Robust visual SLAM system based on fuzzy classification and differential deblurring

Abstract

The invention discloses a robust visual SLAM system based on fuzzy classification and differential deblurring, and belongs to the technical field of image processing, multi-sensor fusion and visual SLAM. The system constructs a closed-loop processing framework of fuzzy discrimination-differential deblurring-feature matching enhancement-SLAM integration, and comprises an image and IMU data acquisition module, a fuzzy discrimination module, a differential deblurring module, an improved GMS feature matching module and a visual SLAM core module. The blurring discrimination module combines image gradient features and IMU motion information to realize blurring degree discrimination of the image, and further distinguishes global blurring and local blurring for repairable blurring; the differential deblurring module introduces IMU constraint deblurring or lightweight deblurring processing for different types of blurring so as to meet the SLAM feature extraction requirement; the improved GMS feature matching module improves the matching stability in a fuzzy scene through a multi-scale and adaptive neighborhood mechanism. And the visual SLAM core module fuses the clear image and the deblurred image to realize high-precision positioning and mapping. According to the method, the robustness and real-time performance of visual SLAM in dynamic and fast motion scenes are effectively improved, and the method is suitable for application scenes such as robot navigation, automatic driving, unmanned aerial vehicles and AR / VR.

Description

Technical Field

[0001] This invention belongs to the fields of image processing, multi-sensor fusion and visual SLAM technology. Specifically, it relates to a robust visual SLAM system based on fuzzy classification and differential deblurring, which is particularly suitable for robots, autonomous driving, drones and other devices that need to deal with motion fuzz and interference from dynamic objects in dynamic or fast-moving scenarios to achieve high-precision positioning and map building. Background Technology

[0002] Visual Simultaneous Localization and Mapping (VisualSLAM) is a core supporting technology in fields such as autonomous driving, robot autonomous navigation, AR / VR, and industrial automation. Its core function is to acquire environmental images through cameras and achieve autonomous device localization and environmental map construction based on image features. However, in real-world applications, rapid camera movement (such as high-speed robot movement, sharp turns of drones, and emergency braking of vehicles) can cause global blurring in the image, while the movement of dynamic objects in the scene (pedestrians, vehicles, obstacles) can cause local blurring. Both types of blurring result in degradation of image edge details and loss of texture information, leading to a series of performance problems in the SLAM system. 1. Feature extraction stage: The edge and texture information of blurred images are damaged, and the FAST corner detection algorithm has difficulty detecting a sufficient number of effective feature points. In some severely blurred areas, feature points are even "blank", resulting in insufficient inter-frame constraints. 2. Feature matching stage: Blurring significantly reduces the discriminative power of ORB descriptors, and the false match rate soars from 3%-5% for clear images to 20%-30%. Traditional matching algorithms (such as brute-force matching and FLANN matching) are unable to effectively eliminate false matches. 3. Pose estimation stage: Insufficient feature points and mismatches will increase the camera pose deviation solved by the PnP algorithm, causing trajectory drift. In severe cases, it will cause the SLAM system to lose tracking and fail to complete localization and mapping. 4. System adaptability: Existing SLAM systems mostly use a single deblurring method to preprocess images, without distinguishing the differences in the causes of global blur and local blur. This results in a lack of motion physical constraints when restoring global blur, and excessive computational overhead when processing local blur, making it difficult to balance restoration quality and real-time performance.

[0003] Existing solutions to the blurring problem in SLAM scenes have the following main shortcomings: Limited accuracy of fuzzy discrimination: Traditional fuzzy discrimination relies solely on the gradient information of the image itself (such as Laplacian variance), without combining IMU motion data. It is easy to misjudge local fuzziness caused by dynamic objects as global fuzziness, or to miss slight global fuzziness, resulting in insufficient targeting of subsequent defuzzification processing. Deblurring methods lack differentiation: Existing methods (such as the basic DeblurGAN-V2 and MPRNet) use a uniform model to handle all types of fuzziness. When restoring global fuzziness, the lack of motion physical constraints makes it easy to produce edge ringing artifacts. When processing local fuzziness, the complexity of the model leads to excessive computational overhead, which cannot meet the real-time requirements of SLAM systems. Insufficient feature matching adaptability: The feature distribution of the deblurred image changes. The traditional GMS algorithm is not optimized for the feature characteristics of the deblurred image. There is still room for improvement in the matching robustness in continuous frames and wide baseline scenarios. It is difficult to make full use of the information of the deblurred image. Low system integration: Existing methods mostly treat deblurring as an independent preprocessing module, without deep integration with the feature extraction, pose estimation, map building and other stages of the SLAM system. The deblurring results are not optimized for the core needs of SLAM, resulting in limited overall performance improvement. Balancing real-time performance and reliability is difficult: While some deep learning deblurring methods can improve restoration quality, they have high computational complexity, with single-frame processing taking more than 50ms, which cannot meet the real-time requirements of SLAM systems above 15Hz; while lightweight methods suffer from insufficient restoration accuracy.

