Research On Loop Closure Detection Algorithm Based On Improved Bag Of Words Model

SLAM(Simultaneous Localization and Mapping)refers to a technology in which devices acquire environmental data through their own sensors(such as cameras,inertial measurement units and lidar)during movement,and then use the data to build an environmental map and conduct independent positioning.Visual Simultaneous Localization and Mapping system refers to a SLAM system in which the sensor for data acquisition is a camera.At present,the visual SLAM system is usually divided into two parts: the front end and the back end.The front end is mainly a visual odometer,whose function is to estimate camera motion based on the images captured by the camera.Due to the inevitable errors in front-end estimation,which will accumulate continuously,the motion estimation of the system after a long run is not accurate,so it is necessary to optimize the back-end.Before the appearance of PTAM system,the back-end optimization scheme usually used a Filter dominated by Extended Kalman Filter(EKF).Starting from PTAM system,the nonlinear optimization scheme began to replace the Filter.In order to avoid serious deviation in the convergence of back-end optimization,the loop closure detection module is introduced in the SLAM system.The return loop closure detection module is a relatively independent module in the current SLAM system,whose main purpose is to identify the scene that the equipment has been to(that is,to form a return loop),and to reduce the cumulative error generated by the visual odometry.Accurate loop detection can effectively improve the mapping quality of visual SLAM system.The word bag model was originally used in the field of natural language processing.Due to its simplicity and flexibility,more and more attention has been paid to it in the computer vision neighborhood.The loop closure detection algorithm based on the bag of words model is one of the applications.In order to improve the performance of the loop closure detection algorithm,this dissertation mainly carried out the following two aspects of research work:(1)Real-time update of the vocabulary tree : Traditional visual bag-of-words model words tree is usually in the offline environment,use has nothing to do with the experimental scenario of a large number of images generated by clustering algorithm,this leads to the information contained in the dictionary and experimental scene information difference is bigger,but the experiment scene image data is unknown,could not use the scene data dictionary.Therefore,by changing the clustering algorithm of dictionary generation and introducing the fusion method of image data,the dictionary can update the data of experimental scenes in the process of loop closure detection.(2)The algorithm based on the modified bag model is proposed: In this dissertation,through the analysis of the loop closure detection process,it is found that the loop closure detection algorithm needs to carry out the same degree of complexity calculation on all the querying images when searching the candidate images for loop closure detection.However,it can be found that the number of the same words in different images is different,and the number of the same words in similar images is significantly greater than the number of the same words in different images.Therefore,this dissertation chooses to simply use the words of the image to filter out a part of the image that is different from the new image,and then carries out the subsequent search process.In this way,two filtering steps with different complexity are used to improve the running efficiency of the algorithm.(3)Comparative experiments of improved algorithms were conducted on the City Center dataset,New College dataset,KITTI00 dataset and KITTI06 dataset.It is verified that the improved algorithm has an average increase in the maximum recall rate of 19.88% with an accuracy of 100% and an average increase in the average recall rate of 43.125% compared with the algorithm before the improvement.The time required for the improved algorithm to process an image is 162.75 ms shorter than the time before the algorithm is improved,and the average time performance is improved by 41.25%.(4)The improved algorithm in this dissertation is compared with other five classic algorithms.The maximum recall rates on the New College dataset,KITTI00 dataset,and KITTI06 dataset were improved by 0.82%,2.69%,and 23.59%,respectively,compared to the best results of the other five algorithms.It proves that the algorithm of this dissertation has achieved the best results on the three datasets.