Loopback detection system and method based on layered characteristics and growth type memory mechanism

Abstract

The invention discloses a loopback detection system and method based on layered features and a growth type memory mechanism, and relates to the technical field of robot positioning and autonomous navigation. According to the method, a three-layer feature layering mechanism is adopted, and the efficiency and precision of loopback detection are improved through progressive matching of coarse-layer rapid search, middle-layer transition optimization and fine-layer precise verification; the characteristics are gradually refined and optimized along with the operation of the system by adopting a growth type memory mechanism, so that the problem of characteristic degradation in long-term operation of a traditional method is effectively solved; dynamic management and optimization of features are achieved through a three-layer memory structure and a self-adaptive updating strategy, and the adaptability of the system to environment changes is improved while the compactness of a memory bank is kept; and multi-modal feature fusion and strict calibration and filtering processing are adopted, so that the robustness of the system in a complex dynamic environment is enhanced.

Description

A loop closure detection system and method based on hierarchical features and growing memory mechanism Technical Field

[0001] This invention relates to the field of robot localization and autonomous navigation technology, specifically to a loop closure detection system and method based on hierarchical features and a growing memory mechanism. Background Technology

[0002] With the continuous development of robot autonomous navigation, unmanned driving, and 3D environment mapping technologies, localization and perception systems based on Simultaneous Localization and Mapping (SLAM) have become a core research direction. Among them, loop closure detection, as a key component of SLAM systems, is used to identify the same scene visited by the robot at different times, so as to correct the accumulated drift error and maintain the consistency of the global map.

[0003] Most existing loop closure detection methods rely on fixed feature representations, such as single-layer feature matching based on image appearance descriptors or point cloud geometry. This static feature representation is prone to feature degradation, fuzzy matching, and strong environmental dependence in long-term operation or large-scale scenarios, leading to a decrease in the stability and accuracy of loop closure detection. Furthermore, existing methods generally lack hierarchical modeling and dynamic update mechanisms for features. The system cannot refine or optimize features based on experience during multiple loop closures, thus limiting its long-term adaptability and robustness.

[0004] In recent years, researchers have begun to explore learning-based feature description and cross-temporal feature fusion methods to improve the stability of loop closure detection in complex scenarios. For example, some scholars have proposed using deep convolutional networks to extract multi-scale features or introducing attention mechanisms to optimize feature matching. However, these methods still have limitations such as static feature updates and difficulty in continuous model evolution. Most systems still rely on fixed feature storage structures during operation, cannot dynamically adjust feature representations based on loop closure detection history, and lack hierarchical memory organization and progressive optimization strategies.

[0005] While existing loop closure detection techniques have made some progress in accuracy and real-time performance, they still have significant shortcomings in long-term operation, multi-scenario variation, and feature memory optimization. Current methods generally lack a mechanism that can progressively optimize feature representation based on system experience and possess long-term memory and feature growth capabilities, making it difficult to maintain high accuracy and robustness in long-term tasks. Therefore, there is an urgent need to propose a loop closure detection method that can achieve hierarchical feature modeling and memory growth updates to enable adaptive feature evolution and long-term optimization, thereby improving the system's global consistency and detection accuracy.

[0006] In conclusion, the loop closure detection method based on hierarchical features and growing memory mechanism is of great significance for solving the above problems. Summary of the Invention

[0007] The invention aims to solve the problems of static feature representation, decreased matching robustness and feature degradation in existing loop closure detection methods. It provides a high-precision and high-robust loop closure detection system and method, which realizes the gradual growth and dynamic evolution of features and improves the long-term stability and detection accuracy of the system in complex environments.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a loop closure detection system based on hierarchical features and a growing memory mechanism, comprising a data acquisition and preprocessing unit, a pose estimation and keyframe generation unit, a feature extraction unit, a multimodal feature fusion unit, a feature filtering unit, a feature growth unit, a candidate generation unit, a matching verification unit, a pose optimization unit, an adaptive update unit, and a memory management unit.

[0009] The data acquisition and preprocessing unit is used to acquire data from multiple sensors and perform time synchronization, spatial calibration, noise filtering, and format unification.

[0010] The pose estimation and keyframe generation unit estimates the robot pose in real time based on standardized multi-sensor data. By setting motion distance or angle thresholds, keyframes are filtered to reduce redundant data and focus on core effective information, providing accurate data support for subsequent feature extraction and loop closure detection.

[0011] The feature extraction unit is used to extract geometric and semantic features. The multimodal feature fusion unit maps the two types of features to a unified feature space and divides them into coarse, medium, and fine features according to scale and semantic level to construct a hierarchical feature system. The multimodal feature fusion unit includes a coarse feature extraction module, a medium feature fusion module, and a fine feature refinement module. The coarse feature extraction module is used to extract coarse features, the medium feature fusion module is used to extract medium features, and the fine feature refinement module is used to extract fine features.

[0012] The feature filtering unit dynamically removes redundant features based on feature confidence and access frequency, keeping the feature set compact and effective, avoiding invalid features from occupying resources, and improving the efficiency of subsequent matching and detection.

[0013] The feature growth unit is used to gradually evolve the features from coarse to fine after loop closure confirmation;

[0014] The candidate generation unit uses coarse-layer features to quickly retrieve potential loop closure candidates, and narrows the matching range through efficient preliminary screening, thereby reducing the computational load for subsequent accurate verification and ensuring the real-time performance of loop closure detection.

[0015] The matching verification unit combines fine-layer features to accurately verify candidate loop closures. During the system initialization phase, the fine-layer feature refinement module of the multimodal feature fusion unit directly generates initial fine-layer features and stores them in the memory management unit. In subsequent operation, the feature growth unit is only responsible for the fusion and update of existing fine-layer features by semantic channel expansion and geometric key point completion. The matching verification unit confirms the loop closure relationship by calculating the similarity of the fine-layer features corresponding to the candidate loop closures, filters out false matches, and ensures the accuracy of loop closure detection.

[0016] The pose optimization unit is used to incorporate the confirmed loop constraints into the graph optimization framework, correct the cumulative drift error generated during robot operation, maintain global map consistency, and improve the localization and mapping accuracy of the SLAM system.

[0017] The adaptive update unit dynamically adjusts the feature growth rate and screening threshold based on the detection confidence and environmental changes;

[0018] The memory management unit constructs a three-layer structure consisting of an immediate feature buffer layer, an evolutionary feature buffer layer, and a global feature map layer, to achieve hierarchical storage and dynamic management of features.

[0019] The memory management unit, feature growth unit, and multimodal feature fusion unit constitute an integrated structure of feature layering and memory growth.

[0020] Furthermore, the data acquisition and preprocessing unit includes a sensor initialization module, a timestamp acquisition module, a time synchronization module, a data acquisition module, a spatial calibration module, a sensor alignment module, a noise filtering module, an outlier detection module, a data cleaning module, a data standardization module, and a feature preparation module.

[0021] Furthermore, the application process of the data acquisition and preprocessing unit is as follows:

[0022] First, complete sensor initialization, data acquisition, and time synchronization:

[0023] The data acquisition module is linked with the sensor initialization module to configure parameters for the camera, lidar, and IMU. The timestamp acquisition module marks the timestamp based on the UTC time of RTK-GNSS. The time synchronization module achieves nanosecond-level synchronization through the GNSS+PTP protocol to form a multi-sensor time synchronization architecture, ensuring that the time error is ≤1ms.

