Research On Visual SLAM Loop Closure Detection Technology Of Indoor Mobile Robot

Loop closure detection,which enables significant decrease of the accumulated error in mapping,is a critical module of visual SLAM(Simultaneous Localization and Mapping).Effective and stable loop closure detection is crucial to visual SLAM.The present bag-of-visual-words based approaches have poor robustness,low accuracy,and take up a lot of memory resources.Therefore,this thesis investigates the feature extraction and matching in the loop closure detection process for the above problems,aiming to raise the robustness and precision of the algorithm and heighten the real-time performance of visual SALM system,which possesses distinguished theoretical meaning and worth of real-world utilization.First of all,the present state of research on visual SLAM and loop closure detection at home and abroad is elaborated.The mathematical model and overall framework of visual SLAM system is introduced with analyzing each module of it.The analysis spotlights the fundamentals of loop closure detection and the approaches of image feature extraction and matching in the process,and the following part of the study concentrates on these two aspects.Then,the image feature extraction of loop closure detection is studied deeply in this thesis.Traditional methods of loop closure detection based on bag-of-visual-words extract features with poor robustness and are prone to detect false-positive loop closure,resulting in poor accuracy.In response to this problem,a stacked assorted auto-encoder(SAAE)network based on unsupervised learning is designed for better feature extraction of visual SLAM scene images.Unlike the traditional stacked auto-encoder,the proposed SAAE is composed denoising auto-encoder,convolutional auto-encoder and sparse auto-encoder in multiple layers,combining the advantages of them,with lower dimensionality of the output features and better robustness.The performance of SAAE is tested on New College and City Centre,which are standard datasets of visual SLAM.And the results indicate that the presented methodology is capable of productively raising the precision and robustness of loop closure detection.Subsequently,a fast loop closure detection algorithm based on improved iterative quantization is proposed in response to the problem that traditional methods of loop closure detection based on bag-of-visual-words usually use the cosine distance to measure the similarity during feature matching,which leads to high memory resource overhead and long time consumption.By improving the projection matrix and orthogonal rotation matrix of the iterative quantization,the time and space complexity of the calculation are further reduced.The improved iterative quantization is used to map the features of visual SLAM scene images extracted by the proposed SAAE to hash codes,and the similarity is calculated by weighted Hamming distance,thereby speeding up the feature matching process of loop closure detection.The performance of the fast loop closure detection algorithm based on improved iterative quantization is tested on New College and City Centre,which are standard datasets of visual SLAM.And the experimental results show that the algorithm can effectively improve the time performance of loop closure detection.Finally,a mobile robot platform of visual SLAM is established with related equipment,and the design of loop closure detection algorithm module is completed.With the purpose of testing the performance of the presented approach,the visual SLAM map construction experiment of mobile robot is indoor conducted.The results indicate that the presented methodology enables significant decrease of the accumulated error in mapping,and productively heighten the real-time performance of visual SLAM system.