A laser SLAM loop detection method based on multi-modal information fusion

Abstract

This invention discloses a laser SLAM (Simultaneous Localization and Mapping) loop closure detection method based on multimodal information fusion, belonging to the field of laser SLAM technology. Addressing the shortcomings of existing technologies in complex scenarios, such as false detections and missed detections, this method constructs a hierarchical dual-channel collaborative decision-making framework: First, it extracts the Scan-Context global appearance descriptor from the laser point cloud and quickly filters candidate loop closure frames through cyclic shift matching; then, it uses a point cloud registration algorithm to calculate the geometric fitting error, which is converted into normalized geometric similarity using a Gaussian kernel function; finally, based on a Bayesian inference framework, it achieves adaptive weighted fusion of appearance and geometric features in the form of logarithmic odds, dynamically adjusts the fusion weights according to the scene, determines loop closures through a fusion confidence threshold, and outputs pose constraints, correcting accumulated errors through factor graph optimization. This invention improves the accuracy and robustness of loop closure detection and can be widely applied to autonomous navigation scenarios such as autonomous driving and mobile robots.

Description

Technical Field

[0001] This invention belongs to the field of laser SLAM (Simultaneous Localization and Mapping) technology, specifically relating to a laser SLAM loop closure detection method based on multimodal information fusion. Background Technology

[0002] In Simultaneous Localization and Mapping (SLAM) systems, loop closure detection is a key technology for eliminating accumulated errors and ensuring global map consistency. Current loop closure detection methods in laser SLAM mainly fall into two categories: methods based on global appearance descriptors (such as ScanContext) and methods based on local geometric features (such as Euclidean distance). Methods like ScanContext, by constructing rotation-invariant scene descriptors, exhibit good retrieval efficiency in structured environments, but are prone to false detections in scenarios such as long corridors and repetitive buildings. While geometric matching methods like Euclidean distance are insensitive to changes in scene appearance, they are prone to missed detections under conditions of large viewpoint changes or dynamic occlusion, and also incur significant computational overhead.

[0003] Existing fusion methods often employ heuristic weighted or cascaded decision-making, lacking a rigorous theoretical framework to guide weight allocation. This results in an inability to adaptively adjust the confidence levels of different modal information in complex scenarios. Consequently, the system struggles to balance detection accuracy and recall in practical applications, limiting the robustness of SLAM systems in fields such as autonomous driving and mobile robotics. Summary of the Invention

[0004] In view of this, the purpose of this invention is to provide a laser SLAM loop closure detection method based on multimodal information fusion, aiming to solve the problems of false detection, missed detection, and poor adaptability caused by the lack of theoretical fusion basis in existing loop closure detection methods in complex scenarios. The core of this invention lies in constructing a probabilistic fusion model and adaptive decision-making mechanism based on Bayesian inference. Its advantages include rigorous theory and strong interpretability, and the ability to dynamically adjust the fusion strategy according to scene characteristics. The system is highly flexible and can adapt to changes in different environments and sensor configurations without requiring a redesign or significant adjustment to the overall framework of the detection algorithm.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A laser SLAM loop closure detection method based on multimodal information fusion includes the following steps:

[0007] S1. Global appearance quick filtering;

[0008] The input laser point cloud frame is processed to extract the global appearance descriptor Scan-Context, which is obtained by projecting the 3D point cloud into a 2D matrix. ;in, The number of divisions in the radial direction represents the height dimension of the scene; The number of divisions in the angular direction represents the density dimension of the scene; express It is OK, A real matrix of columns; the global appearance descriptor encodes the height and density information of the scene; the current frame descriptor in the column direction. With historical frame descriptors Perform cyclic shift matching and calculate the appearance similarity score according to the formula. Select those with values ​​higher than the initial screening threshold T. prune The set of candidate frames;

[0009]

[0010] In the formula, This represents the Scan Context feature matrix of the current frame; The Scan Context feature matrix representing historical frames; This is the result of performing k cyclic shifts on the feature matrix of historical frames; k represents the number of cyclic shifts; N represents the total number of elements in the feature matrix, used for normalization. This indicates taking the value with the smallest difference among all shift counts;

[0011] S2. Precise Local Geometric Authentication;

[0012] For each candidate frame in the candidate frame set, a point cloud registration algorithm is used to calculate the geometric fitting error between the current frame and the candidate frame. Then, use the Gaussian kernel function to... Converted to normalized geometric similarity score ;

[0013]

[0014] In the formula, This represents the tolerance parameter for geometric matching, used to control the width of the Gaussian kernel;

[0015] S3. Multimodal probability fusion;

[0016] Based on the Bayesian inference framework, and As independent observational evidence under the loop closure hypothesis, the fusion confidence level is calculated using logarithmic odds and is equivalently converted into a weighted fusion model. ;

[0017]

[0018] In the formula, Represents the constant bias term; Indicates the fusion weight, fusion weight It will dynamically adjust according to the characteristics of the scene;

[0019] S4. Loop Final Decision: Integrate Confidence Scores Compared with the preset decision threshold, if If the decision threshold is exceeded, a loop closure is determined to be valid. The loop closure matching pair and relative pose constraints are output to the SLAM backend. Accumulated errors are corrected through factor graph optimization to ensure global accuracy. Figure 1 To the point of being responsive.

