| With the development of society,for unmanned autonomous mobile system,if only relying on simultaneous positioning and mapping technology can not meet the needs of complex environment.How to improve the perception ability of unmanned system to the external environment and the robustness of location mapping is an urgent problem to be solved in complex tasks,and it is also a difficult problem to be overcome by semantic slam.In order to improve the robustness and accuracy of positioning and mapping,and accelerate the understanding of the semantic information of the surrounding environment,this paper mainly studies the visual mileage calculation method of dynamic feature point elimination and the thread design of semantic map construction based on STDC network.In the complex dynamic environment,due to the existence of dynamic feature points,the accuracy of system positioning and mapping is greatly reduced.This paper proposes a front-end visual odometer design algorithm based on dynamic feature point elimination.In order to reduce the impact of noise generated by dynamic feature points on system positioning and mapping and ensure the operation efficiency of the system,yolov5 s,which is fast in the field of computer vision,has a small network and low AP accuracy,is used as the target detection network to track feature points through LK optical flow pyramid,The distance between the epipolar line and each feature point is used as the judgment criterion of dynamic feature points,and the dynamic feature points generated by dynamic objects are eliminated in combination with the boundary box obtained by target detection.In order to accelerate the running speed of the front-end visual odometer of the whole visual slam system on the premise of ensuring the positioning and mapping accuracy,the quadtree homogenization operation is carried out for all the extracted static points in the form of grid,and some useless static feature points are eliminated by retaining only the feature points with the highest response value in the quadtree grid,Thus,the time of feature matching and descriptor calculation is saved,and the operation efficiency of the whole system is improved.Compared with orb-slam2 system without removing dynamic feature points,their absolute trajectory error is reduced by 5.43% on average,and the speed of image processing for each frame is increased by about 12% compared with that without homogenizing feature points,It is effectively proved that the proposed algorithm improves the robustness of unmanned system positioning and mapping.The traditional visual slam system can not meet the needs of performing specific tasks in some complex environments.Based on the traditional visual slam system,this paper combines visual slam with deep learning by using convolution neural network technology.Through the lightweight semantic segmentation model STDC network,this paper opens up a semantic mapping thread to semantically segment the key frames passed from the tracking thread to obtain the segmented probability map.Through the mapping relationship between2 D pixels and 3D map points,The obtained probability map is mapped to three-dimensional map points,the semantic information of spatial map points is determined,the three-dimensional spatial map points corresponding to all key frames are jointly estimated,and the map points corresponding to all key frames with probability information are fused to obtain the final three-dimensional dense semantic map.The semantic map construction thread designed in this paper is tested through four kinds of image sequences in the tum dataset.The dense and semantic map construction experiments are carried out for each dataset.The results show that the algorithm not only realizes the semantic information rendering for spatial three-dimensional points,but also accelerates the efficiency of semantic map construction based on STDC lightweight network model,The construction of lightweight semantic map is completed. |
PDF Full Text Request