Extraction And Completion Of Vehicles In Underground Parking Lots Based On Graph Convolutional Network

For the automatic parking in an underground parking lot,accurate detection of vehicles near the parking space will greatly help to achieve the successful parking.Therefore,it is very necessary to study the algorithms for extracting vehicles in underground parking lot based on 3D point clouds.The point clouds of vehicles in the parking lot are easy to be incomplete because of the occlusion.The point cloud completion of incomplete vehicle is,thus,important for the visualization and 3D modeling of underground parking lots.This thesis then aims at developing deep learning point cloud techniques for the vehicle extraction and completion.For the extraction of underground parking vehicles,we add a dimensional feature module to the graph convolutional network to employ the local shape of the point clouds to help the semantic segmentation.In order to further improve the segmentation accuracy,we propose a method to use the minimum bounding box of the vehicle to improve the semantic segmentation results.Experiments show that our method can effectively extract the vehicles in parking lots.The vehicle segmentation accuracy is 99.6%,and the mean IOU is 98.5%.In the work of the point clouds completion of incomplete vehicles,an attitude adjustment module is added to the completion network to enhance its adaptability to the multi-pose incomplete vehicles.This thesis introduces an IOU loss function to accelerate the convergence of the network.Experiments show that our methods can effectively implement the completion of vehicles in the parking lots.Compared with the baseline algorithm,the chamfer distance is reduced to 41%.