Research On Closed-Loop Detection Algorithm Of Visual SLAM Based On Machine Learning

Nowadays,in the field of autonomous robots and unmanned driving,vision simultaneous localization and mapping(VSLAM)technology is one of the cutting-edge core technologies.In visual SLAM,the calculation error of the previous frame will continue to the calculation of the current frame,which will cause the accumulation of errors after a long run,resulting in a large error between the predicted value and the real value of the system.To solve this problem,the closed-loop detection technology can effectively eliminate the cumulative error by establishing the inter-frame relationship between the current frame and the historical frame.The commonly used closed-loop detection algorithms extract the local or global features of the image,and judge whether a closed loop occurs according to the similarity matching.Loop closure detection mainly includes methods based on the bag of words model and methods based on deep learning.Thesis studies these two methods separately,and the research contents are as follows:(1)Through the research of closed-loop detection based on ORB features and offline bag of words,it is found that when using offline bag of words,there will be a phenomenon of mismatching of different key frames during system operation,which will affect the effect of closed-loop detection.First of all,the data of the offline bag-of-words model is fixed,but it is constantly changing in actual application scenarios.The fixed bag-of-words model cannot well reflect the image features in the scene,and will cause a large number of mismatches.Therefore,for the DBo W2 bag-of-words model,Thesis designs a closedloop detection algorithm for updating the bag-of-words model online,so that the bag-ofwords model can contain the characteristics of the application scene,and the closed-loop detection is more accurate.The improved algorithm is applied to the ORB-SLAM3 system to test on the public data set.The experimental results show that the error between the system’s predicted trajectory and the real trajectory is reduced.In order to verify the actual effect of the improved algorithm,the algorithm was transplanted to the sweeping robot for testing,and the results showed that the improved error was 17.58% lower than the original model,and the stability is better.(2)Through the research on the Net VLAD network,it is found that the network cannot fully extract information such as image position,channel and space when extracting image features,which will lead to a decrease in image matching rate during closed-loop detection.Therefore,thesis proposes a feature extraction network model combining Par C＿Net network,Conv Ne Xt-Tiny network and CBAM attention mechanism.This network is combined with Net VLAD network for image matching in loop closure detection,and Par C＿Net network is used to extract image feature position information.Then,the channel and spatial information of the feature map of the CBAM attention module are added.Finally,the features are aggregated by VLAD,and the local features are fused into global features.By comparing the improved network with multiple image matching networks,the results show that the network proposed in thesis has the highest accuracy when the recall number N is the same.To sum up,thesis studies and proposes improvements to the key issues in the two closed-loop detection methods based on the bag-of-words model and deep learning.The improved method basically achieves better matching accuracy in datasets and practical applications,the optimization effect is better.