Application Of Binary Features In Loop Closure Detection

With the rapid development of technologies such as computer vision and robot design,the development of robots with autonomous capabilities has become one of the hot research topics in the field of artificial intelligence today,simultaneous localization and mapping(SLAM)is the key technology to realize robot autonomy.When the robots do simultaneous localization and mapping,they need to determine whether the current location has been visited before,that is the problem of loop closure detection in SLAM.The main task of loop closure detection is to search the database to find the most similar picture to the current picture,and determine whether these two pictures come from the same scene.This thesis first introduces the background and significance of the research,describes the brief algorithms of SLAM,as well as the models and related problems of loop closure detection in these algorithms.Provides the theoretical model for the research of loop closure detection.Then,we introduce the commonly used methods of image feature extraction,ORB,SIFT,SURF extraction methods and their advantages,describes their description methods,and according to their characteristics design their binary operation.Thirdly,aiming at the problem of large amount of information and large number of image features extracted in images,using the visual word bag model(BoVW),the extracted visual features are converted into visual words and stored in the visual dictionary.According to the characteristics of binary visual words,we modify the K-means structure used by the traditional method,get a new visual dictionary,improve the dictionary's storage efficiency and the speed of words matching.Finally,we use the TF-IDF score criterion to calculate the similarity probability between the pictures,and send the probability to Bayesian filter.According to the characteristics of video images,we adjust the relevant parameters in the filter,propose a loop closure detection with higher scene adaptability,then complete the image similarity calculation.By introducing the epipolar geometric constraint and random sample consensus,we improve the accuracy of detection.The effectiveness of the algorithm was verified in several datasets experiments.