Research On Visual SLAM Loop Closure Detection

In recent years,with the development of social technology,people’s functional requirements for intelligent robots are also gradually increasing.In this context,SLAM(Simultaneous Localization and Mapping)technology has emerged,which refers to intelligent robots collecting information about their surroundings through sensors and using the collected information to localize themselves and build maps.Since the camera has more smooth and rich information acquisition technology,the visual SLAM technology using camera as sensor is becoming mainstream.In the visual SLAM system,the front-end visual odometry will inevitably generate errors when estimating the pose of the adjacent images captured by the camera,and with the accumulation of errors,the system will not be able to build trajectories and maps that are consistent with the actual situation.To address this problem,the loop closure detection module is introduced in visual SLAM,whose function is to recognize the robot when it passes through the same scene and correlate the current scene information with the historical data,so as to build an optimization problem to eliminate the accumulated errors at the front end and build a trajectory and map consistent with the actual situation,and then improve the robustness of the system.In loop closure detection,bagof-words model algorithm is the current mainstream algorithm.In order to improve the performance of loop closure detection in visual SLAM,this thesis conducts research work on the bag-of-words model algorithm in the following aspects.1)In the loop closure detection based on bag-of-words model algorithm,the lexicon is formed by the K-means clustering algorithm directly on the descriptors clustering.However,the K-means clustering algorithm itself performs poorly on linearly indistinguishable data and is also sensitive to outliers in the descriptors,resulting in the construction of a dictionary with a low word finding rate.2)To address the dictionary construction problem,this thesis improves the bag-of-words model algorithm.In this thesis,firstly,the descriptors are clustered by the DBSCAN clustering algorithm,secondly,the outlier descriptors are re-clustered by the improved DBSCAN clustering algorithm,and finally,the clusters are partitioned by the dichotomous K-means algorithm using distance as a criterion,and the descriptors are selected as words from the partitioned clusters.In the improved bag-of-words model algorithm,the DBSCAN clustering algorithm can cluster data sets of various shapes,and outliers can be found and retained for them.The dichotomous Kmeans algorithm can be used without specifying the number of clustering layers and the number of split nodes,so the number of words composed by this algorithm can be changed according to the actual scenario.3)The dictionary is constructed by the improved bag-of-words model algorithm before and after the improvement,and the improved dictionary is experimented with similarity and number of word matches,and the improved dictionary can be obtained from the experiment,and the word finding efficiency is higher.For example,in comparing the image similarity value and the number of word matches,the improved bag-of-words model algorithm is higher than the one before the improvement.4)Through the similarity experiments,the loop closure threshold is set in this thesis,and the improved algorithm can be obtained from the experiments,whether on a single image or on the whole data set,its loop closure detection ability is much higher than that of the pre-improved algorithm.