| A 3D laser scanning point cloud scene provides an objective description and digital reproduction of the scanning scene in the real world.By analyzing the positional relationship between objects in a selected scene(e.g.,a road scene)and improving the accuracy of occluded objects,we can clarify the spatial factors leading to occlusion,thus presenting great academic and applied value.However,some technical problems and gaps still exist in the visibility analysis of 3D point cloud scenes.First,some semantic objects suffering occlusion blend into the background,making extraction of these separate semantic objects difficult.Second,there is lack of a systematic automatic visibility analysis method to build observational models and deal with occlusion calculations in complex and variable point cloud scenes.Third,in a practical sense,mass point clouds require more efficient occlusion analysis and object detection methods.As a demonstration,this research work analyzes the occlusion of traffic signs and ground caused by vegetation canopy.Thus,this thesis describes the development of an accurate and effective visibility analysis method for the selected scene in point clouds,from the following three aspects:Firstly,this thesis proposes,for dense laser point clouds,a rapid vegetation crown extraction algorithm based on integral approximation.This algorithm starts with the voxelization operation and then constructs a spherical voxel neighborhood space for each non-empty voxel.The similar values between those two voxels are calculated.The similar value of the entire spherical space is thus accumulated.Thereafter,voxels are identified for obtaining the crown set of the point clouds according to given structural rules.The F1-asure value of this proposed method remains above 0.8.The run time is kept below 0.2 minutes in the mature urban scene.The results demonstrate that our proposed method is feasible for conducting rapid visibility analyses.Secondly,this thesis presents an illumination occlusion analysis method that considers vegetation objects based on laser scanning point clouds.At first,the method extracts a region of interest by using salient features and non-maximum inhibition.For the region of interest,a furthest point greedy strategy is then employed to obtain the number and position of the base points.Thus,by using the geographic information model and the base points,the solar position of the point cloud scene is calculated.Finally,the Generalized Hidden Point Removal algorithm(GHPR)is adopted to perform the occlusion analysis.Based on our experiments,the average accuracy of this method is about 2.71 meters,demonstrating its feasibility.Thirdly,this thesis proposes a driver perspective based traffic sign occlusion detection method.This method uses multi-conditional filtering to obtain traffic signs,and then a trajectory-based observation planning method is proposed to select the traffic signs to be observed and also to locate the corresponding viewpoints.Next,the hidden point removal algorithm is used for occlusion calculation,followed by occlusion analysis and occluder extraction.The F1-measure value of this method is about 0.85,which is effective for analyzing vegetation occluded traffic facilities.In summary,the proposed rapid crown extraction method provides the semantic information in a scene for visibility analysis.The illumination occlusion analysis method overcomes the irregularity of vegetation contours and provides technical support for illumination occlusion analysis of a ground point cloud scene.The traffic sign occlusion detection method provides technical support for visibility analysis of the road scene in point clouds.Therefore,this thesis constructs a systematic visibility analysis method for a point cloud scene.It is believed that this research improves the visibility analysis technology for point cloud scenes,and promotes the application of the point cloud processing technology in various fields,e.g.,landscaping,transportation planning,renewable energy and architecture. |
PDF Full Text Request