Research On Loop Closure Detection In Visual SLAM Based On Auto-encoder

Warehousing mobile robots are representative of automation equipment in the intelligent warehouse system of the "Industrial 4.0" era,and are widely used in various material handling tasks.Achieving autonomous navigation of mobile robots is a problem that must be solved to ensure its flexibility and adapt to various work scenarios.Simultaneous Localization and Mapping(SLAM)is the key technology for autonomous navigation of mobile robots.As an important part of the visual SLAM framework,Loop closure detection is the key to reducing the cumulative error of the robot pose and constructing a global consistency map.In this thesis,loop closure detection in visual SLAM is taken as the research object,exploring the feasibility and method of achieving the loop closure detection by extracting image features with auto-encoders and calculating image feature similarity,based on the analysis of the research status of loop closure detection at home and abroad.The main contents of this paper include:(1)Design the structure of the auto-encoder used to extract image feature.The auto-encoder is an unsupervised neural network model that can implement image compression coding and feature extraction.Based on the shallow auto-encoder,the structures of deep auto-encoder and convolutional auto-encoder are designed respectively,and the deep auto-encoder and convolution auto-encoder are added to ability of removing image noise and extracting sparse features by modifying the loss function.(2)The method of loop closure detection based on auto-encoder is designed from three aspects:image feature extraction,similarity score calculation and loops accuracy verification.The TensorFlow library is used to implement and train the auto-encoder designed in this paper to extract features,and a method for calculating image feature similarity score and verifying loops which were found is proposed(3)Verify the performance of the loop closure detection method based on the auto-encoder on the public data set,and compare it with the visual word bag method(BoW)based on the artificial design feature.The experimental results show that the loop closure detection performance of the deep auto-encoder is slightly better than that of BoW,and the loop closure detection performance of the convolutional auto-encoder is better than that of BoW.The loop closure detection method based on auto-encoder proposed in this paper has higher accuracy and recall rate than the "Bag of words" method,and extracting image feature with auto-encoder designed in this thesis takes less time.The loop closure detection method proposed in this paper can better meet the accuracy and real-time requirements of the visual SLAM system,and become an alternative to the "Bag of words" method.