Monocular neural rendering SLAM method and system in complex environment

Abstract

The invention relates to a monocular neural rendering SLAM method and system in a complex environment. According to the method, an initial image sequence in a complex environment is obtained and preprocessed, then a pre-trained semantic segmentation model is input, and a multi-dynamic object mask is obtained; extracting point features and line features from the preprocessed image sequence, and eliminating corresponding dynamic features by using a multi-dynamic object mask to obtain static features; performing feature matching on the static features, solving a conversion matrix, and obtaining an estimated pose by optimizing a point-line combined re-projection error function; obtaining sparse superprimitives, recognizing a to-be-densified area according to the re-projection error of the points, and generating a new superprimitive set in combination with unmatched feature points; and obtaining a rendered image, and performing multi-scale Gaussian pyramid supervised training on the rendered image to obtain a final global map. Compared with the prior art, the method has the advantages of improving the pose estimation accuracy and map characterization capability of the SLAM in a complex environment and the like.

Description

Technical Field

[0001] The present invention relates to the field of robot visual positioning and mapping technology, and in particular to a monocular neural rendering SLAM method and system in a complex environment. Background Art

[0002] In recent years, mobile robots have been used more and more widely in our daily lives. Traditional laser-based SLAM systems have been very mature in indoor positioning and mapping. However, in special application scenarios, dynamic object interference or degraded environments such as long corridors will also face huge challenges. In addition, laser sensors are expensive and have limited detection range. Therefore, more and more SLAM solutions tend to rely on vision to build systems.

[0003] Traditional point-feature-based visual simultaneous localization and mapping (SLAM) systems also face huge challenges in dynamic environments. Due to the presence of dynamic objects, they often fail to track or have a significant drop in accuracy. Many methods use semantic segmentation networks to identify dynamic objects in images and delete dynamic feature points in the subsequent tracking process, but this method is difficult to promote due to the speed and accuracy of the segmentation network. There are also related methods that propose to remove dynamic points based on the error size of each point, but this method relies on a relatively accurate initial pose and has to introduce new sensors to provide initial value estimates, which greatly increases the complexity of the system. Consider a special indoor scene that contains highly dynamic moving objects, such as crowds. And there are a large number of weak texture areas in the scene. In this case, many methods based on dynamic prior filtering are difficult to operate normally, because after removing the features of dynamic objects, the remaining static background cannot extract enough feature points for matching and initializing the map.

[0004] Therefore, how to improve the pose estimation accuracy and map representation capability of SLAM in complex environments, such as those containing dynamic and weakly textured areas, has become a problem that needs to be solved in this field. Summary of the invention

[0005] The purpose of the present invention is to overcome the defects of the above-mentioned prior art in complex environments containing dynamic and weak texture areas, such as inaccurate pose estimation and insufficient map representation, and to provide a monocular neural rendering SLAM method and system in complex environments.

[0006] The purpose of the present invention can be achieved by the following technical solutions:

[0007] According to a first aspect of the present invention, a monocular neural rendering SLAM method in a complex environment is provided, comprising the following steps: a data acquisition and preprocessing step: acquiring an initial image sequence in a complex environment and performing preprocessing; a dynamic object mask acquisition step: inputting the preprocessed image sequence into a pretrained semantic segmentation model to obtain a multi-dynamic object mask; a feature extraction and dynamic elimination step: extracting point features and line features from the preprocessed image sequence, and using the multi-dynamic object mask to eliminate the corresponding dynamic features to obtain static features; a pose estimation step: based on the point features and the line features, performing feature matching on the static features, solving the transformation matrix, and obtaining an estimated pose by optimizing the reprojection error function of the point-line union; a geometric mapping step: obtaining a sparse hyper-primitive, identifying an area to be densified according to the reprojection error of the point, and generating a new hyper-primitive set in combination with unmatched feature points; a scene mapping step: obtaining a rendered image, and performing multi-scale Gaussian pyramid supervised training on the rendered image to obtain a final global map.

[0008] As a preferred technical solution, the method also includes a loop detection step: loop detection is performed using a preset bag-of-words model, and when a loop is detected, similar changes are used to adjust the current frame and super primitives to correct the posture and correct the accumulated error.