[0024] Next, spatial calibration, data cleansing, and alignment will be carried out:

[0025] The spatial calibration module uses the checkerboard method, dynamic trajectory optimization, and the Kalibr toolkit to complete the calibration of multi-sensor extrinsics.

[0026] Sensor alignment module Map the data to the world coordinate system to ensure a spatial overlap rate of ≥95%;

[0027] The noise filtering module is used for targeted filtering;

[0028] The outlier detection module removes outlier data based on the σ principle, while the data cleaning module interpolates and adds a small number of outliers, removes a large number of outliers, and records the data in a log.

[0029] Finally, data standardization and feature preparation are performed:

[0030] Data standardization module Map the features to the range [0,1]. These are the standardized eigenvalues. The minimum value in the original feature data. The maximum value in the original feature data;

[0031] The feature preparation module extracts point cloud geometric features, image visual features, and IMU motion features, classifies and stores them in a buffer, laying the foundation for multimodal fusion.

[0032] Furthermore, the time synchronization module is provided with multiple sub-modules, including an RTK-GNSS module, a PTP time server, an MCU module, an execution module, a time synchronization / delay correction module, and a unified timestamp output module;

[0033] The RTK-GNSS module is used to provide an absolute time reference and serves as the data source support for time synchronization.

[0034] The PTP time server and MCU module are the core execution components of the time synchronization module, realizing the generation of time synchronization signals and synchronous triggering.

[0035] The time synchronization / delay correction module is an error correction submodule of the time synchronization module, which ensures synchronization accuracy through Kalman filtering;

[0036] The execution module includes a camera, a lidar, and an IMU;

[0037] The unified timestamp output module connects time synchronization with subsequent data processing, serving as the final step in data preprocessing and ensuring the consistency of output data time.

[0038] Furthermore, the application process of the time synchronization module is as follows:

[0039] The RTK-GNSS module receives high-precision UTC time signals from the Global Navigation Satellite System, which serve as the absolute time reference for the entire synchronization architecture, and outputs time references to the PTP time server.

[0040] The PTP time server generates standardized timing signals based on this benchmark and transmits them to the MCU module.

[0041] The MCU module acts as the control center, converting the timing signal into a synchronous trigger command, which is then sent to the camera, lidar, and IMU to drive each sensor to collect data at a uniform time rhythm.

[0042] Then, the time synchronization / delay correction module monitors the timestamp deviation of each sensor data in real time, and uses the Kalman filter algorithm to dynamically correct the device clock drift and data transmission delay, ensuring that the timestamp consistency error of the camera, lidar and IMU is ≤1ms.

[0043] The original data from each sensor, after time delay correction, are uniformly input into a unified timestamp output module, where standardized timestamps are added and the data is categorized and organized according to sensor type. The final output is a multi-source dataset with consistent time synchronization.

[0044] Furthermore, the hierarchical feature system achieves precise feature optimization through a three-level progression of coarse, medium, and fine layers;

[0045] The coarse-layer feature extraction module includes a data pre-filtering submodule, a global feature encoding submodule, a spatial relationship modeling submodule, a coarse feature fusion submodule, and a coarse-layer feature output module;

[0046] The coarse feature extraction module first uses the feature alignment module to suppress point cloud noise with adaptive Gaussian filtering, then the global feature encoding submodule generates a 512-dimensional global feature vector, the coarse feature fusion submodule dynamically weights the vectors according to the scene, and finally the coarse feature output module normalizes the output.

[0047] The mid-layer feature fusion module includes a feature alignment module, a semantic enhancement module, a temporal correlation module, a feature fusion module, an inter-layer feature update module, and a confidence feedback module.

[0048] The mid-level feature fusion module acts as a transition between coarse and fine-level features in the entire feature hierarchy system, while the feature alignment module accepts coarse features. With instant features Through nonlinear transformation Perform space calibration to ensure And strictly require that the IoU of the feature mask must be To eliminate spatial misalignment; subsequently, the semantic enhancement module introduces a channel-level self-attention mechanism. The module enhances key semantic representation, while the temporal correlation module calculates feature differences. Only retain the L1 norm The correlation maintains time stability;

[0049] The preprocessed features enter the feature fusion module, which employs a gated fusion mechanism. Generate mid-layer fusion features Its expression is ,in, The Sigmoid activation function is used. The weight matrix of the gated branch, For time-series correlation features, For the bias term of the gated branch, This is element-wise multiplication. The hyperbolic tangent activation function is used. The weight matrix for the auxiliary branches, As an auxiliary feature, The bias term for auxiliary branches;

[0050] The fusion features are fed into the confidence feedback module for quality assessment, and the confidence level is calculated. And set the confidence threshold to 0.80, where, for Gradient magnitude, The entropy value of the semantic segmentation graph. The average crossover ratio;

[0051] Finally, the inter-layer feature update module dynamically adjusts the incremental update rate based on the value of C. Incremental updates ,in, This is a characteristic of the old middle layer. Incremental update rate For the updated mid-level features;

[0052] The fine-layer feature refinement module serves as the core of high-precision output. The fine-layer feature refinement module includes a local region reconstruction module, a feature residual correction module, a semantic consistency optimization module, a stability evaluation module, a feature growth scheduling module, and an optimization feedback module.

[0053] The local region reconstruction module reconstructs and enhances local spatial information for high-frequency missing or blurred areas in the output of the mid-level feature fusion module; the feature residual correction module calculates and applies feature residuals to correct the current features to minimize deviation; and the semantic consistency optimization module introduces semantic constraints to ensure that the refined features are highly consistent with the high-level semantics in terms of category boundaries and internal representation.

[0054] Subsequently, the feature enters the critical quality control closed loop, where the reliability and stability across time steps of the feature are quantitatively evaluated by the stability assessment module, and the evaluation results drive the feature growth scheduling module in real time.

[0055] When the features fail to meet the preset stability or accuracy standards, the feature growth scheduling module sends the iteration commands and optimization parameters back to the local region reconstruction module through the optimization feedback module, thereby forming a local iterative optimization loop until the features are stable and meet the standards. Finally, the feature growth scheduling module outputs fine-layer features for the final task.

[0056] Furthermore, the feature growth unit is a memory growth module, which includes a feature receiving module, a short-term feature update module, a feature confidence evaluation module, a feature growth determination module, a migration triggering and consolidation module, a long-term optimization module, and a memory feedback and adjustment module.

[0057] The application process of the feature growth unit is as follows:

[0058] The feature receiving module receives the current mid-layer fusion features from the inter-layer feature update module and sends them to the short-term feature update module for fast, local cache updates.

[0059] Following this, the features enter the crucial evaluation and decision-making stage:

[0060] The feature confidence assessment module comprehensively evaluates the quality of newly received features and their matching degree with existing memories, using a formula. ,in To represent the overall confidence level, S is the feature matching success rate, and A is the cross-scene adaptability. The spatial dispersion of the feature;

[0061] The evaluation results guide the feature growth determination module to decide whether to trigger memory growth based on preset indicators, specifically through a comprehensive growth determination scoring formula. A judgment is made, in which, For growth assessment scoring, For the feature access frequency, when Time triggers growth;

[0062] Once growth is determined, the migration triggering and consolidation module migrates the current features from the short-term cache to the long-term optimization module for deep fusion optimization and integration.

