Adaptive dynamic SLAM method based on Gaussian distribution

Abstract

The invention discloses a self-adaptive dynamic SLAM method based on Gaussian distribution, and relates to the technical field of dynamic environments. According to the method, the dynamic environment positioning precision is improved, the system can accurately recognize and eliminate interference of dynamic objects through a dynamic target elimination strategy based on CRF in combination with YOLO detection and conditional random field optimization, and a traditional SLAM system is likely to be affected by moving objects in the dynamic environment to generate pose offset, so that the system is not easy to operate. Reprojection error evaluation and feature point weighting processing are introduced, differential modeling is carried out on static points and dynamic points, the dependence of pose estimation on static features is strengthened, estimation errors caused by dynamic interference are effectively reduced, the positioning precision and stability of the system in a complex dynamic scene are improved, and the positioning accuracy of the system in the complex dynamic scene is improved. Moreover, the consistency and integrity of the map are enhanced, and the mapping efficiency and scene adaptability are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of dynamic environments, and specifically to an adaptive dynamic SLAM method based on Gaussian distribution. Background Art

[0002] Existing SLAM mapping technologies mainly include dense point cloud mapping, voxel / mesh-based mapping methods, and 3D reconstruction schemes based on Gaussian mapping.

[0003] The dense point cloud mapping method is a traditional mainstream mapping solution. This type of method realizes the reconstruction of a 3D scene by performing frame-by-frame registration and accumulation on the RGB-D data or lidar point cloud obtained by the sensor. For example, ElasticFusion uses sparse pose graph optimization and loop closure detection to achieve real-time dense reconstruction, and ORB-SLAM3 combines IMU information to improve mapping accuracy and robustness. However, in a dynamic environment, the dense point cloud is easily contaminated by dynamic objects, resulting in trajectory drift and map distortion.

[0004] The voxel or mesh-based mapping method models by dividing the space into regular voxels or triangular meshes, improving the space occupancy efficiency. TSDF (Truncated Signed Distance Function) and Octomap are typical representatives. The former is suitable for voxel-level dense mapping, and the latter can achieve semantic level expression. However, this type of method has limited detail expression ability in complex texture scenes, and still needs to additionally remove dynamic points in dynamic scenes to ensure consistency.

[0005] The 3D mapping method based on Gaussian mapping has gradually become a research hotspot in recent years. This method uses a differentiable 3D Gaussian distribution to replace the discrete point cloud as the scene representation, which not only saves memory overhead but also supports real-time rendering and optimization. Neural3D Gaussian Splatting realizes high-fidelity and high-efficiency scene modeling by assigning attributes such as color, transparency, and uncertainty to each Gaussian point. However, the current reconstruction methods based on Gaussian mapping mostly rely on an offline optimization process and lack the adaptability to online SLAM tasks. Especially in a dynamic environment, they lack a robust dynamic point processing and adaptive point management mechanism and are difficult to achieve continuous update of a high-quality map.

[0006] In summary:

[0007] Traditional SLAM methods usually assume a static environment. However, in real-world scenarios, the presence of dynamic objects can lead to tracking drift and map distortion. In recent years, 3D Gaussian Splatting (3DGS) technology has demonstrated superior rendering performance in the field of visual SLAM. Nevertheless, it still faces the problem of dynamic object interference in dynamic environments. Existing 3DGS-SLAM methods are difficult to effectively separate dynamic and static scenes, resulting in increased pose estimation errors and degraded map quality.

[0008] Therefore, a new solution to the above problems is needed. Summary of the Invention

[0009] The object of the present invention is to provide an adaptive dynamic SLAM method based on Gaussian distribution. Through innovative strategies such as Gaussian point adaptive management, dynamic target elimination, and optical flow-based background compensation, high-fidelity map reconstruction and robust pose estimation are achieved to solve the technical problems presented in the background art.

[0010] To achieve the above object, the present invention provides the following technical solution: An adaptive dynamic SLAM method based on Gaussian distribution, at least including the following steps:

[0011] S1: Build a dynamic target elimination module based on CRF. Through the introduction of YOLO for preliminary dynamic detection, combine CRF segmentation to optimize the dynamic region mask, and eliminate dynamic points through depth constraint and reprojection error to improve the consistency of the map and avoid the interference of dynamic objects on the SLAM system. The CRF is a conditional random field;

[0012] S2: Build an optical flow-based background repair module. Existing SLAM systems are difficult to remove artifacts brought by moving objects in dynamic environments, affecting the mapping quality. Adopt the optical flow analysis method to detect the inter-frame motion consistency and combine it with the background model for compensation, so that a more stable mapping effect can be obtained for the static environment;

[0013] S3: Adaptive 3D Gaussian mapping initialization. Traditional SLAM mapping methods rely on dense point clouds or voxels, resulting in large storage and computational overheads. Use 3D Gaussian points to represent the scene and adaptively adjust the initialization density according to color gradients and depth changes to improve the reconstruction accuracy of key regions while reducing the storage cost of low-texture regions;

[0014] S4: Introduce a Gaussian point density adaptive insertion strategy. In response to the complexity changes of the scene, increase the Gaussian point density in high-texture regions to enhance the detail restoration ability; reduce the point density in low-texture regions to improve the computational efficiency; at the same time, adaptively insert new Gaussian points based on parameters such as the number of observations and transparency to ensure the accuracy of map updates;

[0015] S5: Further propose a Gaussian point deletion mechanism based on transparency control. Different from traditional SLAM which requires complex keyframe management, by introducing a Gaussian point transparency metric, adaptively delete points that have not been observed for a long time or have low transparency to reduce redundant information and improve the storage efficiency of mapping.

