Research On Loop Closure Detection Algorithm Of Visual SLAM

In recent years,the application of autonomous mobile robots has grown rapidly,affecting all aspects of society and playing an increasingly important role in it.Visual simultaneous localization and mapping(SLAM)is the core technology of autonomous mobile robots.Loop closure detection(LCD),one of the key methods,assists the construc-tion of a globally consistent map by identifying whether the object has visited the current position.However,in practical application,changes in the environmental appearance and different viewpoints,caused by the illumination,seasons,and weather,make it difficult for the LCD algorithm to run steadily with high precision for a long time.To tackle these problems,the main research contents in this thesis are as follows:Firstly,to cope with the environmental changes,an end-to-end LCD algorithm(named MetricNet)based on patch mapping similarity matrix is designed in the thesis.The algorithm takes full account of the inherent relevance of feature extraction and similarity calculation,and introduces the channel weighting mechanism to make the feature extraction module focus on more discriminative regions.Meanwhile,the patch mapping similarity matrix is designed to increase the feature constraints of the spatial dimension,so as to construct an adaptive similarity calculation method.Many comparative experiments with several state-of-the-art algorithms are complemented in this thesis,and results verifies the performance of MetricNet on three different datasets(Gardens Point,Nordland,and St.Lucia),and its average precision(AP)is improved by 7.51%,3.03%,and 5.33%,respectively.Secondly,to further improve the performance of the algorithm,an end-to-end LCD algorithm(named DeepMetricNet)based on deep metric learning is designed in this thesis.Based on MetricNet,the deep metric learning mechanism is introduced to learn a transferable similarity metric function through the neural network,multi-level features are fused to enrich the scene space and semantic information,and one-dimensional convolution is used to reduce the computational complexity.Meanwhile,inner product operation is combined to optimize and accelerate the construction of similarity matrix.Compared with MetricNet,the average matching time(10.6ms)of DeepMetricNet is improved by 69.4%,and AP is improved by 9.78%on multiple datasets.In summary,introducing similarity matrix and deep metric learning,MetricNet and DeepMetricNet algorithms which give consideration to both feature extraction and similarity calculation are designed in this thesis.It proves that the joint optimization of these two modules contributes to improving the performance of the algorithm.Experiments show that DeepMetricNet can handle environmental changes effectively,and has high detection efficiency.