[0063] Finally, the long-term optimization module sends the optimized memory to the memory feedback and adjustment module, which is responsible for feeding back the updated global features to the global feature map layer, forming a closed loop of long-term feature optimization and cyclical updates throughout the process.

[0064] Furthermore, the memory management unit includes a global feature map layer, an evolutionary feature cache layer, an instantaneous feature buffer, a feature evaluation module, and a system update module;

[0065] The instant feature buffer includes a feature receiving module, a feature buffer, a feature evaluation module, a short-term replacement module, a feature update module, a confidence weighting module, and a feature migration trigger.

[0066] The evolutionary feature cache layer includes a feature optimization module, a feature filtering and updating module, a loop closure matching optimization module, a feature stability evaluation module, a feature migration triggering and release module, and a feature association and historical data feedback module;

[0067] The global feature map layer includes a migration receiver, a stable feature set, a global index table, a long-term optimization module, a decay evaluation module, a long-term replacement strategy module, and a feedback output module.

[0068] The application process of the memory management unit is as follows:

[0069] Starting with the instantaneous feature buffer, the feature receiving module loads new feature input features received from the feature evaluation module into the feature buffer. Subsequently, the feature evaluation module immediately evaluates the feature quality and redundancy using the following formula: ,in, The gradient magnitude represents the feature and measures the richness of information. Representatives and caching The correlation is used to suppress redundancy, and short-term replacement modules are based on... Update the buffer; For new features;

[0070] For features deemed to have higher value, after the feature update module performs fine-tuning and the confidence weighting module assigns importance weights, the feature migration trigger determines whether the conditions for long-term memory are met based on the score, thereby triggering the migration of the feature to the next level of cache.

[0071] The triggered features enter the evolutionary feature cache layer, which is responsible for mid-term deep optimization:

[0072] The feature optimization module and the feature filtering and update module perform multiple rounds of iterative optimization on the features; meanwhile, the feature stability evaluation module continuously monitors the volatility of the features in consecutive time steps. : ,in, The total number of feature samples participating in the stability assessment. Let it be a distance function. For the first Features of each time step The mean feature is the feature of N consecutive time steps; the output of the feature stability evaluation module directly drives the feature migration trigger and release module to push the features that have reached the stability standard to the global feature map layer; in addition, the loop closure matching optimization module and the feature association and historical data feedback module ensure consistency in the evolution process;

[0073] Ultimately, the stable features collected through the migration receivers constitute the system's long-term knowledge base—the global feature graph layer;

[0074] The global feature map layer is efficiently maintained and redundancy removed through a decay evaluation module and a long-term replacement strategy module based on value judgment. It uses a global index table to achieve fast querying and continuously refines the stable feature set through a long-term optimization module. Finally, the features that have undergone long-term optimization and indexing are output through a feedback output module to drive the overall improvement of the system update module.

[0075] The loop closure detection method based on hierarchical features and growing memory mechanism includes at least the following steps:

[0076] S1: Data acquisition and time synchronization: Acquire visual images and LiDAR point cloud data during robot operation and synchronize them in time.

[0077] S2: Spatial calibration and data preprocessing, performing spatial calibration, noise filtering and format unification on multi-sensor data;

[0078] S3: Feature extraction and multimodal fusion, extracting geometric and semantic features, and mapping the two types of features to a unified feature space;

[0079] S4: Feature layering, which is divided into coarse, medium and fine features based on scale and semantic level;

[0080] S5: Loop closure candidate generation and matching verification: Quickly search for candidates using coarse-layer features and accurately verify loop closure relationships by combining fine-layer features;

[0081] S6: Feature growth and memory management. Feature growth is triggered after loop closure confirmation. Features are stored in a three-layer memory structure according to their attributes to realize dynamic feature migration and cleanup.

[0082] S7: Pose optimization and adaptive update, using loop closure constraints to correct cumulative errors, and adjusting feature growth rate and screening threshold according to detection confidence;

[0083] S8: Long-term operation and maintenance, periodically integrating highly stable features in the global feature graph layer to construct a long-term knowledge graph.

[0084] Compared with the prior art, the beneficial effects of the present invention are:

[0085] 1. This invention adopts a three-layer feature layering mechanism, which improves the efficiency and accuracy of loop closure detection through progressive matching of coarse-layer fast search, mid-layer transition optimization, and fine-layer accurate verification.

[0086] 2. The growth-based memory mechanism used in this invention enables features to be gradually refined and optimized as the system runs, effectively solving the feature degradation problem of traditional methods in long-term operation;

[0087] 3. The three-layer memory structure and adaptive update strategy of this invention realize the dynamic management and optimization of features, while maintaining the compactness of the memory bank and improving the system's adaptability to environmental changes;

[0088] 4. This invention employs multimodal feature fusion and rigorous calibration and filtering processes to enhance the robustness of the system in complex dynamic environments;

[0089] 5. The present invention employs the construction of a long-term knowledge graph to enable the system to continuously learn, further improving the stability and detection accuracy of long-term operation. Attached Figure Description

[0090] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0091] Figure 1 is a schematic diagram of the system flow of the present invention;

[0092] Figure 2 is a schematic diagram of the data acquisition and preprocessing unit of the present invention;

[0093] Figure 3 is a schematic diagram of the multi-sensor time synchronization architecture of the present invention;

[0094] Figure 4 is a schematic diagram of the integrated structure of feature layering and memory growth of the present invention;

[0095] Figure 5 is a schematic diagram of the memory management unit of the present invention.

[0096] in: Detailed Implementation

[0097] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0098] The proposed setup is applied to a multi-sensor fusion SLAM system to improve loop closure detection accuracy and robustness.

[0099] Example 1:

[0100] Please refer to Figure 1. The loop closure detection system based on hierarchical features and growing memory mechanism includes a data acquisition and preprocessing unit 111, a pose estimation and keyframe generation unit 112, a feature extraction unit 113, a multimodal feature fusion unit 114, a feature filtering unit 121, a feature growth unit 122, a candidate generation unit 123, a matching verification unit 124, a pose optimization unit 125, an adaptive update unit 126, and a memory management unit 127.

[0101] The data acquisition and preprocessing unit 111 is used to acquire data from multiple sensors and perform time synchronization, spatial calibration, noise filtering, and format unification.

[0102] The pose estimation and keyframe generation unit 112 estimates the robot pose in real time based on standardized multi-sensor data. By setting motion distance or angle thresholds, it filters keyframes, reduces redundant data, focuses on core effective information, and provides accurate data support for subsequent feature extraction and loop closure detection.

[0103] The feature extraction unit 113 is used to extract geometric features and semantic features. The multimodal feature fusion unit 114 maps the two types of features to a unified feature space and divides them into coarse, medium and fine features according to scale and semantic level to construct a hierarchical feature system. The multimodal feature fusion unit 114 includes a coarse feature extraction module, a medium feature fusion module and a fine feature refinement module. The coarse feature extraction module is used to extract coarse features, the medium feature fusion module is used to extract medium features, and the fine feature refinement module is used to extract fine features.

[0104] The feature filtering unit 121 dynamically removes redundant features based on feature confidence (≥0.75) and access frequency (≥10 times / 100 frames), keeping the feature set compact and effective, avoiding invalid features from occupying resources, and improving the efficiency of subsequent matching and detection.