[0020] Furthermore, the cyclic shift matching in step S1 is performed along the column direction of the two-dimensional matrix. Perform shift operations of different lengths, one by one with... After similarity calculation, the minimum value is selected as the matching result.

[0021] Furthermore, the point cloud registration algorithm in step S2 is either the Iterative Closest Point (ICP) algorithm or the Normal Distribution Transform (NDT) algorithm.

[0022] Furthermore, in step S3, the fusion weights The dynamic adjustment rule is: in well-structured scenarios, increase To improve The fusion weights; in scenes with rich geometric features but repetitive textures, reduce the fusion weights. To improve The fusion weight.

[0023] The present invention also provides a laser SLAM loop closure detection system based on multimodal information fusion for performing the above method, the system comprising:

[0024] Global appearance filtering module: used to extract the global appearance descriptor of the laser point cloud and match it with historical frames to filter out a set of candidate frames;

[0025] Local geometric authentication module: used to perform point cloud registration on candidate frames and calculate geometric similarity scores;

[0026] Bayesian Fusion Module: Used to perform weighted fusion of appearance similarity score and geometric similarity score based on the Bayesian inference framework to generate a fusion confidence score;

[0027] The loop closure decision module compares the fusion confidence score with a threshold to determine whether a loop closure is valid, and outputs pose constraints to the SLAM backend.

[0028] Beneficial effects:

[0029] 1. Unified Probabilistic Fusion Framework: This invention constructs a unified probabilistic framework based on Bayesian inference, which can simultaneously fuse detection information from two complementary modalities: scene appearance (Scan Context) and local geometry (Euclidean distance), solving the problem of the lack of theoretical support in traditional linear weighting.

[0030] 2. Dual-channel collaborative mechanism: The appearance information processing channel utilizes the rotation invariance of global descriptors to achieve fast scene retrieval and initial screening (high recall), while the geometric information processing channel utilizes Euclidean distance matching between feature points to provide accurate pose constraint verification (high accuracy).

[0031] 3. Adaptive weight decision: The core fusion module can dynamically calculate the confidence weights of appearance and geometric information based on the feature distribution of the scene (such as the degree of environmental degradation), and finally output an optimal and interpretable loop closure judgment result, which can effectively adapt to dynamic and repetitive texture environments.

[0032] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0033] Figure 1 This is a flowchart of a laser SLAM loop closure detection method based on multimodal information fusion according to the present invention. Detailed Implementation

[0034] To make the technical solutions, advantages, and objectives of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the protection scope of this application.

[0035] The core working principle of the invention lies in constructing a hierarchical, dual-channel collaborative processing decision-making system, which performs theoretically optimal fusion at the top level. Its process is as follows: Figure 1 As shown, this can be compared to a sophisticated information filtering and verification mechanism.

[0036] This invention discloses a laser SLAM loop closure detection method based on multimodal information fusion, comprising the following steps:

[0037] S1. First-level processing: Global appearance-based rapid filtering (high recall guarantee)

[0038] The system first processes the input laser point cloud frame to extract its global appearance descriptor, Scan Context. This descriptor projects the 3D point cloud into a 2D matrix. This encodes the scene's height and density information. To overcome viewpoint rotation, the system modulates the current frame descriptor along the column direction. With historical frame descriptors Perform cyclic shift matching. Appearance similarity score. Defined as:

[0039]

[0040] This step acts like a highly efficient "coarse sieve," quickly eliminating irrelevant scenes with drastically different appearances and filtering out those with a similarity higher than the initial screening threshold. The set of candidate frames.

[0041] S2. Second-level processing: Local geometric precision authentication and probabilistic fusion (high-precision decision making)

[0042] For candidate frames that pass the first-level screening, the system initiates a more precise local geometric feature extraction and matching process. Point cloud registration algorithms (such as ICP or NDT) are used to calculate the geometric fitting error (mean Euclidean distance) between the current frame and the candidate frames. To standardize the units of measurement for easier subsequent fusion, a Gaussian kernel function is used to transform the geometric distance into a normalized geometric similarity probability. :

[0043]

[0044] in, This is the tolerance parameter for geometric matching.

[0045] S3. Third-level processing: Probabilistic fusion based on Bayesian inference (core decision)

[0046] This module, based on a Bayesian inference framework, treats the two types of information as independent observational evidence under the loopback hypothesis. Let... Assuming the existence of loops, the observed data are... The system calculates the posterior probability that the current observation data supports the loop closure hypothesis. To simplify the calculation and enhance numerical stability, the final fusion confidence is calculated using log-Odds form:

[0047]

[0048] In the formula, Indicates that when observed and Under the condition that "a loop exists", the posterior probability is given. Indicates that when observed and Under the condition that "there is no loop", the posterior probability of "no loop".