[0004] Therefore, designing an integrated robust SLAM system that can accurately identify fuzzy types, perform targeted differentiated defuzzing, optimize feature matching strategies, and deeply integrate with the SLAM system has become the key to solving the above-mentioned technical bottlenecks. Summary of the Invention

[0005] This invention proposes a robust visual SLAM system based on fuzzy classification and differential deblurring, aiming to solve the problems of insufficient feature extraction, low matching accuracy, positioning drift, and tracking loss caused by global fuzziness (camera motion) and local fuzziness (dynamic objects) in dynamic or fast-moving scenes in traditional visual SLAM systems. It achieves more accurate fuzzy discrimination, deblurring differentiation, robust feature matching, and deeper SLAM integration, thereby improving the positioning accuracy, stability, and real-time performance of SLAM systems in complex fuzzy scenes.

[0006] To achieve the above objectives, this invention provides a robust visual SLAM system based on fuzzy classification and differential deblurring. This system forms a closed-loop end-to-end system from data acquisition, fuzzy discrimination, differential deblurring, feature matching to SLAM integration, with each module working collaboratively. The specific technical solution includes, in sequence, an image and IMU data acquisition module, a fuzzy discriminator module, a differential deblurring module, an improved GMS feature matching module, and a visual SLAM core module. The overall system workflow is as follows: (1) The image and IMU data acquisition module is used to synchronously acquire RGB image sequences and IMU raw data; (2) The blur discriminator module is used to judge the blur degree of RGB images and further judge the repairable blur images as global blur images or local blur images; (3) The differential deblurring module is used to perform corresponding processing on the image based on the fuzzy classification results; (4) Improve the GMS feature matching module for feature extraction and matching of the processed image; (5) The core module of visual SLAM is used to complete pose estimation and map construction based on feature matching results.

[0007] Furthermore, the image and IMU data acquisition modules achieve synchronous acquisition through a hardware triggering mechanism, and achieve time alignment between image data and IMU data through timestamp calibration, with an alignment error of no more than 0.1ms.

[0008] Furthermore, the hardware configuration of the image and IMU data acquisition module includes: an RGB camera (resolution ≥ 1280×720, frame rate ≥ 15Hz), a six-axis IMU (sampling rate 200Hz, supports angular velocity and acceleration acquisition, zero bias stability ≤ 0.5° / h), and a synchronization trigger module (supports dual synchronization of hardware trigger and software timestamp calibration); the preprocessing of the IMU raw data includes sliding window averaging zero bias compensation (window size 50), outlier removal (3σ criterion), and data normalization. The preprocessed data is used as motion constraint input for optical flow field generation and deblurring processing.

[0009] Furthermore, the blur discriminator module analyzes the high-frequency information of the image based on the Laplacian operator. By calculating the variance of the Laplacian response and combining it with an adaptive threshold, the image is classified as a clear image, a moderately or mildly repairable blur image, or a severely unrepairable blur image.

