor located at a fixed height is combined with a rotatable turntable to reconstruct the point cloud model,and its surface features will be missing to some extent.To solve this problem,an improved spherical projection algorithm is proposed in this paper.Firstly,the algorithm locates the missing area of the point cloud model.Secondly,according to the location of the missing area,the positioning posture of the lock pin on the turntable is changed,and the point cloud model under this posture is established.Finally,the point cloud model reconstructed under multiple poses is integrated to make up for the missing surface features.(ⅱ)Based on the reconstructed lock pin point cloud model,a multi-view point cloud database is established.In the dock identification scene,the Angle of view of the lock pin in the identification area is variable.In order to improve the recognition rate of the lock pin,the self-built point cloud database should contain as many data of the lock pin point cloud from different angles as possible.This paper selects the multi-view point cloud acquisition algorithm of the Point Cloud Library(PCL).Affected by the resolution and working principle of different cameras,the established point cloud database cannot identify the lock pin point cloud data collected by other cameras well,and the expansion of the database is also limited.In view of the large difference in density and size between cross-source point clouds(point clouds collected by different cameras),the improved cross-source point cloud preprocessing algorithm was proposed to maximize the elimination of cross-source point cloud differences.(ⅲ)Based on 3D point cloud descriptor,the identification of lock pin point cloud is studied.Considering that descriptors of deep learning have high requirements on data sets and need to retrain the model when encountering unknown lock pin,which is timeconsuming,it cannot satisfy the recognition system’s rapid adaptation to the unknown lock pin.In this paper,a variety of traditional descriptors are selected to carry out object recognition experiments.According to the experimental results of the Self-built Point Cloud Library and the Stanford point cloud model library,the advantages and disadvantages of different descriptors are analyzed and the usability of the self-built point cloud library is verified.Aiming at the complex structural of the lock pin,this paper firstly determines whether the point cloud to be identified is blocked by the Euclidean Clustering Segmentation.Based on the judgment results,a hybrid Feature descriptor based on the Global Aligned Spatial Distribution(GASD)and the Clustered Viewpoint Feature Histogram(CVFH)is proposed.The experimental results show that the algorithm can maintain a high recognition accuracy when the point cloud is blocked or not blocked.