[0016] Furthermore, the CRF-based dynamic object elimination module is used for dynamic object elimination. Performing dynamic object elimination at least includes the following steps:

[0017] First, extract the regions that may contain dynamic objects based on the YOLO detection boxes, and select a set of feature points P = {p1, p2, p3,..., p n}, and optimize using the geometric consistency and observation attributes of the feature points

[0018] In the CRF, each feature point p i is used as a node, and its label x i ∈ {0, 1}, where 0 represents the static background and 1 represents the dynamic foreground. The entire CRF optimization objective is to minimize the Gibbs energy function:

[0019]

[0020] Among them, ψ u (x i ) is the single-point potential energy, measuring the possibility of each point belonging to the foreground or background; ψ p (x i , x j ) is the two-point potential energy, used to model the spatial relationship between adjacent points;

[0021] The single-point potential energy term is modeled according to the geometric characteristics and motion characteristics of each point, and is expressed as follows:

[0022] ψ u (x i ) = -logP static (p i )

[0023] Among them, the static probability P static (p i ) is modeled by four Gaussian distributions of reprojection error, depth change, number of observations, and epipolar distance:

[0024]

[0025] Among them, represents the k-th Gaussian distribution, representing the observation model of the point p i under the static background; π k is the weight of each Gaussian distribution; μ k and They are the mean and variance of the k-th distribution respectively;

[0026] The Gaussian kernel function is used to model the similarity between points, and the two-point potential term is defined as follows:

[0027]

[0028] where μ(x i , x j ) adopts the Potts model, and when the labels of two points are different (i.e., one is dynamic and the other is static), this term imposes a higher penalty; ω m is the weight of each kernel function; k m (f i , f j ) represents the similarity kernel function between similar points and includes:

[0029] Feature similarity kernel:

[0030]

[0031] where α i and α j are the reprojection errors; γ i and γ j are the number of observations; σ α and σ γ are the standard deviations used to control the differences in reprojection errors and the number of observations;

[0032]

[0033] where P i and P j are the point cloud coordinates, and p i and p j are the image coordinates;

[0034] To avoid miseliminating potential static points, a reprojection detection and pose estimation mechanism for dynamic environments is proposed in the dynamic object elimination module based on CRF to improve the camera pose estimation accuracy of the system under the interference of moving objects. As the core component in the tracking thread, this mechanism mainly includes four stages: static feature point screening, reprojection error evaluation, key frame pose optimization, and feature point weighting processing;

[0035] First, receive the background probability map output from the dynamic elimination module, assign static confidence weights to the feature points in the image, and use this as the basis for distinguishing dynamic and static features. By performing staticity analysis on all candidate feature points in the key frame, and without completely excluding potential static points, preferentially retain the feature points with higher static confidence;

[0036] In the pose estimation stage, according to the reprojection error performance between the front and rear frame images, dynamic feature points with significant errors are filtered out, and only static points are retained to participate in the pose solution process. Subsequently, the relative pose estimation between key frames is constructed with static feature points, and an optimization strategy is adopted to jointly solve the rotation and displacement parameters of the camera, and the map point projection and feature point weights are globally adjusted;

[0037] The reprojection detection and pose estimation mechanism for dynamic environments models the differences between dynamic and static feature points, introduces a weighted method based on background credibility, strengthens the dependence of pose estimation on static points, effectively reduces the estimation deviation caused by dynamic interference. At the same time, the system avoids the blind elimination of feature points in traditional methods and improves the tracking accuracy and mapping stability.

[0038] Furthermore, a background repair mechanism combining optical flow analysis and dynamic region reconstruction is designed in the optical flow-based background repair module;

[0039] The optical flow-based background repair module receives the static mask map Md output from the CRF-based dynamic object elimination module, extracts sparse feature points between key frames, and constructs a static optical flow vector V based on the optical flow difference between the front and rear frames static , and its calculation formula is:

[0040] V static =V scenes -V observed

[0041] where V scenes is the overall scene optical flow caused by camera movement, and V observed is the actually measured optical flow;

[0042] Subsequently, according to the pixel optical flow intensity and the static optical flow threshold δ, all pixels are divided into three types of static pixel sets, low-dynamic pixel sets, and high-dynamic pixel sets;

[0043] To improve the temporal consistency and spatial accuracy of background restoration, a sliding window mechanism is designed in the optical flow-based background repair module, that is, the local image group is composed of the nearest 20 key frames, and their relative poses are calculated in turn. For the areas belonging to low-dynamic pixels, interpolation and reconstruction are performed in combination with the front and rear frame images to complete the compensation of the static background;

[0044] Finally, the repaired image and depth map are used for the subsequent Gaussian mapping process to ensure that the static background in the dynamic scene remains continuous and consistent.