[0105] The feature growth unit 122 is used to gradually evolve the features from coarse to fine after loop closure confirmation;

[0106] Candidate generation unit 123 uses coarse-layer features to quickly retrieve potential loop closure candidates, narrows the matching range through efficient preliminary screening, reduces the computational load for subsequent accurate verification, and ensures the real-time performance of loop closure detection.

[0107] The matching verification unit 124 combines fine-layer features to accurately verify candidate loop closures. During the system initialization phase, the fine-layer feature refinement module of the multimodal feature fusion unit 114 directly generates initial fine-layer features and stores them in the memory management unit 127. In subsequent operation, the feature growth unit 124 is only responsible for the fusion and update of the existing fine-layer features by semantic channel expansion and geometric key point completion. The matching verification unit 124 confirms the loop closure relationship by calculating the fine-layer feature similarity threshold ≥0.85 corresponding to the candidate loop closure, filtering out false matches, and ensuring the accuracy of loop closure detection.

[0108] The pose optimization unit 125 is used to incorporate the confirmed loop closure constraints into the graph optimization framework, correct the cumulative drift error generated during robot operation, maintain global map consistency, and improve the localization and mapping accuracy of the SLAM system.

[0109] The adaptive update unit 126 dynamically adjusts the feature growth rate and screening threshold according to the detection confidence and environmental changes;

[0110] The memory management unit 127 constructs a three-layer structure consisting of an immediate feature buffer layer, an evolutionary feature cache layer, and a global feature map layer, thereby realizing hierarchical storage and dynamic management of features.

[0111] The memory management unit 127, the feature growth unit 122, and the multimodal feature fusion unit 114 constitute an integrated structure of feature layering and memory growth. The integrated structure of feature layering and memory growth takes the progressive optimization of layered features and the dynamic growth and evolution of memory as its core logic. Through the deep collaboration of the three feature layers and memory management, it solves the problems of feature degradation and insufficient long-term robustness in traditional loop closure detection, and forms a complete closed loop of feature evolution, memory storage and feedback optimization.

[0112] The system of this invention uses data processing, feature evolution, loop closure detection, and optimization feedback as its core logic. Through the collaborative work of each unit, it achieves high-precision and robust loop closure detection, effectively solving problems such as feature degradation and insufficient long-term operational stability in traditional methods. Its specific implementation process is as follows:

[0113] First, the data acquisition, preprocessing, and pose estimation keyframe generation are completed. The data acquisition and preprocessing unit 111 follows the process of sensor initialization, time synchronization, spatial calibration, noise filtering, data cleaning, and feature preparation, outputting multi-source data that is time-consistent, spatially unified, and highly pure, providing reliable input for subsequent steps. The pose estimation and keyframe generation unit 112 estimates the robot's pose in real time based on standardized data, and filters keyframes by setting motion distance or angle thresholds to reduce redundant data and focus on core effective information. The feature extraction unit 113 extracts geometric features (such as point cloud corners and planar points) and semantic features (such as image texture and edges) from the keyframes. The multimodal feature fusion unit 114 maps the two types of features to a unified feature space, dividing them into coarse, medium, and fine features according to scale and semantic level, and constructing a hierarchical feature system.

[0114] Secondly, feature selection and growth, memory management, and candidate matching verification are performed. The feature selection unit 121 dynamically removes redundant features based on feature confidence (≥0.75) and access frequency (≥10 times / 100 frames) to keep the feature set compact and effective. The feature growth unit 122 starts after the initial loop closure candidate is confirmed. Through semantic channel expansion and geometric key point completion, the coarse features are gradually refined into fine features to solve the feature degradation problem in long-term operation. The memory management unit 127 constructs a three-layer structure of immediate feature buffer layer, evolutionary feature buffer layer, and global feature map layer to realize the temporary storage, medium-term consolidation and long-term accumulation of features. The memory bank is kept dynamically balanced through feature migration and decay elimination mechanisms. The candidate generation unit 123 uses coarse features to quickly retrieve potential loop closure candidates. The matching verification unit 124 combines fine features for accurate verification and confirms the loop closure relationship by calculating feature similarity (threshold ≥0.85).

[0115] Finally, pose optimization and system adaptive update are completed. The pose optimization unit 125 incorporates the confirmed loop closure constraints into the graph optimization framework, corrects the cumulative drift error, and maintains global map consistency. The adaptive update unit 126, as the core of system feedback, dynamically adjusts the feature growth rate (0.1 when confidence ≥ 0.9, 0.3 when confidence < 0.9) and the screening threshold based on the loop closure detection confidence and environmental changes, while triggering memory cleanup and feature transfer. It periodically integrates highly stable features (confidence ≥ 0.95, cross-scene adaptability ≥ 0.9) in the global feature graph layer to construct a long-term knowledge graph. Through the closed-loop mechanism of detection, optimization, and feedback, the system's long-term operating capability and loop closure detection accuracy are continuously improved.

[0116] Referring to Figure 2, the data acquisition and preprocessing unit 111 adopts a progressive logic from data acquisition synchronization to spatial calibration and purification to standardization and feature preparation. Through multi-module collaboration, it achieves temporal unification, spatial alignment, noise purification, and feature adaptation of multi-sensor data, providing a high-quality data source for subsequent loop closure detection.

[0117] The data acquisition and preprocessing unit 111 includes a sensor initialization module 211, a timestamp acquisition module 212, a time synchronization module 213, a data acquisition module 221, a spatial calibration module 222, a sensor alignment module 223, a noise filtering module 224, an outlier detection module 225, a data cleaning module 226, a data standardization module 231, and a feature preparation module 232.

[0118] The application process of the data acquisition and preprocessing unit 111 is as follows:

[0119] First, complete sensor initialization, data acquisition, and time synchronization:

[0120] The data acquisition module 221 is linked with the sensor initialization module 211 to configure parameters for the camera, lidar and IMU. The timestamp acquisition module 212 marks the timestamp based on the UTC time of RTK-GNSS. The time synchronization module 213 achieves nanosecond-level synchronization through the GNSS+PTP protocol to form a multi-sensor time synchronization architecture and ensures that the time error is ≤1ms.

[0121] Next, spatial calibration, data cleansing, and alignment will be carried out:

[0122] The spatial calibration module 222 uses the checkerboard method (100mm×100mm), dynamic trajectory optimization, and the Kalibr toolkit to complete the calibration of multi-sensor extrinsic parameters;

[0123] Sensor alignment module 223 press Map the data to the world coordinate system to ensure a spatial overlap rate of ≥95%;

[0124] Noise filtering module 224 is used for targeted filtering (point cloud Gaussian filtering, image median filtering, IMU sliding window filtering);

[0125] The outlier detection module 225 removes outlier data based on the 3σ principle, and the data cleaning module 226 interpolates to supplement a small number of outliers, removes a large number of outliers, and records logs.

[0126] Finally, data standardization and feature preparation are performed:

[0127] Data standardization module 231 Map the features to the range [0,1]. These are the standardized eigenvalues. The minimum value in the original feature data. The maximum value in the original feature data;

[0128] The feature preparation module 232 extracts point cloud geometric features, image visual features, and IMU motion features, classifies and stores them in a buffer, laying the foundation for multimodal fusion.

