Research Of Algorithm And System Implementation On Depth Completion

In many computer vision applications,such as 3D object detection,3D instance segmentation and slam tasks,accurate depth value is very important to improve accuracy.Especially for autopilot vehicle scenarios,many applications rely on the perception of distance between objects and use depth information reasoning and corresponding responses.The task of depth completion is to obtain dense and accurate depth map.The method based on multi-sensor fusion is the mainstream solution at present.In the field of autopilot,Lidar is used to fuse with RGB image.Lidar’s point cloud is accurate but sparse.RGB image is dense but does not contain depth information directly.The fusion of them can play a complementary role.The traditional depth completion method ignores the completion effect of the edge part of the object and the importance of the closer object to the practical significance.In addition,there is a lack of cheap and easy-to-use autopilot engineering verification machines equipped with real-time depth completion services in the market.In view of the above problems and needs,this paper introduces the work in the following two aspects:First,improved depth completion algorithm.This paper improves the previous depth completion algorithm from two aspects:depth interval division,depth value discretization and reconstruction;New evaluation indexes are proposed for the depth edge of the object and the depth of the near object.The experimental results show that the algorithm model improves the edge blur of dense depth map,and improves the performance of dense depth map on closer objects.Second,developed the prototype system of mobile robot car with deep complement.In industrial production,a simple and cheap way is needed to verify the feasibility of engineering design.This paper selects low-cost hardware equipment to build a mobile robot car platform and deploy navigation services.Using the master-slave distributed ROS cluster,the real-time depth completion service is developed,and two upper layer applications based on dense depth map are deployed.The prototype has the advantages of simple use,powerful function,strong expansibility and low cost.It has the advantages of loose coupling and pluggable for the verification of application layer tasks,and has high practical application value.