[0045] Furthermore, the adaptive 3D Gaussian mapping initialization solves the problems of high redundancy in initialization and poor reconstruction quality in low-texture areas in traditional SLAM systems. The adaptive 3D Gaussian mapping initialization is a method for constructing adaptive 3D Gaussian primitives based on multi-modal perception, and specifically includes the following steps:

[0046] Use a 3D Gaussian ellipsoid with geometric and optical properties as the basic mapping unit;

[0047] By constructing an image color gradient field and a depth change field, establish a pixel-level texture complexity quantization model. Use a compact Gaussian distribution in high-frequency texture areas (such as edges and detailed textures), and expand the primitive coverage range in low-frequency areas (such as smooth surfaces) to achieve a balance between storage efficiency and reconstruction accuracy. The specific formula is:

[0048]

[0049] Where, σ represents the standard deviation of the Gaussian ellipsoid; represents the image color gradient; λ is the expansion factor;

[0050] During the initialization of the first frame, eliminate dynamic interference pixels through a motion mask, and perform multi-scale feature sampling on the static background area;

[0051] Establish the topological structure of the Gaussian primitive space in the world coordinate system through back-projection. The center position of the primitive is jointly solved by the pixel depth and the internal and external camera parameters to ensure geometric consistency. The formula is:

[0052] X = K -1 ·p·d

[0053] Where, X is the center position of the primitive; K is the camera internal parameter matrix; p is the image pixel coordinate; d is the pixel depth;

[0054] Define a differentiable attribute set for each Gaussian primitive. The differentiable attribute set includes a position covariance matrix, an anisotropic scaling factor, and a spherical harmonic illumination coefficient;

[0055] The covariance matrix is automatically solved through the spatial distribution characteristics of neighboring point clouds, and the spherical harmonic coefficients are initialized based on the local illumination distribution. The calculation formula of the covariance matrix is:

[0056]

[0057] Where, x′ i is the coordinate of the neighboring point; N is the number of neighboring points; μ is the mean vector of the neighboring points; T represents the transpose operation.

[0058] Furthermore, the Gaussian point density adaptive insertion strategy is used to achieve the adaptive update of the map in a dynamic scene. The Gaussian point density adaptive insertion strategy is a density control strategy based on the bidirectional feedback of observation stability and geometric error. By calculating the color entropy value and depth variance of the image block online, the scene is divided into a high-dynamic area, a medium-detail area, and a smooth area;

[0059] Adjust the sampling interval t according to the color complexity of the image and the change of depth. Refer to the following formula:

[0060]

[0061] where t min represents the sampling interval of the high-dynamic area, α and β are control factors, Entropy(I) and Var(D) are the image entropy value and depth variance respectively;

[0062] Design a double-threshold trigger mechanism. When the cumulative opacity of pixels is lower than the preset reliability threshold or the depth residual exceeds the adaptive error threshold, generate candidate Gaussian primitives. The candidate primitives need to pass the geometric consistency verification. The formula is:

[0063] ||P render -P obs ||2 < λ reproj

[0064] where P render is the position of the primitive in the rendered image; P obs is the position in the observed image; λ reproj is the reprojection error threshold.

[0065] Furthermore, the Gaussian point deletion mechanism based on transparency control is used to solve the problem of map degradation in a dynamic environment. The Gaussian point deletion mechanism based on transparency control is a Gaussian primitive elimination strategy based on probability inference and spatio-temporal consistency;

[0066] The transparency attenuation model in the Gaussian primitive elimination strategy based on probability inference and spatio-temporal consistency adopts a dynamic attenuation function:

[0067] α(t) = α0·e -kt

[0068] where α0 is the initial transparency, e is the base of the natural logarithm, t is the time variable, and the attenuation coefficient k is jointly determined by the observation frequency of the primitive and the uncertainty of pose estimation. When the cumulative transparency is lower than the reliability threshold, it is determined that the primitive is a potential dynamic interference or an outlier, and the deletion operation is triggered;

[0069] Online monitor the anisotropic parameters of the Gaussian ellipsoid. When the length of the main axis exceeds the scene scale constraint or the ratio of the main axis to the minor axis exceeds the safety threshold, perform rigid shape correction. The primitives that cannot meet the constraints will be marked as invalid. At the same time, construct a sliding key-frame window and project the candidate primitives to be deleted onto all key frames;

[0070] If the reprojection consistency condition cannot be met in the exceeded key frames, then determine that the primitive is a dynamic residual artifact and perform permanent deletion;

[0071] Through the cooperation of differentiable Gaussian representation and probabilistic elimination mechanism, realize the autonomous evolution of the map in a dynamic environment.

