Research On Real-Time Occupancy Grid Mapping For LSD-SLAM

LSD-SLAM(Large-Scale Direct Monocular Simultaneous Localization and Mapping)is based on the monocular vision sensor data to estimate the robot pose and reconstruct the environmental scene in an unknown environment.Research on GPS-independent navigation methods is currently a research hotspot in the field of autonomous unmanned systems.In this paper,LSD-SLAM is studied and the reconstruction of 3D scenes is studied.Based on this starting point,the following research work was mainly completed.Firstly,for the monocular vision SLAM system,LSD-SLAM is determined as the research object.The environmental dataset was selected by contrast method.LSD-SLAM was used as a direct method independent of features to process global image pixels and create semi-dense three-dimensional maps in real time.Secondly,for the problem that the semi-dense point cloud image of LSD-SLAM output is not applicable to the autonomous navigation system,the reverse sensor model(ISM)can be used to characterize the uncertainty of the specific sensor,real-time creation of occupancy grid mapping using a combination of inverse sensor model and LSD-SLAM.Then,by comparing several typical techniques,the octree occupancy grid mapping with full 3D modeling,updateability,flexibility and compactness is determined as the 3D environment reconstruction method in this paper.After discussing the key parameters of the existing Gaussian inverse sensor model that can be used for monocular vision,a real-time three-dimensional occupancy grid mapping is generated.Andert's measurement inverse sensor model for stereo vision is improved.After analyzing the influence of key parameters of the improved measurement inverse sensor model,the real-time octree occupancy grid mapping of the improved measurement inverse sensor model is realized.Finally,the three-dimensional modeling of Gaussian inverse model is compared with the improved three-dimensional modeling of measurement inverse model from the perspectives of memory consumption,real-time,qualitative analysis and quantitative analysis.The three-dimensional occupancy grid mapping under LSD-SLAM is analyzed.The effect of the map and the factors affecting the three-dimensional modeling of the monocular vision system werediscussed.Experiments have shown that LSD-SLAM can be combined with the inverse sensor model to create an accurate semi-dense grid map for autonomous navigation.For large data sets,the Gaussian inverse sensor model performs well.For small data sets,the improved measurement inverse sensor model can completely replace the Gaussian inverse sensor model and map quality at low resolution(128mm).Better than the Gaussian inverse sensor model.