Research And Implementation On 3D Scene Reconstruction Based On Multi-source Sensor Data Fusion

With the rapid development of technology,3D scene reconstruction,as a promising solution for a lot of practical applications,has been proved to be of great prospect in many research areas.3D reconstruction using the data from lidar is widely used as a technical method in recent years.However,target redundancy will occur in certain scenarios.To cope with this challenge,a fusion algorithm using RGB bitmap and data from lidar is proposed to eliminate the irrelevant targets in the specific scenes for 3D scene reconstruction.Firstly,a lightweight simultaneous localization and mapping(SLAM)algorithm is utilized to extract and match different types of feature points before merging point cloud at different time slots for 3D reconstruction.Next,for the irrelevant targets existing in the constructed point cloud map,deep learning application technologies in computer vision area are implemented to process multi-source sensor data for the object detection and elimination the in three-dimensional space.For the two different procedures of point cloud modeling and target detection,this paper adopts point cloud matching to fuse the algorithm and subsequently the 3D scene reproduction in the campus environment is conducted.The method proposed in this paper can be applied to smart cities,unmanned driving and other valuable practical application.In summary,the contributions of this paper are listed as follows:1)A hardware platform and software environment for 3D scene reproduction are set up.The hardware platform is mainly composed of a lidar,a camera and other assembly equipment.The software environment is consisted of the processing of point cloud and environment for algorithm.2)A low-drift,low-power 3D point cloud map modeling algorithm is implemented.The simulation results of existing point cloud modeling algorithms LOAM and Le GO-LOAM positioning and mapping algorithms are exhibited.The applicability of the proposed algorithm is proved by comparing with the effects of different algorithm.3)A D-FPN algorithm for object detection and elimination in 3D scene based onmulti-source data fusion is proposed.Multi-source data fusion can greatly reduce the range of target existence in point cloud data search,and then the point cloud object detection algorithm for object classification and 3D bounding box regression is used to guarantees the accuracy and efficiency of target detection of different sizes.4)Through an ICP point cloud registration method that combines point cloud modeling maps and single-frame target detection and removal results,the 3D scene reproduction without redundant information is finally completed.