[0072] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0073] 1. The present invention improves the positioning accuracy in a dynamic environment. Through the dynamic target elimination strategy based on CRF, combined with YOLO detection and conditional random field optimization, the system can accurately identify and eliminate the interference of dynamic objects. Traditional SLAM systems are prone to pose offset affected by moving objects in a dynamic environment. However, the present invention introduces reprojection error evaluation and feature point weighting processing, performs differential modeling on static points and dynamic points, strengthens the dependence of pose estimation on static features, effectively reduces the estimation error caused by dynamic interference, and improves the positioning accuracy and stability of the system in complex dynamic scenarios;

[0074] 2. The present invention enhances the... Figure 1 consistency and integrity. When traditional SLAM systems process dynamic scenes, they often simply eliminate dynamic regions and leave gaps, affecting the integrity of the map. The present invention, through the background repair mechanism based on optical flow, analyzes the difference between camera movement and actual optical flow, divides pixels into static, low-dynamic, and high-dynamic categories, and uses a sliding window to interpolate and reconstruct the front and back frame images, effectively restoring the static background occluded by dynamic objects, ensuring that the map remains continuously consistent in dynamic scenes, and improving the reconstruction quality and visual effect;

[0075] 3. The invention improves the mapping efficiency and scene adaptability. The present invention uses a 3D Gaussian ellipsoid as the basic mapping unit and designs an adaptive Gaussian mapping initialization, density adjustment, and deletion mechanism, significantly improving the system performance. Through pixel-level texture complexity analysis, the system uses a compact Gaussian distribution in high-frequency texture regions and expands the coverage range in low-frequency regions, balancing the storage efficiency and reconstruction accuracy. At the same time, the Gaussian point management mechanism based on transparency control can intelligently eliminate potential dynamic interference or outliers, combined with an adaptive point density insertion strategy, realizing the autonomous evolution ability of the map. Description of the Drawings

[0076] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the accompanying drawings required for the description of the embodiments. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can be obtained based on these drawings.

[0077] Figure 1 It is a schematic diagram of the overall structure of the present invention. Specific embodiments

[0078] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments.

[0079] Please refer to Figure 1 , the adaptive dynamic SLAM method based on Gaussian distribution includes at least the following steps:

[0080] S1: Build a dynamic target elimination module based on CRF. Through introducing YOLO for preliminary dynamic detection, combining CRF segmentation to optimize the dynamic region mask, and performing dynamic point elimination through depth constraint and reprojection error to improve the consistency of the map and avoid the interference of dynamic objects on the SLAM system. CRF is a conditional random field;

[0081] S2: Build a background repair module based on optical flow. Existing SLAM systems are difficult to remove artifacts brought by moving objects in a dynamic environment, which affects the mapping quality. Adopt the optical flow analysis method to detect the frame-to-frame motion consistency and combine with the background model for compensation, so that a more stable mapping effect can be obtained for the static environment;

[0082] S3: Adaptive 3D Gaussian mapping initialization. Traditional SLAM mapping methods rely on dense point clouds or voxels, resulting in large storage and computational overheads. Use 3D Gaussian points to represent the scene and adaptively adjust the initialization density according to color gradients and depth changes to improve the reconstruction accuracy of key regions while reducing the storage cost of low-texture regions;

[0083] S4: Introduce a Gaussian point density adaptive insertion strategy. For the complexity changes of the scene, increase the Gaussian point density in high-texture regions to improve the detail restoration ability; reduce the point density in low-texture regions to improve the computational efficiency; at the same time, adaptively insert new Gaussian points based on parameters such as the number of observations and transparency to ensure the accuracy of map updates;

[0084] S5: Further propose a Gaussian point deletion mechanism based on transparency control. Different from traditional SLAM which requires complex keyframe management, by introducing a Gaussian point transparency metric, adaptively delete points that have not been observed for a long time or have low transparency to reduce redundant information and improve the storage efficiency of mapping.

[0085] The dynamic object elimination module based on CRF is used for dynamic object elimination. Performing dynamic object elimination includes at least the following steps:

[0086] First, extract the region that may contain dynamic objects based on the YOLO detection box, and select a set of feature points P = {p1, p2, p3,..., p n}, and optimize using the geometric consistency and observation attributes of the feature points

[0087] In CRF, each feature point p i is used as a node, and its label x i ∈ {0, 1}, where 0 represents the static background and 1 represents the dynamic foreground. The entire CRF optimization objective is to minimize the Gibbs energy function:

[0088]

[0089] where, ψ u (x i ) is the single-point potential energy, measuring the possibility of each point belonging to the foreground or background; ψ p (x i , x j ) is the two-point potential energy, used to model the spatial relationship between adjacent points;

[0090] The single-point potential energy term is modeled according to the geometric characteristics and motion characteristics of each point, and is expressed as follows:

[0091] ψ u (x i ) = -logP static (p i )

[0092] where, the static probability P static (p i ) is modeled by four Gaussian distributions of reprojection error, depth change, number of observations, and epipolar distance:

[0093]

[0094] where, represents the k-th Gaussian distribution, representing the observation model of point p i under the static background; π k is the weight of each Gaussian distribution; μ k and They are the mean and variance of the k-th distribution respectively;

[0095] The Gaussian kernel function is used to model the similarity between points. The two-point potential term is defined as follows:

[0096]

