Research On Visual SLAM Loop Closure Detection Based On Machine Learning

With the further development of mobile robot system and the constant integration of 3D imaging technology,visual Simultaneously Localization And Mapping(visual SLAM)has been highly valued by governments,society and enterprises,and also attracted active participation of numerous producers in the relevant industry chain and links.Visual SLAM acquires image information through different specific cameras to achieve functions such as establishing the environment while estimating its own movement during the motion without the prior information of the environment.The key application basis of visual SLAM industrialization process lies in how to ensure and improve the mapping accuracy of mobile robot under the interference of complex environment.The intelligent mobile robot estimates its movement relying on its own location and visual sensor in the unknown environment,in the meantime,positions itself and establishes the map of its surrounding environment in a smart way,of which the process is called Simultaneously Localization And Mapping(SLAM),namely positioning and mapping at the same time.Compared with laser SLAM,visual SLAM is featured with simpler and lighter structure,no limit on probing distance of sensor,low cost,extractable semantic information and considerable market potential,therefore,visual SLAM has attracted great attention from various industries today.Visual SLAM is mainly divided into five major parts: sensor data,visual odometer,back-end optimization,loop closure detection and final mapping.These five parts are closely linked with each other,for example,during visual odometer,the intelligent mobile robot will constantly accumulate errors,whether it is caused by visual sensor or manual work,which is unavoidable.At this time,it is necessary to detect the entire map of the robot through loop closure detection.When the loop closure effect is detected,mapping errors can be reduced through loop closure location,thereby increasing the robustness and accuracy of the whole map.The traditional loop closure detection method is mainly the Bag of Words(Bo W),which is a relatively concise method with good operating process and experimental result.However it requires lots of features set artificially to improve and perfect the algorithm,which greatly confines the effect of the whole SLAM.Additionally,there will be deviation for the effect when the lighting is changing obviously,therefore a new idea is needed to study on loop closure detection.Now,deep learning technology has been applied in many aspects,mainly including image classification,data fitting,etc.Loop closure detection is,in principle,a classification process of image,thus,it is theoretically a feasible and more effective way to apply deep learning technology into loop closure detection.This paper proposes to apply multiple convolutional neural networks(CNN)of deep learning to loop closure detection of visual SLAM.Taking CNN networks as image processor,the author will only care about the input and output of the processor.Different CNN networks can get their respective output.The author will set output as one-dimensional vector,then process the output to get more effective and concise data,and finally compare the similarity of these data through loop closure detection to show corresponding results in various ways.This paper provides a new and innovative method of loop closure detection for the study of visual SLAM,and its main research work and results are as follows:1.This paper describes the five major parts of visual SLAM,including sensor data,visual odometer,back-end optimization,loop closure detection and final mapping,and puts more emphasis on the relevant theories of loop closure detection.2.This paper elaborates the mathematical description of motion estimation of visual SLAM and the theories of Bo W,the traditional method of loop closure detection,and obtains the corresponding motion model and observation model mainly through describing the mathematical theories in the process of robot motion and observation in visual SLAM to solve the problems appearing in the whole SLAM.3.This paper makes a detailed description on the algorithm of Bow,a loop closure detection method based on artificial design feature,and analyzes a series of methods of feature descriptor,mainly including SIFT,SURF and ORB,among which,the detail of ORB algorithm is mainly illustrated.This paper also introduces the wholeimplementation method of Bo W algorithm,including the formation of the traditional Bo W algorithm,the algorithm process of Bo W loop closure detection.Moreover,this paper carries out experiment to show the results through multiple data,thereby fully displaying the advantages and disadvantages of the Bo W algorithm of the traditional loop closure detection.4.This paper proposes the combination of deep learning and visual SLAM so as to apply a variety of CNN networks into loop closure detection of visual SLAM,and provides a method for laboratory use to improve loop closure detection,that is to train different CNN architectures and process the output data with the methods of normalization,PCA and albefaction,etc.,then calculate the distance between two images with Normalized Euclidian Distance,and finally carry out experiment on the standard data set of New College and City Centre.This experiment compares the traditional algorithm(Bo W)based on artificial design features and several CNN models of before-and-after improvement.Based on the experimental results,the multiple CNN networks have more advantages in similar matrix,PR curve,average accuracy and feature extraction time performance during loop closure detection by adopting improved algorithm.