Research On Camera Pose Computation And Point Cloud Perception Algorithm In 3D Vision

In recent years,with the upsurge of artificial intelligence,the research on computer vision has become more and more diversified in academia and industry.The technology of Simultaneous Localization and Mapping(SLAM)has attracted wide attention,and this technology is an important step for robots to become intelligent.Mobile robots need to estimate their location in an unknown environment and be able to map out the surrounding environment so that specific tasks can be accomplished without human interaction.Early mobile robots mainly rely on laser radar,sonar,moving in space,and finally mapping spatial points to twodimensional maps.With the increase in the variety of cameras and the development of computer vision,the SLAM system relying on cameras is increasingly rich.Compared with laser radar,the camera is cheap,low-power,and rich in information,making the camera-equipped SLAM system more and more widely used in academia and industry.The vision-based SLAM system is called V-SLAM(Visual SLAM).This paper focuses on two issues of camera pose computation and point cloud perception.The main innovations are as follows: 1)In this thesis,a new camera pose calculation method based on plane markers is proposed for camera pose computation.Also,a new marker with two separated coplanar circles are designed.Based on the circular image,the parameters of camera pose can be directly calculated.In the method we proposed,multiple points do not need to be matched with each other,and only using a single picture can estimate the camera pose.In addition,owing to the usage of edge information of the designed coplanar circle,it can also guarantee the robustness of calculating the camera pose even under the condition of fast motion and image blur.The algorithm of this paper compares the simulation data and the real data.The experimental results show that the method we proposed achieves a high degree of precision and stability.2)According to the the existing three-dimensional data structure,we proposed a new network structure with point cloud as the input.The structure includes a channel attention module and the Vector of Locally Aggregated Descriptor(VLAD)network module.Based on the PointNet,test experiments are accomplished on the ModelNet40 and S3 DIS datasets,and the results prove the validity of the method we proposed.