[0097] where μ(x i , x j ) adopts the Potts model, which imposes a higher penalty when the labels of two points are different (i.e., one is dynamic and the other is static); ω m is the weight of each kernel function; k m (f i , f j ) represents the similarity kernel function between similar points, including:

[0098] Feature similarity kernel:

[0099]

[0100] where α i and α j are the reprojection errors; γ i and γ j are the number of observations; σ α and σ γ are the standard deviations used to control the differences in reprojection errors and the number of observations;

[0101]

[0102] where P i and P j are the point cloud coordinates, and p i and p j are the image coordinates;

[0103] To avoid miseliminating potential static points, a reprojection detection and pose estimation mechanism for dynamic environments is proposed in the dynamic object elimination module based on CRF to improve the camera pose estimation accuracy of the system under the interference of moving objects. As the core component in the tracking thread, this mechanism mainly includes four stages: static feature point screening, reprojection error evaluation, key frame pose optimization, and feature point weighting processing;

[0104] First, receive the background probability map output from the dynamic elimination module, assign static confidence weights to the feature points in the image, and use this as the basis for distinguishing dynamic and static features. By performing staticity analysis on all candidate feature points in the key frame, and without completely excluding potential static points, preferentially retain the feature points with higher static confidence;

[0105] In the pose estimation stage, according to the reprojection error performance between the front and rear frame images, dynamic feature points with significant errors are filtered out, and only static points are retained to participate in the pose solution process. Subsequently, the relative pose estimation between key frames is constructed with static feature points, and an optimization strategy is used to jointly solve the rotation and displacement parameters of the camera, and the map point projection and feature point weights are globally adjusted;

[0106] The reprojection detection and pose estimation mechanism for dynamic environments models the differences between dynamic and static feature points, introduces a weighted method based on background credibility, strengthens the dependence of pose estimation on static points, effectively reduces the estimation deviation caused by dynamic interference. At the same time, the system avoids the blind elimination of feature points in traditional methods and improves the tracking accuracy and mapping stability.

[0107] There is a background repair mechanism that combines optical flow analysis and dynamic area reconstruction designed in the optical flow-based background repair module;

[0108] The optical flow-based background repair module receives the static mask map Md output from the CRF-based dynamic object elimination module, extracts sparse feature points between key frames, and constructs a static optical flow vector V based on the optical flow difference between the front and rear frames static , and its calculation formula is:

[0109] V static =V scenes -V oserved

[0110] where V scenes is the overall scene optical flow caused by camera movement, and V observed is the actually measured optical flow;

[0111] Subsequently, according to the optical flow intensity of pixel points and the static optical flow threshold δ, all pixels are divided into three types of static pixel sets, low-dynamic pixel sets, and high-dynamic pixel sets;

[0112] To improve the temporal consistency and spatial accuracy of background recovery, a sliding window mechanism is designed in the optical flow-based background repair module, that is, the local image group is composed of the nearest 20 key frames, and their relative poses are calculated in turn. For the areas belonging to low-dynamic pixels, interpolation and reconstruction are performed in combination with the front and rear frame images to complete the compensation of the static background;

[0113] Finally, the repaired image and depth map are used for the subsequent Gaussian mapping process to ensure that the static background in the dynamic scene remains continuous and consistent.

[0114] Adaptive 3D Gaussian mapping initialization solves the problems of high initialization redundancy and poor reconstruction quality in low-texture areas of traditional SLAM systems in dynamic scenarios. Adaptive 3D Gaussian mapping initialization is a method for constructing adaptive 3D Gaussian primitives based on multi-modal perception, which specifically includes the following steps:

[0115] Use a 3D Gaussian ellipsoid with geometric and optical properties as the basic mapping unit;

[0116] By constructing an image color gradient field and a depth change field, a pixel-level texture complexity quantization model is established. A compact Gaussian distribution is used in high-frequency texture areas (such as edges and detailed textures), while the primitive coverage range is extended in low-frequency areas (such as smooth surfaces) to achieve a balance between storage efficiency and reconstruction accuracy. The specific formula is:

[0117]

[0118] where, σ represents the standard deviation of the Gaussian ellipsoid; represents the image color gradient; λ is the expansion factor;

[0119] During the initialization of the first frame, dynamic interference pixels are removed through a motion mask, and multi-scale feature sampling is performed on the static background area;

[0120] Through back-projection, a topological structure of the Gaussian primitive space in the world coordinate system is established. The center position of the primitive is jointly solved by the pixel depth and the internal and external camera parameters to ensure geometric consistency. The formula is:

[0121] X = K -1 ·p·d

[0122] where, X is the center position of the primitive; K is the camera internal parameter matrix; p is the image pixel coordinate; d is the pixel depth;

[0123] Define a differentiable attribute set for each Gaussian primitive. The differentiable attribute set includes a position covariance matrix, an anisotropic scaling factor, and spherical harmonic illumination coefficients;

[0124] The covariance matrix is automatically solved through the spatial distribution characteristics of the neighborhood point cloud, and the spherical harmonic coefficients are initialized based on the local illumination distribution. The calculation formula of the covariance matrix is:

[0125]

[0126] where, x’ i is the coordinate of the neighborhood point; N is the number of neighborhood points; μ is the mean vector of the neighborhood points; T represents the transpose operation.

