Research On The Method Of Loop Closure Detection Of Visual SLAM Based On BoW

Visual SLAM(Simultaneous Localization And Mapping)is the technology that based on camera as a sensor,researches simultaneous localization and mapping of autonomous mobile robot.Nowadays,the development of artificial intelligence is changing with each passing day.Visual SLAM plays an important role in the advancement of popular applications such as self-driving and AR and so on.Loop closure detection is the method of checking whether the moving trajectory of the mobile robot coincides with the previous place in the Visual SLAM in real time.It will correct the cumulative errors of position and orientation of the mobile robot and the building map in the process of robot'smovement,so it is an essential task for Visual SLAM.Loop closure detection should be characterized by accuracy,high efficiency,and real time.At the same time,it is faced with the reasonable choice of key frames,the perception confusion caused by repeated scenes,the selection of loop closure detection's method,and the inspection for robustness of loop closure detection's result.For the requirements and main problems of loop closure detection,this article researches the BoW model which is used to match image and improves it for loop closure detection,then puts forward an effective method of loop closure detection.The article introduces and practices the process of loop closure detection through three aspects.First,describe the specific process of creating the BoW model.The principle of three types of algorithms about feature points extraction is introduced and test is performed,then select the ORB algorithm that best meets the requirements of loop closure detection.Using the method of k-means to cluster feature points into visual words,then create the visual dictionary tree and establish the feature index to enhance BoW's searching efficiency on features.Secondly,analyze and summarize the defects of the BoW model,and the improved method for adding feature space information is proposed.That is,dividing the image into 4*4 grid areas in the same size,and counting the score of visual words that are quantified in each area by the features inside,to increase the spatial relationship of the features.Then show the implementation details of feature space vector's representation,extraction and quantificative comparison.Finally,introduce the verification methods of time consistency and geometric consistency,which are used to eliminate the mismatched loop closure's image,thus ensure loop closure detection with high robustness.The experiments show that,selecting ORB algorithm to extract feature points can construct the BoW model quickly.Adding feature space information to obtain feature space vector,which is utilized to match image in loop closure detection.The method can improve perceived ambiguity and increase the performance of loop closure detection,also meet real-time requirements.