[0129] Referring to Figure 3, the multi-sensor time synchronization architecture is based on GNSS reference, PTP protocol and time delay correction. The multi-sensor time synchronization architecture achieves high-precision time synchronization of IMU, camera and lidar through multi-level collaboration, ensuring the consistency of timestamps of multi-source data and providing time reference support for subsequent data fusion and processing.

[0130] The time synchronization module 213 has multiple sub-modules, including an RTK-GNSS module 31, a PTP time server 32, an MCU module 33, an execution module 34, a time synchronization / delay correction module 35, and a unified timestamp output module 36.

[0131] The RTK-GNSS module 31 is used to provide an absolute time reference and is the data source support for time synchronization;

[0132] PTP time server 32 and MCU module 33 are the core execution components of time synchronization module 213, realizing the generation of time synchronization signals and synchronous triggering;

[0133] The time synchronization / delay correction module 35 is an error correction submodule of the time synchronization module 213, which ensures synchronization accuracy through Kalman filtering;

[0134] Execution module 34 includes camera 341, lidar 342 and IMU 343;

[0135] The unified timestamp output module 36 connects time synchronization and subsequent data processing, serving as the final step in data preprocessing and ensuring the consistency of output data time.

[0136] The application process of time synchronization module 213 is as follows:

[0137] The RTK-GNSS module 31 receives the high-precision UTC time signal from the global navigation satellite system, which serves as the absolute time reference for the entire synchronization architecture, and outputs the time reference to the PTP time server 32.

[0138] PTP time server 32 generates standardized timing signals based on this reference and transmits them to MCU module 33;

[0139] The MCU module 33 acts as the control center, converting the timing signal into a synchronous trigger command, which is then sent to the camera 341, the lidar 342, and the IMU 343 to drive each sensor to collect data at a uniform time rhythm.

[0140] Then, the time synchronization / delay correction module 35 monitors the timestamp deviation of each sensor data in real time, and uses the Kalman filter algorithm to dynamically correct the device clock drift and data transmission delay, ensuring that the timestamp consistency error of the camera 341, lidar 342 and IMU 343 is ≤1ms.

[0141] After time delay correction, the raw data from each sensor are uniformly transmitted to the unified timestamp output module 36, where standardized timestamps are added and the data is sorted and organized according to sensor type. Finally, a multi-source dataset with consistent time synchronization is output to meet the time consistency requirements of subsequent spatial calibration, data fusion and other processes.

[0142] The hierarchical feature system achieves precise feature optimization through a three-level progression of coarse, medium, and fine layers.

[0143] The coarse feature extraction module includes a data pre-filtering submodule 4311, a global feature encoding submodule 4312, a spatial relationship modeling submodule 4313, a coarse feature fusion submodule 4314, and a coarse feature output module 4315.

[0144] The coarse feature extraction module first uses the feature alignment module 4321 to suppress point cloud noise with adaptive Gaussian filtering (kernel width 0.02-0.08m), then the global feature encoding submodule 4312 generates a 512-dimensional global feature vector, the coarse feature fusion submodule 4314 dynamically weights the features according to the scene (dynamic scene geometric weight 0.6, static scene semantic weight 0.7), and finally the coarse feature output module 4315 normalizes the output.

[0145] The mid-layer feature fusion module includes a feature alignment module 4321, a semantic enhancement module 4322, a temporal correlation module 4323, a feature fusion module 4324, an inter-layer feature update module 4325, and a confidence feedback module 4326;

[0146] The mid-level feature fusion module acts as a transition between coarse-level features (high robustness) and fine-level features (high precision) in the entire feature hierarchy system, while the feature alignment module 4321 accepts coarse-level features. With instant features Through nonlinear transformation Perform space calibration to ensure And strictly require that the IoU of the feature mask must be To eliminate spatial misalignment; subsequently, the semantic enhancement module 4322 introduces a channel-level self-attention mechanism. The key semantic expression is enhanced, while the temporal correlation module 4323 calculates feature differences. Only retain the L1 norm The correlation maintains time stability;

[0147] The preprocessed features enter the feature fusion module 4324, where a gated fusion mechanism is employed. Generate mid-layer fusion features Its expression is ,in, The Sigmoid activation function is used. The weight matrix of the gated branch, For time-series correlation features, For the bias term of the gated branch, This is element-wise multiplication. The hyperbolic tangent activation function is used. The weight matrix for the auxiliary branches, As an auxiliary feature, The bias term for auxiliary branches;

[0148] The fused features are fed into the confidence feedback module 4326 for quality assessment, and the confidence level is calculated. And set the confidence threshold to 0.80, where, for Gradient magnitude, The entropy value of the semantic segmentation graph. The average crossover ratio;

[0149] Finally, the inter-layer feature update module 4325 dynamically adjusts the incremental update rate based on the value of C. Incremental updates ,in, This is a characteristic of the old middle layer. Incremental update rate For the updated mid-level features;

[0150] This enables efficient high-precision integration and dynamic adaptive optimization of mid-level features across spatial, semantic, and temporal dimensions, laying a high-quality feature foundation for subsequent refinement layer optimization.

[0151] As the core of high-precision output, the fine-layer feature refinement module includes a local region reconstruction module 4331, a feature residual correction module 4332, a semantic consistency optimization module 4333, a stability evaluation module 4334, a feature growth scheduling module 4335, and an optimization feedback module 4336.

[0152] The local region reconstruction module 4331 reconstructs and enhances the local spatial information of high-frequency missing or blurred areas in the output of the mid-level feature fusion module; the feature residual correction module 4332 calculates and applies the feature residual to correct the current feature to minimize the deviation; and the semantic consistency optimization module 4333 introduces semantic constraints to ensure that the refined features are highly consistent with the high-level semantics in terms of category boundaries and internal representation.

[0153] Subsequently, the feature enters the critical quality control closed loop, where the reliability and stability across time steps of the feature are quantitatively evaluated by the stability evaluation module 4334, and the evaluation results drive the feature growth scheduling module 4335 in real time.

[0154] When the features fail to meet the preset stability or accuracy standards, the feature growth scheduling module 4335 sends the iteration command and optimization parameters back to the local region reconstruction module 4331 through the optimization feedback module 4336, thereby forming a local iterative optimization loop until the features are stable and meet the standards. Finally, the feature growth scheduling module 4335 outputs high-precision and high-robust fine-layer features for the final task.

[0155] Referring to Figure 4, the feature growth unit 122 is the memory growth module. The feature growth unit 122 includes a feature receiving module 421, a short-term feature update module 422, a feature confidence assessment module 423, a feature growth determination module 424, a migration triggering and consolidation module 425, a long-term optimization module 426, and a memory feedback and adjustment module 427.

[0156] The application process of feature growth unit 122 is as follows:

[0157] The feature receiving module 421 receives the current mid-layer fusion features from the inter-layer feature update module 4325 and sends them to the short-term feature update module 422 for fast, local cache update.

[0158] Following this, the features enter the crucial evaluation and decision-making stage:

[0159] The feature confidence assessment module 423 comprehensively evaluates the quality of newly received features and the matching degree between existing memories, using a formula. ,in To represent the overall confidence level, S is the feature matching success rate, and A is the cross-scene adaptability. The spatial dispersion of the feature;

[0160] The evaluation results guide the feature growth determination module 424 to decide whether to trigger memory growth based on preset indicators, specifically through the growth determination comprehensive scoring formula. A judgment is made, in which, For growth assessment scoring, For the feature access frequency, when Time triggers growth;

[0161] Once growth is determined, the migration triggering and consolidation module 425 migrates the current feature from the short-term cache to the long-term optimization module 426 for deep fusion optimization and integration.