[0127] The Gaussian point density adaptive insertion strategy is used to achieve the adaptive update of the map in a dynamic scene. The Gaussian point density adaptive insertion strategy is a density control strategy based on the two-way feedback of observation stability and geometric error. By calculating the color entropy value and depth variance of the image block online, the scene is divided into a high-dynamic area, a medium-detail area, and a smooth area;

[0128] Adjust the sampling interval t according to the color complexity of the image and the change of depth. See the following formula:

[0129]

[0130] where t min represents the sampling interval of the high-dynamic area, α and β are control factors, Entropy(I) and Var(D) are the image entropy value and depth variance respectively;

[0131] Design a double-threshold trigger mechanism. When the cumulative opacity of pixels is lower than the preset reliability threshold or the depth residual exceeds the adaptive error threshold, generate candidate Gaussian primitives. The candidate primitives need to pass the geometric consistency verification. The formula is:

[0132] ||P render -P obs ||2 < λ reproj

[0133] where P render is the position of the primitive in the rendered image; P obs is the position in the observed image; λ reproj is the reprojection error threshold.

[0134] The Gaussian point deletion mechanism based on transparency control is used to solve the problem of map degradation in a dynamic environment. The Gaussian point deletion mechanism based on transparency control is a Gaussian primitive elimination strategy based on probability inference and spatio-temporal consistency;

[0135] The transparency attenuation model in the Gaussian primitive elimination strategy based on probability inference and spatio-temporal consistency adopts a dynamic attenuation function:

[0136] α(t) = α0·e -kt

[0137] where α0 is the initial transparency, e is the base, t is the time variable, and the attenuation coefficient k is jointly determined by the observation frequency of the primitive and the uncertainty of pose estimation. When the cumulative transparency is lower than the reliability threshold, determine that the primitive is a potential dynamic interference or outlier and trigger the deletion operation;

[0138] Online monitor the anisotropic parameters of the Gaussian ellipsoid. When the length of the main axis exceeds the scene scale constraint or the ratio of the main axis to the secondary axis exceeds the safety threshold, perform rigid shape correction. The primitives that cannot meet the constraints will be marked as invalid. At the same time, construct a sliding key-frame window and project the candidate primitives to be deleted onto all key frames;

[0139] If the reprojection consistency condition cannot be met in the exceeded key frames, then determine that the primitive is a dynamic residual artifact and perform permanent deletion;

[0140] Through the cooperation of differentiable Gaussian representation and probabilistic elimination mechanism, realize the autonomous evolution of the map in a dynamic environment.

[0141] To sum up:

[0142] Experiments show that under the same reconstruction quality, the storage requirement of the present invention is reduced by about 30% - 45% compared with the traditional point cloud representation, the computing efficiency is significantly improved, and it can adapt to various application scenarios from static indoor to complex dynamic outdoor, and maintain the stability and adaptability of the system during long-term operation.

[0143] Furthermore, the following advanced technologies are reflected:

[0144] 1. Dynamic object elimination technology based on CRF

[0145] The present invention proposes a dynamic object elimination method combining YOLO detection and conditional random field (CRF) optimization. This method first detects potential dynamic regions through YOLO, and then uses the CRF model to further optimize the dynamic region mask. A multi-dimensional Gaussian distribution model based on reprojection error, depth change, number of observations, and epipolar distance is established to quantify the static probability. The system is optimized through the geometric consistency and observation attributes of feature points, and a reprojection detection and pose estimation mechanism for dynamic environments is introduced to avoid mis-eliminating potential static points and improve the camera pose estimation accuracy of the system under the interference of moving objects.

[0146] 2. Adaptive 3D Gaussian mapping construction and management method

[0147] The present invention uses a 3D Gaussian ellipsoid as the basic mapping unit and proposes a complete set of Gaussian point adaptive management strategies. This method establishes a pixel-level texture complexity quantification model through the image color gradient field and depth change field, uses a compact Gaussian distribution in high-texture regions, and expands the coverage range in low-texture regions. The system divides the scene complexity using image entropy values and depth variances, adopts different sampling intervals for different regions, and monitors and deletes Gaussian points that may be dynamic residues or invalid through a transparency control mechanism. This set of methods significantly optimizes the storage and computing efficiency of the system, realizes the autonomous evolution ability of the map, and enables the system to adapt to various complex dynamic scenarios and maintain stability during long-term operation.

[0148] It is obvious to those skilled in the art that the present invention is not limited to the details of the above-described exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, in any regard, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Accordingly, all changes that fall within the meaning and scope of the equivalent elements of the claims are intended to be embraced within the present invention. Any reference signs in the claims should not be construed as limiting the claims involved.