[0010] Furthermore, the blur discriminator module calculates a dense optical flow field for the repairable blurred image and generates a desired optical flow field based on camera motion parameters obtained through IMU pre-integration. By comparing the consistency between the two, the repairable blurred image is classified as either a globally blurred image or a locally blurred image. The specific process includes: 1) Image preprocessing: Convert the RGB image into a single-channel image by weighted grayscale to reduce redundant calculations; 2) Blur degree judgment: The high-frequency components are extracted using a diagonal-sensitive Laplacian operator. The logarithmic transformation is applied to the operator response values to compress the dynamic range, and the variance V of the transformed response values is calculated. Based on the Otsu algorithm, the initial threshold T0 = 4.2 is optimized, and the adaptive threshold T = T0 + 0.1×v is dynamically adjusted in combination with the average optical flow velocity v of the image. According to the relationship between V and T, V1 (severe blur threshold, V1 = T0×0.5), the classification is as follows: V≥T is a clear image, V1≤V<T is a moderately / lightly repairable blurred image, and V<V1 is a severely irreparable blurred image; 3) Blur type discrimination: For repairable blurred images, the LK optical flow algorithm is used to calculate the dense optical flow field F. Based on the camera motion parameters (rotation matrix R, translation vector T) obtained by IMU pre-integration, the expected optical flow field F_IMU of the IMU is generated through the pinhole camera model. The weighted L2 norm difference D (higher weights are assigned to high-gradient regions) and the cosine similarity S are calculated. When D<D0 (D0 = 0.3) and S>S0 (S0 = 0.7), it is determined as a globally blurred image; otherwise, it is a locally blurred image; 4) Discriminator optimization: A sliding window cache (window size 10 frames) is introduced to perform consistency verification on the discrimination results of consecutive frames, avoiding the SLAM tracking fluctuations caused by single-frame misjudgment.

[0011] Furthermore, the differential deblurring module processes images according to the following rules: Clear images are directly input into the core module of visual SLAM; severely irreparable blurred images are directly removed; globally blurred images are deblurred by introducing IMU motion constraints; locally blurred images are deblurred by lightweight deblurring.

[0012] Furthermore, the deblurring of globally blurred images and locally blurred images is both implemented based on the DeblurGAN-V2 network. Among them, IMU motion constraints are introduced in the deblurring of globally blurred images, and a network structure with simplified parameters is adopted in the deblurring of locally blurred images.

[0013] Furthermore, the improved GMS feature matching module extracts ORB features from clear images and deblurred images, and realizes feature matching through multi-scale grid division, dynamic neighborhood adjustment, and clustering consistency verification. The specific implementation is as follows: 1) Feature extraction adaptation: ORB feature extraction is used for clear images and deblurred images. The FAST corner detection threshold is set to 20, the maximum number of feature points is 2000, and the patch size is expanded to 41×41 to improve the feature stability in blurred scenarios; 2) Multi-scene matching strategy: For continuous frame scenes (sharp images / deblurred continuous frames): multi-scale grid partitioning (15×15, 20×20, 25×25) is adopted, multi-scale support is fused (weights 0.3, 0.5, 0.2), bidirectional GMS verification (forward + reverse matching consistency) is added, and one-way mismatches are eliminated; For wide baseline scenes (after removing heavily blurred frames): DBoW2 bag-of-words retrieval (vocabulary size 10,000) is introduced to filter candidate matches, reducing invalid matches by 70%; the neighborhood type (4 neighborhood for φ<30°, 6 neighborhood for 30°≤φ<60°, and 8 neighborhood for φ≥60°) and the number of rotations (3, 5, and 8 times) are dynamically adjusted according to the viewing angle φ, and weak perspective transformation constraints are added for refinement (reprojection error ≤1.5 pixels). 3) Post-processing optimization: DBSCAN clustering verification is adopted (neighborhood radius ε=15 pixels, minimum number of cluster points minPts=5) to eliminate isolated mismatches; temporal smoothing filter (Kalman filter) is introduced to correct the position of matching points in consecutive frames and improve matching consistency. 4) Parameter Adaptation: The support threshold is dynamically adjusted according to the degree of image blur (T=8 for clear images, T=6 for slightly blurred images, and T=4 for moderately blurred images) to adapt to different input image qualities.

[0014] Furthermore, the IMU motion information is obtained through the IMU pre-integration method, used to generate the desired optical flow field and as the motion constraint input for global fuzzing and defuzzing processing.