[0162] Finally, the long-term optimization module 426 sends the optimized memory to the memory feedback and adjustment module 427, which is responsible for feeding back the updated global features to the global feature map layer. In the whole process, a closed loop of long-term feature optimization and cyclical update is formed, realizing the continuous growth of memory and the systematic consolidation of knowledge.

[0163] The memory management unit 127 includes a global feature map layer 411, an evolutionary feature cache layer 412, an instant feature buffer 413, a feature evaluation module 51, and a system update module 55;

[0164] Referring to Figure 5, the memory management unit 127 is the core hub for the feature learning system to achieve multi-level, dynamic memory maintenance;

[0165] The instant feature buffer 413 includes a feature receiving module 521, a feature buffer 522, a feature evaluation module 523, a short-term replacement module 524, a feature update module 525, a confidence weighting module 526, and a feature migration trigger 527;

[0166] The evolutionary feature cache layer 412 includes a feature optimization module 531, a feature filtering and updating module 532, a loop closure matching optimization module 533, a feature stability evaluation module 534, a feature migration triggering and release module 535, and a feature association and historical data feedback module 536;

[0167] The global feature map layer 411 includes a migration receiver 541, a stable feature set 542, a global index table 543, a long-term optimization module 544, a decay evaluation module 545, a long-term replacement strategy module 546, and a feedback output module 547.

[0168] The memory management unit 127 is the core of knowledge accumulation and long-term optimization in the entire feature layering process. Its function is to manage and update the global feature map layer in the memory management unit, and ensure the efficiency and stability of the system's long-term features through a series of continuous evaluation and decision-making steps.

[0169] The application process of memory management unit 127 is as follows:

[0170] Starting with the instant feature buffer 413, the feature receiving module 521 loads new feature input features received from the feature evaluation module 51 into the feature buffer 522. Subsequently, the feature evaluation module 523 immediately evaluates the feature quality and redundancy using the following formula: ,in, The gradient magnitude represents the feature and measures the richness of information. Representatives and caching The correlation is used to suppress redundancy, and the short-term replacement module 524 is based on... Update the buffer; For new features;

[0171] For features deemed to have higher value, after the feature update module 525 performs fine-tuning and the confidence weighting module 526 assigns importance weights, the feature migration trigger 527 determines whether the conditions for long-term memory are met based on the score, thereby triggering the migration of the feature to the next level of cache.

[0172] The triggered features enter the evolutionary feature cache layer 412, which is responsible for mid-term depth optimization:

[0173] The feature optimization module 531 and the feature filtering and updating module 532 perform multiple rounds of iterative optimization on the features to improve their robustness; at the same time, the feature stability evaluation module 534 continuously monitors the volatility of the features in consecutive time steps. : ,in, The total number of feature samples participating in the stability assessment. Let it be a distance function. For the first Features of each time step The mean feature of N consecutive time steps is used; the output of the feature stability evaluation module 534 directly drives the feature migration triggering and release module 535 to push the features that have reached the stability standard to the global feature map layer 411; in addition, the loop matching optimization module 533 and the feature association and historical data feedback module 536 ensure consistency in the evolution process;

[0174] Ultimately, the stable features collected by the migration receiver 541 constitute the system's long-term knowledge base—the global feature graph layer 411;

[0175] The global feature map layer 411 is efficiently maintained and redundancy removed through the decay evaluation module 545 and the long-term replacement strategy module 546 based on value judgment. It uses the global index table 543 to achieve fast querying and continuously improves the stable feature set 542 through the long-term optimization module 544. Finally, the features that have been optimized and indexed through the long-term optimization are output through the feedback output module 547 to drive the overall improvement of the system update module 55.

[0176] Example 2:

[0177] Based on the above embodiment 1, a loop closure detection method based on hierarchical features and a growing memory mechanism is proposed, which includes at least the following steps:

[0178] S1: Data acquisition and time synchronization: Acquire visual images and LiDAR point cloud data during robot operation and synchronize them in time.

[0179] S2: Spatial calibration and data preprocessing, performing spatial calibration, noise filtering and format unification on multi-sensor data;

[0180] S3: Feature extraction and multimodal fusion, extracting geometric and semantic features, and mapping the two types of features to a unified feature space;

[0181] S4: Feature layering, which is divided into coarse, medium and fine features based on scale and semantic level;

[0182] S5: Loop closure candidate generation and matching verification: Quickly search for candidates using coarse-layer features and accurately verify loop closure relationships by combining fine-layer features;

[0183] S6: Feature growth and memory management. Feature growth is triggered after loop closure confirmation. Features are stored in a three-layer memory structure according to their attributes to realize dynamic feature migration and cleanup.

[0184] S7: Pose optimization and adaptive update, using loop closure constraints to correct cumulative errors, and adjusting feature growth rate and screening threshold according to detection confidence;

[0185] S8: Long-term operation and maintenance, periodically integrating highly stable features in the global feature graph layer to construct a long-term knowledge graph.

[0186] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A loop closure detection system based on hierarchical features and a growing memory mechanism, characterized in that: The system includes a data acquisition and preprocessing unit (111), a pose estimation and keyframe generation unit (112), a feature extraction unit (113), a multimodal feature fusion unit (114), a feature selection unit (121), a feature growth unit (122), a candidate generation unit (123), a matching verification unit (124), a pose optimization unit (125), an adaptive update unit (126), and a memory management unit (127). The data acquisition and preprocessing unit (111) is used to acquire multi-sensor data and complete time synchronization, spatial calibration, noise filtering, and format unification. The pose estimation and keyframe generation unit (112) is based on the standardized multi-sensor data. Based on real-time estimation of robot pose, keyframes are filtered by setting motion distance or angle thresholds to reduce redundant data and focus on core effective information, providing accurate data support for subsequent feature extraction and loop closure detection; the feature extraction unit (113) is used to extract geometric and semantic features, and the multimodal feature fusion unit (114) maps the two types of features to a unified feature space, dividing them into coarse, medium, and fine features according to scale and semantic level, and constructing a hierarchical feature system. The multimodal feature fusion unit (114) includes a coarse feature extraction module, a medium feature fusion module, and a fine feature refinement module. The coarse feature extraction module is used to extract coarse features, and the medium feature fusion module is used to extract coarse features. The block is used to extract mid-level features, and the fine-level feature refinement module is used to extract fine-level features; the feature filtering unit (121) dynamically removes redundant features based on feature confidence and access frequency, keeping the feature set compact and effective, avoiding invalid features from occupying resources, and improving the efficiency of subsequent matching and detection; the feature growth unit (122) is used to gradually evolve features from coarse to fine after loop closure confirmation; the candidate generation unit (123) uses coarse-level features to quickly retrieve potential loop closure candidates, narrows the matching range through efficient preliminary screening, reduces the computational load for subsequent accurate verification, and ensures the real-time performance of loop closure detection; the matching verification unit (124) combines fine-level features to accurately verify candidate loop closures. In the system initialization phase, the fine-layer feature refinement module of the multimodal feature fusion unit (114) directly generates initial fine-layer features and stores them in the memory management unit (127). In subsequent operation, the feature growth unit (124) is only responsible for the fusion update of the existing fine-layer features by semantic channel expansion and geometric key point completion. The matching verification unit (124) confirms the loop relationship by calculating the fine-layer feature similarity corresponding to the candidate loop, filters false matches, and ensures the accuracy of loop detection. The pose optimization unit (125) is used to incorporate the confirmed loop constraints into the graph optimization framework, correct the cumulative drift error generated during robot operation, maintain global map consistency, and improve the localization and mapping accuracy of the SLAM system. The adaptive update unit (126) dynamically adjusts the feature growth rate and screening threshold according to the detection confidence and environmental changes.The memory management unit (127) constructs a three-layer structure consisting of an immediate feature buffer layer, an evolutionary feature buffer layer, and a global feature map layer, realizing hierarchical storage and dynamic management of features; the memory management unit (127), the feature growth unit (122), and the multimodal feature fusion unit (114) constitute an integrated structure of feature hierarchical storage and memory growth.