Claims

1. An adaptive dynamic SLAM method based on Gaussian distribution, characterized in that: At least include the following steps: S1: Build a dynamic object rejection module based on CRF. Through introducing YOLO for preliminary dynamic detection, combine CRF segmentation to optimize the dynamic region mask, and perform dynamic point rejection through depth constraint and reprojection error to improve the consistency of the map and avoid the interference of dynamic objects on the SLAM system. The CRF is the conditional random field; S2: Build an optical flow-based background repair module. Adopt the optical flow analysis method to detect the inter-frame motion consistency and combine with the background model for compensation, so that a more stable mapping effect can be obtained for the static environment; S3: Adaptive 3D Gaussian mapping initialization. Use 3D Gaussian points to represent the scene and adaptively adjust the initialization density according to the color gradient and depth change to improve the reconstruction accuracy of key regions and reduce the storage cost of low-texture regions at the same time; S4: Introduce a Gaussian point density adaptive insertion strategy. For the complexity change of the scene, increase the Gaussian point density in high-texture regions to improve the detail restoration ability; reduce the point density in low-texture regions to improve the calculation efficiency; at the same time, based on parameters such as the number of observations and transparency, adaptively insert new Gaussian points to ensure the accuracy of map update; S5: Further propose a Gaussian point deletion mechanism based on transparency control. By introducing the Gaussian point transparency metric, adaptively delete the points that have not been observed for a long time or have low transparency to reduce redundant information and improve the storage efficiency of mapping.

2. The adaptive dynamic SLAM method based on Gaussian distribution according to claim 1, characterized in that: The dynamic object rejection module based on CRF is used for dynamic object rejection. Performing dynamic object rejection at least includes the following steps: First, based on the YOLO detection boxes, extract the regions that are likely to contain dynamic objects, and select a set of feature points P = {p1, p2, p3,..., p n}, and optimize them using the geometric consistency and observation attributes of the feature points In the CRF, for each feature point p i is regarded as a node, and its label x i ∈ {0, 1} is defined, where 0 represents the static background and 1 represents the dynamic foreground. The overall CRF optimization objective is to minimize the Gibbs energy function: where, ψ u (x i ) is a single-point potential energy that measures the likelihood of each point belonging to the foreground or background; ψ p (x i , x j ) is a two-point potential energy used to model the spatial relationship between adjacent points; The single-point potential term is modeled according to the geometric and motion characteristics of each point and is expressed as follows: ψ u (x i )=-logP static (p i ) Among them, the static probability P static (p i ) is modeled by four Gaussian distributions: reprojection error, depth change, number of observations, and epipolar distance Among them, represents the k-th Gaussian distribution, and represents the point p i is the observation model under a static background; π k is the weight of each Gaussian distribution; μ k and are the mean and variance of the k-th distribution respectively; Use the Gaussian kernel function to model the similarity between points, and the two-point potential term is defined as follows: Among them, μ(x i ,x j ) adopts the Potts model, and when the labels of two points are different, this term imposes a higher penalty; ω m is the weight of each kernel function; k m (f i ,f j ) represents the similarity kernel function between similar points and includes: Feature similarity kernel: where α i and α j are reprojection errors; γ i and γ j are the number of observations; σ α and σ γ are the standard deviations used to control the differences in reprojection errors and the number of observations; Where P i and P j are point cloud coordinates, and p i and p j are image coordinates; To avoid mis-rejecting potential static points, a reprojection detection and pose estimation mechanism for dynamic environments is proposed in the dynamic object rejection module based on CRF, which is used to improve the camera pose estimation accuracy of the system under the interference of moving objects. This mechanism is the core component in the tracking thread and mainly includes four stages: static feature point screening, reprojection error evaluation, key frame pose optimization, and feature point weighting processing; First, receive the background probability map output from the dynamic rejection module, assign static confidence weights to the feature points in the image, and use this as the basis for distinguishing dynamic and static features. Through the staticity analysis of all candidate feature points in the key frame, without completely excluding potential static points, preferentially retain the feature points with higher static confidence; In the pose estimation stage, according to the reprojection error performance between the front and rear frame images, screen out the dynamic feature points with significant errors and only retain the static points to participate in the pose solution process. Subsequently, construct the relative pose estimation between key frames with static feature points, adopt the optimization strategy to jointly solve the rotation and displacement parameters of the camera, and globally adjust the map point projection and feature point weights; The reprojection detection and pose estimation mechanism for dynamic environments models the differences between dynamic and static feature points, introduces a weighted method based on background credibility, strengthens the dependence of pose estimation on static points, effectively reduces the estimation offset caused by dynamic interference. At the same time, the system avoids the blind elimination of feature points in traditional methods, improving the tracking accuracy and mapping stability.

3. The adaptive dynamic SLAM method based on Gaussian distribution according to claim 1, wherein: In the optical flow-based background repair module, a background repair mechanism combining optical flow analysis and dynamic area reconstruction is designed; The background repair module based on optical flow receives the static mask graph Md output by the dynamic target elimination module based on CRF, extracts sparse feature points between key frames, and constructs a static optical flow vector V based on the optical flow difference between the front and back frames static , and its calculation formula is: V static = V scenes - V observed Among them, V scenes is the overall scene optical flow caused by camera movement, and V observed is the actually measured optical flow; Subsequently, according to the optical flow intensity of pixel points and the static optical flow threshold δ, all pixels are divided into three categories: static pixel set, low-dynamic pixel set, and high-dynamic pixel set; To improve the temporal consistency and spatial accuracy of background recovery, a sliding window mechanism is designed in the optical flow-based background repair module, that is, the 20 most recent key frames are used to form a local image group, and their relative poses are calculated in sequence. For the areas belonging to low-dynamic pixels, interpolation and reconstruction are performed by combining the front and back frame images to complete the compensation of the static background; Finally, the repaired image and depth map are used for the subsequent Gaussian mapping process to ensure the continuity and consistency of the static background in the dynamic scene.

