Research On 3D Scene Reconstruction Method For Mobile Robot In Unstructured Environment

With the improvement of people's pursuit of quality of life,the continuous increase of industrial labor costs,the urgent need for industrial upgrading,artificial intelligence technology has made great progress,the application of mobile robots has become more and more extensive,and gradually replacing or assisting humans to complete many difficulties.Being able to move autonomously in all kinds of complex environments is a prerequisite for mobile robots to complete their tasks.3D reconstruction is an important part of the field of computer vision.It not only has an important position in scientific research,but also is closely related to our daily life.3D reconstruction can provide rich information to help robots perceive the environment,identify scenes and make reasonable judgments to ensure that they can move safely and stably.In this context,it is of great significance to study how to use mobile robots to build 3D models of scenes.In view of this,this paper carries out research work on 3D reconstruction of mobile robots in unstructured scenes.The main research contents of this paper are as follows:(1)The Kinect depth image sensor is used to collect the data of the measured scene,and the acquired multi-frame point cloud data is registered and spliced.A point cloud registration method combining sampling consistency and iterative closest point algorithm is proposed.Extract the Fast Point Features Histograms(FPFH)feature of the point cloud data and calculate the initial transformation of the point cloud by using the Sample Consensus Initial Alignment(SAC-IA).The better registration position solves the problem that the registration process is easy to fall into the local optimal solution.The k-d tree neighbor search method is added to the iterative nearest point algorithm to accelerate the search for the corresponding point pairs,which accelerates the convergence speed.The direction vector threshold is used to effectively eliminate the misalignment and improve the registration accuracy.The experimental results show that the efficiency of this algorithm is 55.28% higher than that of the traditional ICP algorithm,and the registration error is reduced by 69.86% on average,which improves the accuracy and robustness of the overall registration algorithm.(2)The ray projection algorithm is used to extract the depth data information,and the TSDF voxel representation is used to store the data depth information and weights.The surface reconstruction is implemented by the ball rotation surface reconstruction algorithm.Aiming at the problem that the sampling point information data is missing and the noise point data is easy to appear in the ball rotation algorithm,the ball rotation algorithm is improved,and the surface method of adding point cloud data is added in addition to the position information of each spatial point data of the point cloud.The line information removes the error space point by setting the angle between the normal vector directions of the spatial points,and improves the surface model reconstruction accuracy.The experimental results show that the improved spherical rotation algorithm can obtain better surface reconstruction results of geometric models,and real-time scene surface reconstruction can be realized.(3)Texture reconstruction of the surface model of the reconstructed scene is performed using an improved voxel color mixing algorithm.Using the key frame blur detection algorithm to select a color image with higher definition as the input texture atlas can effectively improve the texture blur problem.The dynamic weight distribution function is used to improve the color information update and the distribution of color information weights in voxels.The theoretical analysis and experimental results show that the improved dynamic weight distribution function can reasonably allocate the original weight and update weight,and improve the edge area.Shadow blurring phenomenon.The texture reconstruction experiment achieved a frame rate of 16 fps,enabling clearer scene texture reconstruction in real time.