2. The loop closure detection system based on hierarchical features and growing memory mechanism according to claim 1, characterized in that: The data acquisition and preprocessing unit (111) includes a sensor initialization module (211), a timestamp acquisition module (212), a time synchronization module (213), a data acquisition module (221), a spatial calibration module (222), a sensor alignment module (223), a noise filtering module (224), an outlier detection module (225), a data cleaning module (226), a data standardization module (231), and a feature preparation module (232).

3. The loop closure detection system based on hierarchical features and growing memory mechanism according to claim 2, characterized in that: The application process of the data acquisition and preprocessing unit (111) is as follows: First, sensor initialization, data acquisition and time synchronization are completed: The data acquisition module (221) links with the sensor initialization module (211) to configure parameters for the camera, lidar and IMU. The timestamp acquisition module (212) marks the data based on the UTC time of RTK-GNSS. The time synchronization module (213) achieves nanosecond-level synchronization through the GNSS+PTP protocol to form a multi-sensor time synchronization architecture, ensuring that the time error is ≤1ms. Then, spatial calibration, data purification and alignment are carried out: The spatial calibration module (222) uses the checkerboard method, dynamic trajectory optimization and Kalibr toolkit to complete the calibration of the external parameters of the multi-sensor. The sensor alignment module (223) follows the... The data is mapped to the world coordinate system to ensure a spatial overlap rate of ≥95%; the noise filtering module (224) is used for targeted filtering; the outlier detection module (225) removes outlier data based on the 3σ principle; the data cleaning module (226) interpolates to supplement a small number of outliers, removes a large number of outliers and logs them; finally, data standardization and feature preparation are performed: the data standardization module (231) according to... Map the features to the range [0,1]. These are the standardized eigenvalues. The minimum value in the original feature data. The maximum value in the original feature data; the feature preparation module (232) extracts point cloud geometric features, image visual features, IMU motion features, classifies and stores them in the buffer, laying the foundation for multimodal fusion.

4. The loop closure detection system based on hierarchical features and growing memory mechanism according to claim 3, characterized in that: The time synchronization module (213) is equipped with multiple sub-modules, including an RTK-GNSS module (31), a PTP time server (32), an MCU module (33), an execution module (34), a time synchronization / delay correction module (35), and a unified timestamp output module (36). The RTK-GNSS module (31) is used to provide an absolute time reference and is the data source support for time synchronization. The PTP time server (32) and the MCU module (33) are the core execution components of the time synchronization module (213) to realize the generation of timing signals and synchronous triggering. The time synchronization / delay correction module (35) is the error correction sub-module of the time synchronization module (213) and ensures synchronization accuracy through Kalman filtering. The execution module (34) includes a camera (341), a lidar (342), and an IMU (343). The unified timestamp output module (36) connects time synchronization and subsequent data processing, and is the final link of data preprocessing to ensure the time consistency of output data.

5. The loop closure detection system based on hierarchical features and growing memory mechanism according to claim 4, characterized in that: The application process of the time synchronization module (213) is as follows: the RTK-GNSS module (31) receives the high-precision UTC time signal of the global navigation satellite system as the absolute time reference of the entire synchronization architecture and outputs the time reference to the PTP time server (32); the PTP time server (32) generates a standardized time signal based on the reference and transmits it to the MCU module (33); the MCU module (33) acts as the control center, converts the time signal into a synchronization trigger command, and sends it to the camera (341), the lidar (342), and the IMU (343) respectively. The system drives each sensor to collect data at a uniform time rhythm. Then, the time synchronization / delay correction module (35) monitors the timestamp deviation of each sensor data in real time and uses the Kalman filter algorithm to dynamically correct the device clock drift and data transmission delay, ensuring that the timestamp consistency error of the camera (341), lidar (342) and IMU (343) is ≤1ms. The original data of each sensor after delay correction is uniformly transmitted to the unified timestamp output module (36), marked with a standardized timestamp and classified and organized according to sensor type, and finally outputs a multi-source dataset with consistent time synchronization.

6. The loop closure detection system based on hierarchical features and growing memory mechanism according to claim 1, characterized in that: The hierarchical feature system achieves precise feature optimization through a three-level progression of coarse, medium, and fine layers. The coarse-layer feature extraction module includes a data pre-filtering submodule (4311), a global feature encoding submodule (4312), a spatial relationship modeling submodule (4313), a coarse feature fusion submodule (4314), and a coarse-layer feature output module (4315). The coarse-layer feature extraction module first uses an adaptive Gaussian filter to suppress point cloud noise through a feature alignment module (4321), and then generates a 512-dimensional global feature vector through the global feature encoding submodule (4312). The feature fusion submodule (4314) dynamically weights features according to the scenario, and finally normalizes the output by the coarse feature output module (4315). The mid-level feature fusion module includes a feature alignment module (4321), a semantic enhancement module (4322), a temporal correlation module (4323), a feature fusion module (4324), an inter-layer feature update module (4325), and a confidence feedback module (4326). The mid-level feature fusion module acts as a transition between coarse and fine features in the entire feature hierarchy system. The feature alignment module (4321) accepts coarse features. With instant features Through nonlinear transformation Perform space calibration to ensure And strictly require that the IoU of the feature mask must be To eliminate spatial misalignment; Subsequently, the semantic enhancement module (4322) introduces a channel-level self-attention mechanism. The key semantic expression is enhanced, while the temporal correlation module (4323) calculates feature differences. Only retain the L1 norm The correlation is maintained to ensure temporal stability; the preprocessed features enter the feature fusion module (4324) and adopt a gating fusion mechanism. Generate mid-layer fusion features Its expression is , in, For the Sigmoid activation function, The weight matrix of the gated branch, For time-series correlation features, For the bias term of the gated branch, This is element-wise multiplication. The hyperbolic tangent activation function is used. The weight matrix for the auxiliary branches, As an auxiliary feature, The bias term for the auxiliary branch; the fused features flow into the confidence feedback module (4326) for quality assessment, and the confidence level is calculated. And set the confidence threshold to 0.80, where, for Gradient magnitude, The entropy value of the semantic segmentation graph. To achieve the average intersection-union ratio; finally, the inter-layer feature update module (4325) dynamically adjusts the incremental update rate based on the value of C. Incremental updates ,in, This is a characteristic of the old middle layer. Incremental update rate The updated mid-level features; the fine-level feature refinement module serves as the core of high-precision output, and includes a local region reconstruction module (4331), a feature residual correction module (4332), a semantic consistency optimization module (4333), a stability evaluation module (4334), a feature growth scheduling module (4335), and an optimization feedback module (4336); the local region reconstruction module (4331) reconstructs and enhances local spatial information for high-frequency detail loss or blurred areas in the output of the mid-level feature fusion module; the feature residual correction module (4332) calculates and applies feature residuals to correct the current features to minimize deviation; the semantic consistency optimization module (4333) optimizes the feature growth of the mid-level features by reconstructing and enhancing the mid-level features; the feature residual correction module (4332) calculates and applies feature residuals to correct the current features to minimize deviation; the semantic consistency optimization module (4333) optimizes the feature growth of the mid-level features by reconstructing and enhancing the mid-level features; the feature residual correction module (4332) calculates and applies feature residuals to correct the current features to minimize deviation; the feature growth of the mid-level features is optimized by reconstructing and enhancing the mid-level features ... Semantic constraints are introduced to ensure that the refined features are highly consistent with the high-level semantics in terms of category boundaries and internal representations. Subsequently, the features enter the critical quality control closed loop. The stability evaluation module (4334) quantitatively evaluates the reliability and stability of the features across time steps. The evaluation results drive the feature growth scheduling module (4335) in real time. When the features do not meet the preset stability or accuracy standards, the feature growth scheduling module (4335) sends the iteration command and optimization parameters back to the local region reconstruction module (4331) through the optimization feedback module (4336), thereby forming a local iterative optimization loop until the features are stable and meet the standards. Finally, the feature growth scheduling module (4335) outputs the fine-layer features for the final task.

