Research On Loop Closure Detection Algorithm For Visual SLAM Mobile Robot

Real-time localization and map construction are always the key problems for mobile robots in unknown environment.The errors generated by mobile robots in the process of long time movement will accumulate into the subsequent key frame matching.As a key link in visual SLAM,loop closure detection plays an important role in reducing the accumulated errors in the process of robot movement and back-end optimization.Loop closure detection refers to comparing the similarity between the current image and the historical image obtained by the mobile robot,judging whether the loop closure occurs in the process of the robot moving,and updating the map synchronously according to the detection results to eliminate the accumulated error.In the current research on the loop closure detection algorithm,the loop closure detection algorithm based on visual SLAM is still the mainstream algorithm,and the relevant research results are relatively rich,but there are still many deficiencies.In this paper,the loop closure detection algorithm is optimized and improved by studying and analyzing the bag-of-words model method commonly used in loop closure detection,which is susceptible to the influence of factors such as object change and illumination change in the scene,leading to the problem of misjudgment in loop closure detection.The main work and achievements are as follows:In terms of feature extraction,an improved ORB algorithm based on quad-tree partition was proposed to solve the problems of feature points extracted from the ORB algorithm,which lacked scale invariance and tended to generate stack redundancy.The scale invariance of ORB feature points is increased by building an image pyramid on the images.By the method of quad-tree partition image,the divided into areas of the extracted features respectively,to improve the distribution of the feature points uniformity,and aiming at the problem of excessive quad-tree partitioning,according to the required number of feature points,differentiate set the maximum depth,and according to the correlation between adjacent areas,restricted quad-tree partition depth increase the timeliness and uniformity of the algorithm.Experimental results show that the uniformity and matching accuracy of feature points extracted by the improved ORB algorithm are significantly improved.In loop closure detection,K-medians clustering algorithm is used to optimize the word bag model in order to solve the problem that the clustering algorithm of word bag model is susceptible to noise.Meanwhile,semantic features are introduced to solve the problem that the word bag model does not pay attention to the position and order of words,which is prone to perceptual bias.YOLOv4 algorithm was used to extract semantic features and calculate cosine similarity among various semantic features.By using ORB features to screen candidate loop closures and semantic features to verify candidate loop closures,the computation is reduced.Through experiments on public data set and real shot data set,and comparison with other algorithms,and analysis of algorithm accuracy,recall rate and other evaluation indicators,experimental results show that the improved loop closure detection algorithm performance has been improved.