[0015] Furthermore, the core visual SLAM module is implemented based on the ORB-SLAM3 framework, and completes camera pose estimation and map construction through PnP pose estimation, bundle adjustment optimization, and loop closure detection. The specific process is as follows: 1) Keyframe Management: Set keyframe insertion thresholds (rotation difference ≥ 5°, translation difference ≥ 0.1m or feature matching number ≤ 300) to avoid redundant keyframes; 2) Pose estimation optimization: Based on the inlier matching pairs obtained by the improved GMS, the PnP algorithm is used to initially estimate the pose, and the reprojection error is optimized by Local Bundle Adjustment (LocalBA) (target error ≤ 1.0 pixel). 3) Enhanced loop closure detection: Introducing a feature descriptor library for deblurred images to improve the recall and accuracy of loop closure detection; 4) Map building: A semi-dense map building strategy is adopted, which generates dense map points in areas with dense features and retains key map points in areas with sparse features, thus balancing map accuracy and storage overhead.

[0016] A robust visual SLAM system based on fuzzy classification and differential defuzzification is applied to devices that require high-precision positioning and mapping in dynamic or fast-moving scenarios, such as robot autonomous navigation, autonomous driving, drone inspection, AR / VR positioning, and industrial robot positioning. It is compatible with mobile platforms equipped with RGB cameras and six-axis IMUs.

[0017] The beneficial effects of this invention are: 1. Accurate and efficient fuzzy discrimination: Combining image gradients and IMU motion data, it achieves three-level classification and two-level subdivision with a classification accuracy of ≥94.7% and an F1 score improvement of over 22.9%. It avoids redundant processing of clear images, accidental removal of repairable images, and misjudgment of fuzzy types. The processing time per frame is ≤0.45ms, which fully meets the real-time requirements of SLAM. 2. Highly targeted deblurring: Global blurring combined with IMU motion constraints results in excellent restoration quality (PSNR≥30.75dB, SSIM≥0.9137) with no edge ringing artifacts; Local blurring uses a lightweight network, reducing computation by 60% and processing time per frame ≤15ms, balancing quality and real-time performance. 3. High robustness of feature matching: The improved GMS algorithm adapts to the feature characteristics of deblurred images, increasing the inlier rate by 11.6%-33.1% and reducing matching time by 20%-55.8%, applicable to both continuous frames and wide baseline scenarios; the false match rate is reduced from 20%-30% to 5%-8%. 4. Significantly improved SLAM performance: On the Motionblur1-Motionblur3 sequences of the ground-challenge dataset, the RMSE of SLAM trajectories was reduced by 46.14%, 41.92%, and 51.10%, respectively, and the number of feature matches was increased by 26.59%, 22.04%, and 23.17%, respectively, solving the problems of localization drift and tracking loss caused by blur. 5. Strong engineering practicality: Based on the mature ORB-SLAM3, DeblurGAN-V2 and DBoW2 frameworks, it is easy to integrate with existing robot hardware (such as ROS robots and autonomous vehicles); it supports multiple sensor configurations and scene adaptations, and is suitable for a variety of complex indoor and outdoor environments; 6. Excellent scalability: The deblurring network and feature matching algorithm can be flexibly replaced, and the performance can be further optimized by introducing deep learning feature extraction, semantic segmentation and other technologies. Attached Figure Description

[0018] Figure 1 This is an overall flowchart of a robust visual SLAM system based on fuzzy classification and differential defuzzification according to the present invention. Figure 2 This is a flowchart illustrating the workflow of the fuzzy discriminator of the present invention. Figure 3 The flowchart of the improved GMS algorithm of this invention; Figure 4 This is a comparison chart of the fuzzy classification effects of the present invention; Figure 5 For comparison of the deblurring effect; Figure 6 This is a comparison chart of the feature matching results of this invention on the ground-challenge dataset; Figure 7 This is a trajectory comparison diagram of the present invention on the ground-challenge dataset; Figure 8 This is a graph showing the real-time performance test results of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments 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 embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are intended to be within the scope of the invention and form different embodiments. For example, in the claims, any of the claimed embodiments can be used in any combination.

[0021] The following is a reference to the appendix. Figure 1 To be continued Figure 8 This invention describes a robust visual SLAM system based on fuzzy classification and differential deblurring. The system forms a closed-loop end-to-end from data acquisition, fuzzy discrimination, differential deblurring, feature matching to SLAM integration, with each module working collaboratively. The specific technical solution includes, in sequence, an image and IMU data acquisition module, a fuzzy discriminator module, a differential deblurring module, an improved GMS feature matching module, and a visual SLAM core module. The overall system workflow is as follows: (1) The image and IMU data acquisition module is used to synchronously acquire RGB image sequences and IMU raw data; (2) The blur discriminator module is used to judge the blur degree of RGB images and further judge the repairable blur images as global blur images or local blur images; (3) The differential deblurring module is used to perform corresponding processing on the image based on the fuzzy classification results; (4) Improve the GMS feature matching module for feature extraction and matching of the processed image; (5) The core module of visual SLAM is used to complete pose estimation and map construction based on feature matching results.

