Research On Loop Closure Detection Method In Visual SLAM

SLAM(Simultaneous Localization and Mapping)means simultaneous positioning and map construction.Using SLAM technology,mobile robots can achieve autonomous localization and mapping in previously inaccessible environments.When the sensor is a camera,it's called a visual SLAM.With the long-term movement of the robot,the motion error will accumulate,which makes it impossible to build globally consistent trajectory and map.Therefore,the loop closure detection module is required to detect whether the robot has visited the current scene,reduce the accumulated error,build consistent map,and improve the robustness of the visual SLAM system.Most of the traditional loop closure detection methods are based on the bag of word,and there are also methods to solve loop closure detection problems based on PSO(Particle Swarm Optimization)algorithm.In order to improve the accuracy and efficiency of closed-loop detection,this thesis proposes a closed-loop detection method based on HHO(Harris Hawks Optimization)algorithm,which has achieved good results in a small scale data set.The characteristics of the traditional methods are used to artificial design,artificial design features cannot take full advantage of image information and poor robustness,when large data sets,feature extraction effect is poor,and the convolutional neural network by training to extract the image characteristics of deep and extract features better robustness,can make full use of image information,the feature extraction in massive data set.This thesis presents a loop closure detection method based on improved Res Net(Residual Network)network.The main research and work are as follows:(1)A loop closure detection method based on HHO algorithm is proposed.Firstly,to solve the problem that the original FAST(Features from Accelerated Segment Test)feature extraction algorithm is unstable when the light condition changes,the FAST algorithm adopts the dynamic threshold method to extract image features.Then the image feature descriptor is calculated and compressed to obtain the image robust descriptor.Finally,the loop closure detection problem is transformed into the optimization problem of finding the image with the maximum similarity with the current frame in the historical frame,andHHO algorithm is used to solve the problem.The experimental results show that compared with the loop closure detection method based on bag of word and PSO algorithm,the proposed closed loop detection method based on HHO algorithm in this thesis has higher efficiency and accuracy.(2)A loop closure detection method based on improved Res Net network is proposed.In order to extract multi-scale features of images and make full use of image information,this thesis integrates Res2 Net structure into Res Net network to improve the feature extraction capability of Res Net network.The experimental results show that,compared with the closed loop detection method of the bag of word,the original Res Net network and HHO algorithm,the improved Res Net network detection method proposed in this thesis has higher accuracy and efficiency.