[0009] As a preferred technical solution, the specific implementation process of the dynamic object mask acquisition step includes: tracking and segmenting a single dynamic object in turn to obtain the masks corresponding to all dynamic objects; fusing all masks in each frame to obtain the masks of all dynamic objects in each frame image.

[0010] As a preferred technical solution, in the feature extraction and dynamic elimination steps, the point features are extracted using the ORB detection method, and the line features are extracted using the LSD line detection method.

[0011] As a preferred technical solution, in the pose estimation step, triangulation is performed through feature matching to complete the initialization of the hyper-primitive map, key frames are selected according to the current frame information, and a factor graph is constructed by optimizing the common view point features, line features and key frames. The factor graph is used to optimize the tracking process.

[0012] As a preferred technical solution, the current frame information includes the matching number of point features and line features and feature tracking status.

[0013] As a preferred technical solution, the point-line joint reprojection error function is expressed as:

[0014]

[0015] In the formula, the kth key frame pose is T cw,k= [R, t], the observed i-th spatial point p I The reprojection error is Ep k,i , the reprojection error of the jth spatial line observed in the kth key frame is El k,i , and Represent different observation covariances, ρ p and ρ l They represent different robust kernel functions respectively.

[0016] As a preferred technical solution, the scene mapping step specifically includes: performing Gaussian smoothing and downsampling on the input image to obtain a rendered image; constructing a pyramid to perform supervision at different levels of image resolution; constructing photometric loss and SSIM loss to train and optimize the representation of the rendered image.

[0017] As a preferred technical solution, the luminosity loss and SSIM loss are expressed as:

[0018]

[0019] In the formula, represents the photometric loss and SSIM loss, λ is the preset weight parameter, I r is the rendered image, I gt is the true value of the input image.

[0020] According to a second aspect of the present invention, a monocular neural rendering SLAM system for a complex environment is provided, comprising: a data acquisition and preprocessing module, used to acquire an initial image sequence in a complex environment and perform preprocessing; a dynamic object mask acquisition module, used to input the preprocessed image sequence into a pretrained semantic segmentation model to obtain a multi-dynamic object mask; a feature extraction and dynamic culling module, used to extract point features and line features from the preprocessed image sequence, and use the multi-dynamic object mask to eliminate the corresponding dynamic features to obtain static features; a pose estimation module, used to perform feature matching on the static features based on the point features and the line features, solve the transformation matrix, and obtain an estimated pose by optimizing the reprojection error function of the point-line union; a geometric mapping module, used to acquire sparse hyper-primitives, identify the area to be densified according to the reprojection error of the points, and generate a new hyper-primitive set in combination with unmatched feature points; and a scene mapping module, used to acquire a rendered image, perform multi-scale Gaussian pyramid supervised training on the rendered image, and obtain a final global map.

[0021] Compared with the prior art, the present invention has the following beneficial effects:

[0022] 1. The present invention uses a semantic segmentation model to obtain multiple dynamic object masks, and uses the mask to eliminate the interference of dynamic objects, which can achieve accurate pose estimation in a dynamic environment. At the same time, after removing the dynamic features, the static features are matched, the transformation matrix is ​​solved, and the estimated pose is obtained by optimizing the reprojection error function of the point-line joint. By introducing point features and line features to minimize the error function, accurate pose estimation under weak texture conditions can be achieved. Therefore, the present invention can fully improve the accuracy of pose estimation in complex environments containing dynamic and weak texture areas;

[0023] 2. After obtaining a sparse super-primitive, i.e., a point cloud set, the present invention introduces neural rendering into the SLAM method. After obtaining the rendered image, the image is subjected to multi-scale Gaussian pyramid supervised training to obtain the final global map, which can achieve accurate estimation of the camera pose and enhance the scene representation under multiple perspectives, effectively solving the SLAM problem in dynamic scenes with weak textures;

