Representation Learning And 3D Object Recognition From Mobile LiDAR Point Clouds

Object recognition is the foundational task of 3D point clouds processing.It is of great significance for many applications to recognize road objects accurately,quickly,robustly and efficiently.As the core scientific problem of object recognition,representation learning aims to mine the semantic features of point clouds and reconstruct the compact expression.At present,the characteristics of mobile LiDAR point clouds,such as the large amount of data,uneven density,serious occlusion and noise disturbance,bring great challenges to point clouds intelligent processing.There are still several problems,including the fast recognition of ground objects,the accurate characterization of spatial structure features and efficient recognition of point clouds in complex environments.Therefore,focusing on the 3D point clouds representation learning and recognition in road scene,this paper aims to effectively solve the above problems via theoretical analysis and experimental verification.This doctoral dissertation mainly consists of three aspects:Firstly,because of the large amount and high density of point clouds in road scene,it is difficult to recognize objects quickly.To solve this problem,a 3D point clouds recognition model based on fusing multi-view representation is proposed.Three shallow feature descriptors are designed to obtain the 2D representations of 3D point clouds.Then,building on the 2D representations,a two-step features fusion strategy is proposed,and a point cloud recognition model based on multi-view representation learning is proposed.The experimental results on several datasets show that the proposed model can recognize the road objects quickly,accurately and robustly.Secondly,to solve the problem of descripting accurately the spatial structure features of 3D point clouds in road scenes,a sequential slice representation learning based method is proposed.This method uses recurrent neural network to mine the correlation between sequence slices,and an attention mechanism strategy is designed to fuse the temporal and pointwise features efficiently.In addition,the main direction slicing strategy introduced in the slicing process significantly improves the robustness to the rotation change of 3D objects.And the end-to-end weight automatic learning mechanism designed in the fusion of temporal features effectively uses the structural information contained in different temporal output.By transforming the spatial feature learning into the temporal feature learning,the model can descript point cloud accurately.Finally,because the representations in most existing 3D object recognition models cannot match the discrete property of point clouds closely enough,these models usually have a large amount of parameters and high memory consumption.To solve this problem,a point clouds representation learning model based on local dilated connection and context-aware information is proposed.Specifically,build on the spatial structure features contained in the relationship between points,a local relation extraction strategy,i.e.,local dilated connection,is designed to expand the receptive field.At the same time,a graph node feature expression that integrating spatial location and context information is proposed,and a graph convolution neural network model for point clouds recognition is developed.This model not only maintains its accuracy and robustness,but also improves the memory and training efficiency.The above works focus on 3D point cloud representation learning as well as 3D object recognition.They were evaluated via extensive experiments.The results show the effectiveness of the proposed methods in solving their own problems,and the comparison results with other existing methods also show the superiority of the proposed methods.