[0022] Specifically, the image and IMU data acquisition module hardware uses an RGB camera (1280×720 resolution, 30Hz frame rate, global shutter to avoid distortion caused by rolling shutter) and a six-axis IMU (200Hz sampling rate, angular velocity zero-bias stability ≤0.5° / h, acceleration zero-bias stability ≤1mg), mounted on a mobile platform (robot, vehicle, drone); it adopts hardware trigger synchronization (the camera and IMU are connected through the GPIO interface, and the trigger signal is acquired synchronously) + software timestamp calibration to ensure that the timing of each frame of image is aligned with the corresponding IMU data.

[0023] Specifically, the IMU data preprocessing workflow includes: 1) Zero bias compensation: The zero bias of IMU angular velocity and acceleration is estimated by using the sliding window averaging method (window size 50), and the compensated IMU data is closer to the real motion; 2) Outlier removal: Outliers (data points exceeding the mean ± 3 standard deviations) in the IMU data are removed using the 3σ criterion. 3) Data normalization: Normalize the IMU data to the [-1,1] interval to adapt to the input requirements of the subsequent defuzzification network.

[0024] Specifically, the fuzzy discriminator module adopts a two-level mechanism of "fuzziness degree judgment + fuzziness type judgment", simplifying the formulas throughout and implementing them through logical judgment and numerical thresholds to ensure real-time performance and accuracy. Step 1: Determine the degree of fuzziness 1) Image grayscale conversion: Convert the RGB image into a single-channel image through weighted grayscale conversion (weight coefficients: R=0.299, G=0.587, B=0.114) to reduce redundant calculations; 2) High-frequency component extraction: The diagonal-sensitive Laplacian operator is used, which can capture horizontal, vertical and diagonal edges at the same time, avoiding feature omissions caused by blurred direction, and is especially suitable for the detection needs of key SLAM features such as lane lines and building outlines. 3) Response value optimization: The response value of the Laplacian operator is logarithmically transformed to compress the dynamic range while enhancing the distinguishability of blurred edges and avoiding interference from extreme values; 4) Blur Quantization: Calculate the variance V of the transformed response value. A larger V value indicates richer image edge details and a clearer image; this is combined with an adaptive threshold T for judgment. V≥T: Clear image (directly input into the SLAM system to avoid redundant processing); V1 ≤ V < T: Moderate / Light Repairable Blurred Images (Enter the Type Discrimination Process); V < V1: Severe Unrepairable Blurred Images (Directly Rejected to Avoid Affecting SLAM Performance); Threshold Calibration: The initial threshold T0 = 4.2 (Determined by the Otsu algorithm based on 1000 indoor and outdoor blurred-clear image samples), and T = T0 + 0.1×v is dynamically adjusted in combination with the average optical flow velocity v of the image (The faster the movement, the more relaxed the threshold to avoid misjudgment); V1 is the severe blur threshold, with a value of T0×0.5 = 2.1.

[0025] Second Step: Blur Type Discrimination 1) Dense Optical Flow Field Estimation: The LK optical flow algorithm is used to calculate the pixel motion vector field F of adjacent frames. Based on the assumption of constant brightness, an overdetermined system of equations is constructed and solved within a 3×3 neighborhood to obtain the optical flow vector with sub-pixel accuracy; 2) Generation of IMU Expected Optical Flow Field: Based on the camera motion parameters (rotation matrix R, translation vector T) obtained by IMU pre-integration, the camera motion is mapped to the pixel displacement of the image plane through the pinhole camera model to generate the IMU expected optical flow field F_IMU, which directly reflects the global blur pattern caused by camera motion; 3) Consistency Metric: The weighted L2 norm difference D and cosine similarity S are used as dual metrics to determine the blur type: Weighted L2 Norm Difference D: Higher weights are assigned to high-gradient regions (such as object edges) to highlight the motion consistency in the edge regions; Cosine Similarity S: Measures the direction consistency between the optical flow field F and F_IMU; Decision Rule: When D < D0 (D0 = 0.3) and S > S0 (S0 = 0.7), it is determined as global blur (caused by camera motion); otherwise, it is determined as local blur (caused by dynamic objects).