4. The adaptive dynamic SLAM method based on Gaussian distribution according to claim 1, wherein: The adaptive 3D Gaussian mapping initialization solves the problems of high initialization redundancy and poor reconstruction quality in low-texture areas of traditional SLAM systems in dynamic scenes. The adaptive 3D Gaussian mapping initialization is a method for constructing adaptive 3D Gaussian basis elements based on multi-modal perception, which specifically includes the following steps: Use a 3D Gaussian ellipsoid with geometric and optical properties as the basic mapping unit; By constructing an image color gradient field and a depth change field, a pixel-level texture complexity quantization model is established. A compact Gaussian distribution is used in high-frequency texture areas, while the basis element coverage range is expanded in low-frequency areas to achieve a balance between storage efficiency and reconstruction accuracy. The specific formula is: where σ represents the standard deviation of the Gaussian ellipsoid; represents the image color gradient; λ is the expansion factor; At the initialization of the first frame, dynamic interference pixels are removed through a motion mask, and multi-scale feature sampling is performed on the static background area; The spatial topology of Gaussian basis elements in the world coordinate system is established through back-projection. The center position of the basis element is jointly solved by pixel depth and camera internal and external parameters to ensure geometric consistency. The formula is: X = K -1 ·p·d Among them, X is the center position of the basis element; K is the camera internal parameter matrix; p is the image pixel coordinate; d is the pixel depth; Define a differentiable attribute set for each Gaussian basis element, and the differentiable attribute set includes a position covariance matrix, an anisotropic scaling factor, and spherical harmonic illumination coefficients; The covariance matrix is automatically solved through the spatial distribution characteristics of neighboring point clouds, and the spherical harmonic coefficients are initialized based on the local illumination distribution. The calculation formula of the covariance matrix is: wherein, are the coordinates of the domain points; N is the number of domain points; μ is the mean vector of the neighborhood points; and T represents the transpose operation.

5. The adaptive dynamic SLAM method based on Gaussian distribution according to claim 1, characterized in that: The Gaussian point density adaptive insertion strategy is used to achieve the adaptive update of the map in dynamic scenes. The Gaussian point density adaptive insertion strategy is a density control strategy based on the bidirectional feedback of observation stability and geometric error. By online calculating the color entropy value and depth variance of image blocks, the scene is divided into high-dynamic areas, medium-detail areas, and smooth areas; Adjust the sampling interval t according to the color complexity of the image and the change in depth. Refer to the following formula: where t min represents the sampling interval of the high dynamic range region, α and β are control factors, Entropy(I) and Var(D) are the image entropy value and depth variance respectively; Design a dual-threshold trigger mechanism. When the cumulative opacity of pixels is lower than the preset reliability threshold or the depth residual exceeds the adaptive error threshold, generate candidate Gaussian primitives. The candidate primitives need to pass geometric consistency verification. The formula is as follows: ||P render -P obs ||2<λ reproj Among them, P render is the position of the primitive in the rendered image; P obs is the position in the observed image; λ reproj is the reprojection error threshold.

6. The adaptive dynamic SLAM method based on Gaussian distribution according to claim 1, characterized in that: The Gaussian point deletion mechanism based on transparency control is used to solve the problem of map degradation in dynamic environments. The Gaussian point deletion mechanism based on transparency control is the Gaussian primitive elimination strategy based on probabilistic inference and spatio-temporal consistency; The transparency attenuation model in the Gaussian primitive elimination strategy based on probabilistic inference and spatio-temporal consistency adopts a dynamic attenuation function: α(t) = α0·e -kt Among them, α0 is the initial transparency, e is the base of the natural logarithm, t is the time variable, and the attenuation coefficient k is jointly determined by the observation frequency of the primitive and the uncertainty of pose estimation. When the cumulative transparency is lower than the reliability threshold, it is determined that the primitive is a potential dynamic interference or outlier, and the deletion operation is triggered; Online monitor the anisotropic parameters of the Gaussian ellipsoid. When the length of the major axis exceeds the scene scale constraint or the ratio of the major axis to the minor axis exceeds the safety threshold, perform rigid shape correction. The primitives that cannot meet the constraints will be marked as invalid. At the same time, construct a sliding key-frame window and project the candidate primitives to be deleted to all key frames; When the reprojection consistency condition cannot be met in the exceeded key frames, it is determined that the primitive is a dynamic residual artifact and a permanent deletion is performed; Through the cooperation of the differentiable Gaussian representation and the probabilistic elimination mechanism, the autonomous evolution of the map in a dynamic environment is realized.

Images