Research On Semantic SLAM Based On Lightweight Deep Network In Dynamic Scenes

Simultaneous localization and mapping(SLAM)algorithm is a key technology for mobile robots to act autonomously in unfamiliar environments.However,most of the traditional SLAM algorithms assume that the mobile robot runs in a static environment.When dynamic object interference occurs in the real environment,the accuracy and stability of the SLAM system will be greatly reduced.Therefore,the research of SLAM algorithm in dynamic environment has become a hot research direction.With the rapid development of deep learning,the semantic SLAM system combining deep network with traditional SLAM algorithm has excellent performance in identifying dynamic targets and filtering dynamic feature points in dynamic environment.However,the deep network used at this stage has the disadvantages of complex network structure,high hardware requirements,and long consumption time.It has a new contradiction with the mobile robot with limited computing resources and the SLAM system with high real-time requirements.In order to solve the above problems,the main work of this thesis is as follows:Firstly,in view of the large amount of computation and long reasoning time of the semantic segmentation network of deep learning,this thesis adopts the mainstream target detection network YOLOv5 for lightweight design.The main contents are as follows:(1)The lightweight feature extraction network Shuffle Net v2 is used to replace the original YOLOv5 feature extraction network.(2)Aiming at the detection requirements of SLAM system,a bidirectional weighted feature pyramid network with two detection heads is designed.(3)The loss function of the original bounding box is improved to improve the convergence speed and accuracy of the model.Finally,the accuracy and inference speed of the improved lightweight target detection network YOLOv5-SIENet are verified.The final experiment shows that the m AP(0.5)of the improved YOLOv5-SIENet on the COCO dataset reaches 54.5%,and the FPS of the inference speed reaches 209.Compared to the YOLOv5 s algorithm before improvement,the reasoning speed is enhanced by 2.7 times with 5.6% m AP(0.5)accuracy loss,which meets the real-time requirements of SLAM system.Secondly,in view of the problem that the Semantic information output from target detection may have misjudgment of dynamic feature points and semantic SLAM system to eliminate dynamic points may lead to problems such as system robustness decline.Based on the ORB-SLAM2 system framework,this thesis adds a dynamic feature point screening and adaptive filtering algorithm based on semantic and geometric constraints.The main contents are as follows:(1)The YOLOv5-SIENet target detection network and Depth＿RANSAC algorithm are combined as a new dynamic target detection thread added to the SLAM framework and the epipolar geometric constraint algorithm to construct a dynamic feature point screening module.(2)According to the extraction of ORB feature points and the number of dynamic feature points,a flexible dynamic feature point filtering algorithm is designed to ensure that the number of feature points of the SLAM system is not lower than the minimum requirement of the tracking thread,which improves the robustness of the semantic SLAM system.Finally,on the TUM public dataset,the SLAM system in this thesis is compared with the ORB-SLAM2 system and other excellent dynamic SLAM systems such as DS-SLAM and Dyna SLAM for accuracy and time-consuming analysis.The final experimental results show that in high dynamic scenes,the absolute trajectory error of the SLAM system in this thesis is reduced by more than 80% compared with the ORBSLAM2 system,slightly 5.7% lower than the Dyna SLAM system,and 14.8% lower than the DS-SLAM system.Compared with the ORB-SLAM2 system,the average reduction is 13.1% in low dynamic scenes.In the time efficiency analysis,the average time per frame of the proposed algorithm is reduced by 34.7% compared with DSSLAM and 84.1% compared with Dyna SLAM.The results show that the SLAM algorithm has better accuracy and real-time performance.