Research On Place Recognition Algorithm Based On Deep Learning

Place recognition is a very challenging problem in the field of computer vision.At the same time,place recognition is also an important part of Loop Closure,the core module of the Simultaneous Localization and Mapping(SLAM).Place recognition is one of the key components of the autonomous navigation task of mobile robots,it plays an important role in SLAM technology for global positioning and global correction of maps.and it is also one of the current hot research directions.Place recognition tasks often encounter severe environmental changes such as light,weather,and angles.Place recognition algorithms based on traditional manual features(such as SIFT,ORB,etc.)often have difficulty achieving ideal results.With the great success of deep learning in the field of computer vision,this paper uses the powerful representation capabilities of Convolution Neural Network(CNN)features pre-trained on large data sets as feature descriptors for place images.Proposed a method of extracting image features using convolutional neural network,and combining the Regional Maximum Activation of Convolutions(RMAC)and Hierarchical Navigable Small World graphs(HNSW).Aiming at the problem that the time complexity of the similarity calculation of CNN features with too high dimensions is too high,this paper introduces the RMAC algorithm to reduce the dimensionality of the CNN feature descriptors extracted from the place image to reduce the time complexity of the similarity calculation,thereby improving the The real-time nature of the place recognition algorithm,and in order to avoid the performance degradation of the algorithm caused by the loss of information from CNN feature dimensionality reduction,the algorithm introduces the process of rough matching using dimensionality-reduced features and accurate matching of non-dimensionality reduction features Improve the accuracy of the algorithm.At the same time,in order to solve the problem that the time complexity of the similarity matching query between the current place image feature descriptor and the previously passed place image descriptor is too high,a fast place recognition algorithm based on HNSW is proposed,thereby greatly improving the real-time performance of the algorithm.In addition,this paper proposes a fast loop detection algorithm for loopdetection tasks.In order to avoid detecting false loops,a timing judgment and sequence matching process is added to the algorithm flow to improve the accuracy of the algorithm.In order to verify the effectiveness of the algorithm proposed in this paper,a comparative experiment is performed on the currently disclosed challenging place recognition datasets,and the experiment proves that the algorithm proposed in this paper reduces the accuracy of a small part of the algorithm within an acceptable range.Under the premise,the algorithm's real-time performance has been greatly improved,and it has been proved that the algorithm proposed in this paper has certain significance and value in theoretical innovation and practical application.I hope that the research content in this paper can make future Place recognition research development Some inspiration.