[0026] Third Step: Discriminator Optimization A sliding window cache (window size 10 frames) is introduced to perform consistency verification on the discrimination results of consecutive frames: If the discrimination result of a certain frame image is inconsistent with the majority of frames within the window, its blur parameters V, D, and S are recalculated to avoid SLAM tracking fluctuations caused by single-frame misjudgment; After optimization, the single-frame processing time of the blur discriminator ≤ 0.45ms, the global blur recognition accuracy reaches 94.7%, and the threshold false detection rate in different scenarios is lower than 5.2%.

[0027] Specifically, to address the differences in the causes of global blur and local blur, a differentiated processing strategy is adopted to ensure that the deblurred image can retain the edge and texture features required by SLAM while meeting real-time requirements. The core design is to introduce IMU motion prior constraints to make the deblurring process conform to the camera motion law and avoid restoration without physical basis.

[0028] To address the characteristic of local blurring in dynamic objects (concentrated blurry areas with clear backgrounds), the network is optimized to balance speed and quality.

[0029] Specifically, the following experiment will verify this. First, data acquisition and preprocessing will be performed, as follows: 1. Start the ROS system, initialize the camera and IMU nodes, and achieve data synchronization through hardware triggering. Subscribe to the ` / camera / rgb / image_raw` topic published by the camera and the ` / imu / data` topic published by the IMU, with the timestamp error controlled within 0.1ms; 2. Preprocess the raw IMU data: use the sliding window averaging method (window size 50) to estimate the zero bias, use the 3σ criterion to remove outliers, and normalize to the [-1,1] interval; 3. Preprocess the dataset images: uniformly scale them to 1280×720 resolution, associate them with IMU data by timestamp, and store them in the format of "image path + IMU data file + timestamp + blur type label".

[0030] Then, the fuzzy discriminator is deployed, specifically including: 1. Load the pre-trained fuzzy discrimination model and set the threshold parameters: T0=4.2, V1=2.1, D0=0.3, S0=0.7, and the sliding window size=10 frames; 2. Perform blur level judgment on each frame of image: calculate the Laplacian response variance V, and classify according to the relationship between V and T, V1; 3. Perform type discrimination on repairable blurred images: Calculate the dense optical flow field F and the expected optical flow field F_IMU of the IMU, and determine the blur type by the consistency measure of D and S; 4. Output classification results: Clear images are labeled "Valid - Clear", globally blurred images are labeled "Valid - Globally Blurry", locally blurred images are labeled "Valid - Locally Blurry", and heavily blurred images are labeled "Invalid - Removed". See the appendix for the results. Figure 4 The comparison chart of fuzzy classification results is shown.

[0031] Third, implement differentiated deblurred network deployment, refer to the appendix. Figure 2 The flowchart of the fuzzy discriminator process includes: 1. Call the corresponding network based on the classification result: Global blurred image: Input IMU to enhance DeblurGAN-V2, synchronously input the desired optical flow field to IMU, and output the deblurred image; Locally blurred image: Input a lightweight DeblurGAN-V2, output a deblurred image; 2. Image post-processing: Contrast enhancement (gamma correction, γ=1.2) is performed on the deblurred image to improve the readability of features in weak texture areas and edges, and to ensure the detection rate and stability of ORB feature extraction; 3. Statistical effective input rate: In the Motionblur1 sequence, the effective input rate is ≥85% to ensure that the SLAM system has sufficient inter-frame constraints.

[0032] Fourth, improve the deployment of the GMS feature matching module, referring to the appendix. Figure 3 As shown, the improved GMS algorithm process specifically includes: 1. Configure ORB feature extraction parameters: FAST corner threshold = 20, maximum number of feature points = 2000, patch size = 41×41; 2. Configure improved GMS parameters: multi-scale grid size (15×15, 20×20, 25×25), DBoW2 bag-of-words dictionary (vocabulary size = 10000), DBSCAN clustering parameters (ε = 15 pixels, minPts = 5); 3. Select matching strategy according to scenario: Multi-scale fusion and bidirectional verification are enabled for continuous frame scenarios, and bag-of-words retrieval and dynamic neighborhood adjustment are enabled for wide baseline scenarios; 4. Output valid matching pairs: Count the number of matching pairs and inlier rate for each frame, and save the matching results for subsequent pose estimation.

