Research On Visual SLAM In Complex Environment Based On Deep Learning For Automotive Guided Vehicle

Autonomous mobile robots need accurate positioning information and surrounding environment information to realize planning and control functions,so as to realize more expansion functions.Simultaneous localization and mapping(SLAM)algorithms are mostly designed in static environments,but they often encounter complex dynamic environments in practical applications and cannot effectively deal with dynamic objects and static objects moved by humans.As a result,the performance of positioning and mapping is affected.The deep learning method can obtain the semantic information in the image,and assist SLAM algorithm to realize more accurate and reliable localization and map construction in complex environment.Therefore,aiming at indoor AGVs,this paper studies a synchronous positioning and mapping algorithm based on deep learning in complex environments.The research contents and achievements are as follows:(1)Considering dynamic objects and human moving objects in complex environment,a complete visual SLAM algorithm framework is constructed based on Mask R-CNN neural network and ORB-SLAM3 algorithm,and all aspects of SLAM algorithm are designed,including the following aspects: extracting semantic feature points by using neural network instance segmentation results;Identify and eliminate dynamic objects in the front end of SLAM;Improve the key frame strategy to achieve more reasonable local map construction and optimization;Improved long-term data association of SLAM algorithm in loop and map fusion thread.(2)For dynamic targets,a motion consistency checking algorithm based on semantic information is proposed,and the potential dynamic feature points are detected by using the pole constraint method of adaptive threshold,so as to achieve accurate removal of dynamic feature points in images.The key frame selection strategy is optimized according to the removal of dynamic feature points.The results of data set test and experiment show that this method can eliminate dynamic feature points well,and can stabilize the position and posture of the tracking system.(3)For the artificially moved object(semi-static),the change of its map point coordinates will affect the positioning accuracy during map fusion or loop optimization.An improved long-term data association method is proposed to optimize the loop optimization and map fusion process of the algorithm by optimizing the semi-static information in the key frame.The experimental results of historical map loading show that the algorithm can reduce the influence of semi-static objects on the positioning effect and build high reuse maps when map fusion occurs in the system.(4)The public data set and the AGV experiment vehicle were used to compare and verify the performance of the algorithm.Targets with high dynamic degree exist in TUM/fr3/walking＿xyz data set.Test results show that the proposed algorithm can run stably in high dynamic environment,and the positioning accuracy is 65.6% higher than that of ORB-SLAM3.In the data set TUM/fr3/sitting＿xyz,only objects with low dynamic degree are found,and the test results show that the proposed algorithm still improves the positioning accuracy by 8.5%.Experimental results of real vehicles in dynamic scenarios also show that the proposed algorithm has high positioning accuracy and robustness.The experimental results of historical map loading show that the algorithm still has good map reuse in the long run when there are semi-static objects,and can eliminate the interference of the identified semi-static objects on the pose optimization,which proves that the algorithm in this paper has an obvious effect on the improvement of long-term data association optimization method.