Research On Visual SLAM Closed Loop Detection Algorithm Based On Word Bag Method

Visual-simultaneous location and mapping(V-SLAM)technology is widely used in robots,drones,unmanned driving,AR,VR and other fields.The basis for all types of robots to undertake tasks is to let them Possessing "vision" and relying on sensors can realize the functions of automatic positioning,mapping,path planning and other functions of the machine,thereby obtaining the ability to move freely.In the visual SLAM algorithm,due to the existence of various errors,there is also an error in the pose estimation,and it is transmitted through the recursion of the pose between the frames,so the accumulation of errors is unavoidable,which makes it impossible to construct a consistent map.You can provide more information for back-end optimization through closed-loop detection,and perform global optimization to eliminate the effects of accumulated errors.The key to closed-loop detection is to be able to correctly identify the scenes that have passed through history,and to use effective methods to verify the accuracy of the closed-loop.In view of some problems of visual SLAM back-end optimization and closed-loop detection,this article introduces and expands the work of this article in three aspects: first,in order to solve the problem of feature extraction and aggregation,a feature selection strategy based on quadtree is proposed,It can effectively select evenly distributed feature points and improve the robustness of the features;then,in order to eliminate the traditional visual dictionary,only the quantization of image features on a single scale is considered when matching words,ignoring words with different levels of different representation capabilities Problem,this paper builds a visual dictionary tree with TF-IDF as the weight,proposes to calculate the similarity increment between images from bottom to top layer by layer,and finally establishes a kernel function to integrate the similarity increment of the golden word tower word matching method;finally,In the back-end optimization,a kernel function is proposed to solve the problem of excessive error growth,and the detected closed loop is verified by methods such as time continuity,spatial consistency,and polar geometric constraints to ensure that the closed loop detection is highly robust.Sex.Experiments show that the strategy of selecting feature points in this paper can effectively obtain uniformly distributed feature points,can quickly build a layered calculation of Bo W model,and can effectively detect closed loops by using the similarity incremental gold word tower word matching method,Combined with back-end optimization and closed-loop detection,has achieved good closed-loop identification and verification of the TUM dataset,and can achieve visual SLAM relocation based on the image matching technology of the bagof-words method.