Research On Model Establishment And Recognition Method Of Special Shape Lock Pin Based On 3D Point Cloud

The installation and removal of container lock pins is one of the important tasks of the terminal,and the intelligent and automated loading-unloading system has become the mainstream configuration of the terminal.In the smart unlocking station of the container trucks at the terminal,the conveyor belt transports the lock pins,so recognizing the lock pin type correctly to ensure the correct installation and removal by the manipulator is very important.For the lock pin recognition on the conveyor belt in the container lock pin intelligent unlocking system in this paper,the lock pin scene is fixed and neat and does not contain occlusion.In order to meet the recognition of the low-precision point cloud of the dataset,the limitation of the amount of lock-out data,and the response plan and demand for temporary emergencies for new lock pins at the terminal,it is better to choose the recognition method of global feature descriptor matching instead of deep learning.It maybe take less time and has advantages to deal with the temporary addition of new lock pins.In this paper,the method of global feature descriptor matching of 3D point cloud based on self-built library is used to recognize the lock pin.In the self-built library,there are data set with part of the local perspective point cloud of lock pins and data set with the non-high-precision complete point cloud of lock pins obtained by the depth camera and the acquisition system.When establishing the model,in order to obtain the lock pin model for matching quickly,based on the SAC-IA（Sample Consensus Initial Algorithm）and ICP（Iterative Closet Point）registration methods,this paper proposes a method of quick model establishment based on high-overlap perspectives screening because of the need of high-overlap point clouds of this registration method.When splicing point clouds,the relative size of the bounding box area projected by the lock pin in the z-direction of the camera coordinate system is obtained to estimate the general shape of the lock pin,and an appropriate number of point cloud views with a high overlap are selected to ensure the success rate of registration and reduce timeconsuming based on the difference between the areas of adjacent views.In the recognition process,the category of lock pin is searched through the training model established offline by the VFH（Viewpoint Feature Histogram）or CVFH（Clustered Viewpoint Feature Histogram）of the target point cloud.In order to solve the problem that the point cloud on the plane perpendicular to the viewing angle is filtered together with noises due to sparseness,the method of descriptor selection based on the Euclidean distance clustering number is used.The method has low time complexity and little impact on the recognition efficiency,so that both descriptors can play their respective advantages and ensure the recognition accuracy and increase recognition stability.Finally,for the lock pin category,this paper uses the lock pin’s bounding box to optimize the recognition process.Based on the correlation between the size of the lock pin and the actual use,the lock pin is divided into three clusters according to the size by the clustering method,which reduces the possibility of recognition errors,and at the same time slightly speeds up the time-consuming of a single recognition.That makes the recognition accuracy rate and efficiency improved.