[0033] Fifth, deploy and run the SLAM system, which specifically includes: 1. Configure ORB-SLAM3 core parameters: Keyframe insertion thresholds: rotation difference ≥ 5°, translation difference ≥ 0.1m, feature match count ≤ 300; Pose estimation parameters: PnP iteration count = 1000, reprojection error threshold = 1.0 pixel; Loop closure detection parameters: bag-of-words similarity threshold = 0.7, consistency check count = 3; 2. Start the SLAM system, subscribe to the valid image topic (' / deblur / image_raw') and IMU data topic (' / imu / data'), and execute the following process: 1) Feature extraction: Extract ORB features from the input image; 2) Feature matching: Call the improved GMS feature matching module to obtain valid matching pairs; 3) Pose estimation: Preliminary PnP estimation + LocalBA optimization; 4) Keyframe Management: Determine whether to insert a new keyframe; 5) Map building: Generate / update map points, loop closure detection, and GlobalBA optimization; 3. Record output data: real-time camera pose (' / slam / pose'), environment map (' / slam / map'), feature matching results (' / slam / matches'), trajectory file ('trajectory.txt').

[0034] Finally, verification is performed, specifically including: 1. Verification of Qualitative Results Deblurring effect: See attached document Figure 5 As shown, after processing, the details of the building outlines and road sign text in the globally blurred image are clear and without artifacts; after processing, the details of the dynamic object area are fully restored and the background is not overly smoothed. Feature matching results: See appendix Figure 6 As shown, the matching lines are regular, mismatches are significantly reduced, the number of consecutive frame matching pairs is ≥300, and the number of wide baseline matching pairs is ≥200. Track effect: See attached document Figure 7 As shown in the trajectory comparison image on the ground-challenge dataset, the SLAM trajectory closely matches the ground truth trajectory with no obvious drift and no tracking loss.

[0035] 2. Validation of quantitative results Deblurring metrics: On the extended GOPRO test set, the PSNR of IMU-enhanced DeblurGAN-V2 is 30.75 dB, and the SSIM is 0.9137; the PSNR of lightweight DeblurGAN-V2 is 29.86 dB, and the SSIM is 0.8972. Matching metrics: In the Motionblur1 sequence, the number of feature matches increased from 372.13 to 471.07, an improvement of 26.59%, and the inlier rate increased from 62% to 85%. SLAM metrics: On the Motionblur1-Motionblur3 sequence, the trajectory RMSE decreased by 46.14%, 41.92%, and 51.10%, respectively, and the mean absolute error (MAE) decreased by more than 35%. Real-time performance indicators: see appendix Figure 8 As shown in the real-time test results, the overall processing time for a single frame is ≤33ms, corresponding to a frame rate ≥30Hz, which meets the real-time requirements of SLAM.

[0036] 3. Comparative experimental verification To further verify the performance of the complete "deblurring + improved GMS feature matching" scheme, experiments were conducted on the Motionblur1-Motionblur3 sequences in the GroundChallenge dataset for comparison: The experimental results are shown in Table 1. The positioning accuracy of the complete scheme is significantly better than that of the "deblurring only" scheme: In the Motionblur1 sequence, the root mean square error (rmse) of the complete scheme is improved by 64.80% compared with ORB-SLAM3 (and further improved by 34.65% compared with the "deblurring only + ORB-SLAM3" scheme), and the minimum error (min) is reduced by 94.28%; the rmse optimization rate of the Motionblur2 sequence is 64.37% (improved by 38.65% compared with "deblurring only + ORB-SLAM3"); even in the Motionblur3 sequence with the most significant blurring, the rmse of the complete scheme still achieves a significant improvement of 73.00% (improved by 44.79% compared with "deblurring only + ORB-SLAM3").

