Research On Depth Completion Based On Propagation Mechanism And Laser Intensity Information

In recent years,deep learning technology has flourished,due to which autonomous driving technology and other mobile-robots technology have also made great progress.In these technical fields,dense and precise depth information,namely distance information,plays an important role for various tasks.However,the depth maps collected by Li DARs are sparse.Therefore,depth completion task has gained its popularity.In this paper,two aspects of research are conducted in view of the limitations of the existing depth completion algorithms.Transformers have achieved great success in natural language processing and computer vision tasks.The multi-head attention layer in Transformer can build a global attention matrix and weight the input features,which is suitable to perform spatial propagation of depth information in depth completion tasks.Therefore,in this paper,study on applying Transformer in depth completion task is conducted and the GSPT model is proposed.Firstly,in this part of the research work,the concept of Depth Memory information is firstly proposed and then combined with sparse depth map related information to construct Depth Embedding,and Transformer decoder layer is utilized to perform the global spatial propagation of Depth Embedding;Secondly,in view of the problem that the existing Transformer structure is not fully suitable for depth completion tasks,this paper proposes to improve it and builds a Trans UNet sub-network,with which the propagation of depth value information is completed and global information is extracted.Finally,aiming at the balance of extracting global semantic feature and protecting local depth information,the encoder design of the GSPT model is completed.The proposed GSPT model achieves advanced depth completion accuracy on the KITTI depth completion validation set and selected validation set.Further ablation study results demonstrate the effectiveness of the improvements in this research work.Algorithms associated with point cloud processing such as 3D object detection networks,usually take the intensity data of each point as part of the input information,while none of the existing depth completion algorithms make use of laser intensity information.In view of this limitation,this paper proposes that the use of laser reflectivity data can assist the depth completion algorithm,and conducts research on how to demonstrate the effectiveness of laser reflectivity and how to design a reasonable way to utilize it.In this part of the research work,basing on the pin-hole camera model,laser point cloud is firstly projected to the RGB image coordinate system,generating the corresponding laser intensity map of each sample in KITTI dataset.Then,this paper analyzes the laser intensity data and considers the characteristics of the existing depth completion model,after which the Intensity Feature Extractor and Intensity Feature Utilizer are proposed and then introduced into three depth completion algorithms including GSPT model.Experimental results on the KITTI depth completion dataset show that the proposed IFE and IFU can improve the accuracy of existing depth completion models.The auxiliary effect of laser intensity data on depth completion task is then presented by ablation study results.