7. The loop closure detection system based on hierarchical features and growing memory mechanism according to claim 6, characterized in that: The feature growth unit (122) is the memory growth module. The feature growth unit (122) includes a feature receiving module (421), a short-term feature update module (422), a feature confidence evaluation module (423), a feature growth judgment module (424), a migration triggering and consolidation module (425), a long-term optimization module (426), and a memory feedback and adjustment module (427). The application process of the feature growth unit (122) is as follows: the feature receiving module (421) receives the current mid-layer fusion feature from the inter-layer feature update module (4325) and sends it to the short-term feature update module (422) for fast and local cache update. After that, the feature enters the key evaluation and decision-making stage: the feature confidence evaluation module (423) comprehensively evaluates the quality of the newly received feature and the matching degree of the existing memory, and uses the formula ,in To represent the overall confidence level, S is the feature matching success rate, and A is the cross-scene adaptability. The spatial dispersion of the feature; its evaluation result guides the feature growth determination module (424) to decide whether to trigger memory growth based on preset indicators, specifically through the growth determination comprehensive scoring formula. A judgment is made, in which... For growth assessment scoring, For the feature access frequency, when Growth is triggered when it is determined that growth is needed. Once growth is determined, the migration triggering and consolidation module (425) migrates the current feature from the short-term cache to the long-term optimization module (426) for deep fusion optimization and integration. Finally, the long-term optimization module (426) sends the optimized memory to the memory feedback and adjustment module (427), which is responsible for feeding back the updated global features to the global feature map layer. In the whole process, a closed loop of long-term feature optimization and cyclic update is formed.

8. The loop closure detection system and method based on hierarchical features and growing memory mechanism according to claim 7, characterized in that: The memory management unit (127) includes a global feature map layer (411), an evolutionary feature cache layer (412), an instantaneous feature buffer (413), a feature evaluation module (51), and a system update module (55); the instantaneous feature buffer (413) includes a feature receiving module (521), a feature buffer (522), a feature evaluation module (523), a short-term replacement module (524), a feature update module (525), a confidence weighting module (526), ​​and a feature migration trigger (527); the evolutionary feature cache layer (412) includes a feature optimization module (531), a feature filtering and update module (532), a loop closure matching optimization module (533), a feature stability evaluation module (534), and a feature... The feature migration triggering and release module (535) and the feature association and historical data feedback module (536) are included; the global feature map layer (411) includes a migration receiver (541), a stable feature set (542), a global index table (543), a long-term optimization module (544), a decay evaluation module (545), a long-term replacement strategy module (546), and a feedback output module (547); the application process of the memory management unit (127) is as follows: starting from the immediate feature buffer (413), the feature receiving module (521) loads the new feature input feature received from the feature evaluation module (51) into the feature buffer (522), and then the feature evaluation module (523) immediately evaluates the feature quality and redundancy, and the evaluation formula is: ,in, The gradient magnitude represents the feature and measures the richness of information. Representatives and caching The correlation is used to suppress redundancy, and the short-term replacement module (524) is based on... Update the buffer; For new features, after the feature update module (525) performs fine-tuning and the confidence weighting module (526) assigns importance weights, the feature migration trigger (527) determines whether the conditions for long-term memory are met based on the score, thereby triggering the feature to migrate to the next level of cache; the triggered feature enters the evolutionary feature cache layer (412), which is responsible for mid-term deep optimization: the feature optimization module (531) and the feature filtering and update module (532) perform multiple rounds of iterative optimization on the feature to improve its robustness; at the same time, the feature stability evaluation module (534) continuously monitors the volatility of the feature in continuous time steps. ： ,in, The total number of feature samples participating in the stability assessment. Let be the distance function. For the first Features of each time step The mean feature of N consecutive time steps; the output of the feature stability evaluation module (534) directly drives the feature migration trigger and release module (535) to push the features that have reached the stability standard to the global feature map layer (411); in addition, the loop matching optimization module (533) and the feature association and historical data feedback module (536) ensure consistency in the evolution process; finally, the stable features collected by the migration receiver (541) constitute the long-term knowledge base of the system - the global feature map layer (411); the global feature map layer (411) is efficiently maintained and deredundant through the decay evaluation module (545) and the long-term replacement strategy module (546) based on value judgment, uses the global index table (543) to realize fast query, and continuously improves the stable feature set (542) through the long-term optimization module (544). Finally, the features that have been optimized and indexed by the long-term optimization are output through the feedback output module (547) to drive the overall improvement of the system update module (55).

9. A method for loop closure detection based on hierarchical features and growing memory mechanism, used in the loop closure detection system based on hierarchical features and growing memory mechanism as described in any one of claims 1-8, characterized in that: At least the following steps are included: S1: Data Acquisition and Time Synchronization: Acquire visual images and LiDAR point cloud data during robot operation and synchronize them in time. S2: Spatial Calibration and Data Preprocessing: Perform spatial calibration, noise filtering, and format unification on multi-sensor data. S3: Feature Extraction and Multimodal Fusion: Extract geometric and semantic features and map them to a unified feature space. S4: Feature Hierarchy: Divide the features into coarse, medium, and fine layers based on scale and semantic hierarchy. S5: Loop Closure Candidate Generation and Matching Verification: Quickly search for candidates using coarse-layer features and accurately verify loop closure relationships using fine-layer features. S6: Feature Growth and Memory Management: Trigger feature growth upon loop closure confirmation and store features according to attributes into a three-layer memory structure to achieve dynamic feature migration and cleanup. S7: Pose optimization and adaptive update, using loop closure constraints to correct cumulative errors, and adjusting feature growth rate and screening threshold according to detection confidence; S8: Long-term operation and maintenance, periodically integrating highly stable features in the global feature graph layer to construct a long-term knowledge graph.

Images