[0037] Table 1 Comparison of errors between the method of the present invention and ORB-SLAM3 on the GroundChallenge dataset. The above results demonstrate that deblurring provides a clear visual input foundation for SLAM, while improved GMS feature matching further optimizes the matching accuracy and robustness of feature points. The combined approach maximizes the suppression of motion blur interference and significantly improves the positioning accuracy of the SLAM system. Compared to solutions that only introduce deblurring, the complete solution achieves secondary optimization across all error metrics, fully validating the synergistic gain effect of the "deblurring + improved GMS" technology link and proving the practicality and superiority of the proposed solution in complex motion blur scenarios.

[0038] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A robust visual SLAM system based on fuzzy classification and differential defuzzification, characterized in that, The system comprises, in sequence, an image and IMU data acquisition module, a blur discriminator module, a differential deblurring module, an improved GMS feature matching module, and a visual SLAM core module; the system operates as follows: The image and IMU data acquisition module is used to synchronously acquire RGB image sequences and raw IMU data; The blur discriminator module is used to determine the degree of blur in RGB images, and further determine whether repairable blur images are globally blurred or locally blurred. The differential deblurring module is used to perform corresponding processing on the image based on the fuzzy classification results; The improved GMS feature matching module is used for feature extraction and matching of the processed image; The visual SLAM core module is used to complete pose estimation and map construction based on feature matching results.

2. The robust visual SLAM system based on fuzzy classification and differential defuzzification according to claim 1, characterized in that, The image and IMU data acquisition modules achieve synchronous acquisition through a hardware triggering mechanism, and achieve time alignment between image data and IMU data through timestamp calibration, with an alignment error of no more than 0.1ms.

3. The robust visual SLAM system based on fuzzy classification and differential defuzzification according to claim 1, characterized in that, The blur discriminator module analyzes high-frequency information of the image based on the Laplacian operator. By calculating the Laplacian response variance and combining it with an adaptive threshold, the image is classified as a clear image, a moderately or mildly repairable blur image, or a severely unrepairable blur image.

4. The robust visual SLAM system based on fuzzy classification and differential defuzzification according to claim 3, characterized in that, The blur discriminator module calculates a dense optical flow field for the repairable blur image and generates a desired optical flow field based on the camera motion parameters obtained by IMU pre-integration. By comparing the consistency between the two, the repairable blur image is identified as a globally blurred image or a locally blurred image.

5. A robust visual SLAM system based on fuzzy classification and differential defuzzification as described in claim 1, characterized in that, The differential deblurring module processes the image according to the following rules: Clear images are directly input into the visual SLAM core module; severely blurred images that cannot be repaired are directly removed; globally blurred images are deblurred using IMU motion constraints; and locally blurred images are deblurred using lightweight methods.

6. A robust visual SLAM system based on fuzzy classification and differential defuzzification as described in claim 5, characterized in that, The deblurring of both the globally blurred and locally blurred images is implemented based on the DeblurGAN-V2 network. The deblurring of the globally blurred image introduces IMU motion constraints, while the deblurring of the locally blurred image uses a network structure with simplified parameters.

7. A robust visual SLAM system based on fuzzy classification and differential defuzzification as described in claim 1, characterized in that, The improved GMS feature matching module extracts ORB features from clear and deblurred images, and achieves feature matching through multi-scale grid partitioning, dynamic neighborhood adjustment, and cluster consistency verification.

8. A robust visual SLAM system based on fuzzy classification and differential defuzzification as described in claim 4, characterized in that, The IMU motion information is obtained through the IMU pre-integration method and is used to generate the desired optical flow field and as the motion constraint input for global fuzzing and defuzzing processing.

9. A robust visual SLAM system based on fuzzy classification and differential defuzzification as described in claim 1, characterized in that, The core module of visual SLAM is implemented based on the ORB-SLAM3 framework, and completes camera pose estimation and map construction through PnP pose estimation, bundle adjustment optimization and loop closure detection.

10. A robust visual SLAM system based on fuzzy classification and differential defuzzification according to any one of claims 1-9, characterized in that, It is applied to devices that require high-precision positioning and mapping in dynamic or fast-moving scenarios, such as robot autonomous navigation, autonomous driving, drone inspection, AR / VR positioning, and industrial robot positioning. It is compatible with mobile platforms equipped with RGB cameras and six-axis IMUs.

Images