[0049] Based on the independence assumption, the derivation is ultimately equivalent to the weighted fusion model:

[0050]

[0051] in, Represents the constant bias term; Indicates the fusion weight, fusion weight It is not a fixed value, but is dynamically determined by the observation uncertainty (variance) of the two modes in the current scenario.

[0052] (1) When in a well-structured scene (such as a teaching building area), the appearance descriptor has high confidence. Automatically increases in size.

[0053] (2) When in a scene with rich geometric features but repetitive textures (such as a long straight corridor), the appearance differentiation is low. The system automatically decreases distance, relying more on geometric distance for judgment.

[0054] S4. Final Decision

[0055] The generated fusion confidence score The score is compared with a threshold. If the score exceeds the threshold, a loop closure is determined, and the system outputs the matching pair and relative pose constraints to the SLAM backend for factor graph optimization, thereby correcting accumulated errors and ensuring global accuracy. Figure 1 To the point of being responsive.

[0056] This invention also discloses a laser SLAM loop closure detection system based on multimodal information fusion, used to perform the above method, the system comprising:

[0057] Global appearance filtering module: used to extract the global appearance descriptor of the laser point cloud and match it with historical frames to filter out a set of candidate frames;

[0058] Local geometric authentication module: used to perform point cloud registration on candidate frames and calculate geometric similarity scores;

[0059] Bayesian Fusion Module: Used to perform weighted fusion of appearance similarity score and geometric similarity score based on the Bayesian inference framework to generate a fusion confidence score;

[0060] The loop closure decision module compares the fusion confidence score with a threshold to determine whether a loop closure is valid, and outputs pose constraints to the SLAM backend.

[0061] It is hereby declared that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A laser SLAM loop closure detection method based on multimodal information fusion, characterized in that, Includes the following steps: S1. Global appearance quick filtering; The input laser point cloud frame is processed to extract the global appearance descriptor Scan-Context, which is obtained by projecting the 3D point cloud into a 2D matrix. ;in, The number of divisions in the radial direction represents the height dimension of the scene; The number of divisions in the angular direction represents the density dimension of the scene; express It is OK, A real matrix of columns; the global appearance descriptor encodes the height and density information of the scene; the current frame descriptor in the column direction. With historical frame descriptors Perform cyclic shift matching and calculate the appearance similarity score according to the formula. Select those with values ​​higher than the initial screening threshold T. prune The set of candidate frames; In the formula, This represents the Scan Context feature matrix of the current frame; The Scan Context feature matrix representing historical frames; This is the result of performing k cyclic shifts on the feature matrix of historical frames; k represents the number of cyclic shifts; N represents the total number of elements in the feature matrix, used for normalization. This indicates taking the value with the smallest difference among all shift counts; S2. Precise Local Geometric Authentication; For each candidate frame in the candidate frame set, a point cloud registration algorithm is used to calculate the geometric fitting error between the current frame and the candidate frame. Then, use the Gaussian kernel function to... Converted to normalized geometric similarity score ; In the formula, This represents the tolerance parameter for geometric matching, used to control the width of the Gaussian kernel; S3. Multimodal probability fusion; Based on the Bayesian inference framework, and As independent observational evidence under the loop closure hypothesis, the fusion confidence level is calculated using logarithmic odds and is equivalently converted into a weighted fusion model. ; In the formula, Represents the constant bias term; Indicates the fusion weight. It will dynamically adjust according to the characteristics of the scene; S4. Loop Final Decision: Integrate Confidence Scores Compared with the preset decision threshold, if If the decision threshold is exceeded, a loop closure is determined to be valid. The loop closure matching pair and relative pose constraints are output to the SLAM backend. The accumulated error is corrected through factor graph optimization to ensure global map consistency.

2. The laser SLAM loop closure detection method based on multimodal information fusion according to claim 1, characterized in that, The cyclic shift matching in step S1 is performed along the column direction of the two-dimensional matrix. Perform shift operations of different lengths, one by one with... After similarity calculation, the minimum value is selected as the matching result.

3. The laser SLAM loop closure detection method based on multimodal information fusion according to claim 1, characterized in that, The point cloud registration algorithm in step S2 is either the Iterative Closest Point (ICP) algorithm or the Normal Distribution Transform (NDT) algorithm.

4. The laser SLAM loop closure detection method based on multimodal information fusion according to claim 1, characterized in that, In step S3, the fusion weights are... The dynamic adjustment rule is: in well-structured scenarios, increase To improve The fusion weights; in scenes with rich geometric features but repetitive textures, reduce the fusion weights. To improve The fusion weight.

5. A laser SLAM loop closure detection system based on multimodal information fusion, used to perform the method according to any one of claims 1-4, the system comprising: Global appearance filtering module: used to extract the global appearance descriptor of the laser point cloud and match it with historical frames to filter out a set of candidate frames; Local geometric authentication module: used to perform point cloud registration on candidate frames and calculate geometric similarity scores; Bayesian Fusion Module: Used to perform weighted fusion of appearance similarity score and geometric similarity score based on the Bayesian inference framework to generate a fusion confidence score; The loop closure decision module compares the fusion confidence score with a threshold to determine whether a loop closure is valid, and outputs pose constraints to the SLAM backend.

Images