Research On Point Cloud Processing Algorithms For Port Automatic Unloading

With the rapid development of economy,trade between countries and regions is becoming more frequent.The bulk materials are important resources for import and export.Its unloading process is that driver controls the grab in driver's cab and drops it into cabin of the bulk carrier to grab the material.And then the grab is moved to the bin to release the material.Manual operation is limited by factors such as time and weather,resulting in low efficiency of unloading opeations and high accident rate.Therefore,automatic unloading technology has gradually received attention.Automatic unloading needs to realize the autonomous positioning of the cabin,the grabbing point decision and the safety warning function.Considering that the laser scanner has high degree,wide scanning area and little influence by environmental factors,this paper uses laser scanner to scan the unloading carrier.And based on point cloud data,this paper implements an automatic workflow.The difficulty lies in the complicated port scene,large amount of data and high real-time requirements.The position of the scanner will change and the coordinate system will be displaced.At the same time,the scene will be contaminated by the grab track and interfere with the segmentation and positioning of the cabin.In this paper,aiming at above difficulties,five point cloud segmentation matching problems are proposed for the port scene,and the results are applied to the automatic unloading system.The specific contents are as follows:1.Scene pre-segmentation based on elevation clustering algorithm: the algorithm performs a large range of pre-segmentation on the point cloud scene based on the elevation information and the Euclidean distance.For the port scenario,it can be initially divided into land area,deck area and material area.2.An algorithm for segmentation of irregular connected objects based on the growth of super-voxel clustering regions: the algorithm over-cuts the scene to obtain super-voxels,and then performs regional growth and merging based on super-voxels to segment connected objects.For the port scene,the grab will appear in the scanning scene during the scanning process,which seriously affects the subsequent segmentation recognition algorithm.The grab track can be regarded as an irregular large-area connected object,and the algorithm can achieve efficient segmentation and filtering of the grab track.3.Improved RANSAC based on Mean Shift for target plane segmentation: the algorithm firstly obtain the plane actual normal by Mean Shift,and based on the improved RANSAC to accurately segment the plane.For the port scenario,the positioning of the grabbing area,that is,the cabin area,needs to be completed.The algorithm can achieve accurate segmentation of the cabin plane and greatly improve the efficiency of the segmentation.4.3D Hough transform based on SHOT features for model matching: the algorithm uses the SHOT descriptor to perform rough matching of the model,and then performs3 D Hough transform matching based on the candidate matching points.For the port scenario,the local coordinate system with the scanner as the origin may be displaced,so the objects in the scene need to be mapped to the world coordinates according to the local coordinates.In this paper,the method of positioning the reflector is proposed to calculate the mapping relationship between the two coordinate systems.The algorithm is used to achieve the matching and positioning of the scene reflector.5.PointNet based on local voxel for pedestrian segmentation and recognition: the algorithm extracts global features from PointNet network,and adds feature extraction layer to extract features from local voxels,and combines global and local features to segment pedestrians in the scene.The model is based on open road scenario dataset being trained and tested,which is 2.1% higher than PointNet in IoU.For the port scenario,the algorithm can effectively segment pedestrians in the deck area and improve system security.Finally,this paper develops a complete automated unloading operation system based on Qt cross-platform open source framwork and PCL library.The system encapsulates the above algorithms and provides good human-computer interaction and visual interface functions.The effects of algorithms are verified based on the real port scene data.