[0024] 3. The present invention can significantly improve the scalability of the SLAM system, and the algorithm has been verified on multiple data sets. Compared with the existing algorithms, the absolute trajectory error (ATE) and peak signal-to-noise ratio (PSNR) are significantly improved, proving the effectiveness and practicality of the method. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 A schematic diagram of the architecture of the method provided by the present invention;

[0026] Figure 2 A schematic diagram of an implementation flow of a method is provided for an embodiment of the present invention;

[0027] Figure 3 A multi-dynamic object segmentation result diagram of the method provided by the embodiment of the present invention;

[0028] Figure 4 The following is a schematic diagram of the comparison results of mapping of various SLAM methods based on the PSNR indicator provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0029] For special indoor scenes containing highly dynamic moving objects and a large number of weak texture areas, the existing methods based on dynamic prior filtering have certain limitations in positioning and map representation, which easily reduce the accuracy of visual simultaneous positioning and mapping (SLAM). In this regard, the present invention takes into account the particularity of the scene, uses a semantic segmentation model to obtain multiple dynamic object masks to remove the interference of dynamic objects, and introduces neural rendering technology to improve the SLAM method. Among them, neural rendering, as an emerging technology, combines the ideas of classical computer graphics and machine learning, and reconstructs scene representation from observations through the loss of differentiable rendering, thereby synthesizing photo-realistic images and videos. Due to its characteristics of fast rendering and high-quality image representation, it can be introduced into SLAM, thereby improving the shortcomings of existing SLAM mapping methods and providing a new type of map representation method.

[0030] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0031] Example

[0032] like Figure 1 As shown, this embodiment proposes a monocular neural rendering SLAM method in a complex environment. On the one hand, the method uses an open semantic segmentation model to accurately obtain dynamic object priors in multiple scenes, that is, multiple dynamic segmentation masks; on the other hand, in the traditional point feature-based SLAM process, line feature constraints are added to improve the problem of inaccurate pose estimation caused by insufficient number of feature points in weak texture areas; on this basis, a progressive rendering method is used to achieve photo-level mapping results, and scene information under a new perspective can be obtained. Figure 2 One implementation process of the above method is shown, which is as follows:

[0033] Step S1, data acquisition and preprocessing: acquiring an initial image sequence in a complex environment and performing preprocessing.

[0034] Specifically, in practical applications, the initial image sequence can be obtained using a preset sensor. However, due to the sensor's own structure and the particularity of the scene, the obtained image may have problems such as noise, distortion or illumination changes, and cannot be directly used for tracking and mapping of SLAM methods and systems. Therefore, advanced image processing technology can be used to preprocess the image data to improve the validity of the data. Preprocessing includes steps such as denoising, distortion correction, and illumination adjustment, so that the image is more suitable for subsequent reconstruction and analysis.

[0035] Step S2, dynamic object mask acquisition: input the preprocessed image sequence into the pre-trained semantic segmentation model to obtain multiple dynamic object masks.

[0036] Specifically, in complex scenes containing multiple dynamic objects, the open semantic segmentation model SAM2 is used to segment objects of different categories in parallel, and then multiple object masks are fused to obtain accurate multi-dynamic object masks in consecutive frames. The SAM2 segmentation model is used to perform semantic segmentation on the current frame to obtain the category label id∈{0,1,…N} of each pixel.

[0037] Most existing visual SLAM systems are based on static background assumptions and cannot be directly applied to dynamic scenes. In order to overcome this problem, the simplest way is to remove redundant features through dynamic object priors to prevent calculation errors in the matching process. Specifically, using the pre-trained visual segmentation model SAM2, the perception of dynamic objects is achieved by providing the rough pixel positions of people or other dynamic objects, and the corresponding masks are obtained. This method can segment dynamic objects in near real time, and the segmentation results are relatively accurate, and most of the pixel areas where dynamic objects are located can be obtained.

[0038] However, in actual use, when tracking multiple objects at the same time, the semantic segmentation model is often prone to label jumps and other erroneous segmentation problems. Therefore, this method first tracks and segments a single dynamic object to obtain its mask, then segments the remaining categories of objects to obtain the masks corresponding to all dynamic objects, and then fuses the masks of all dynamic objects in each frame to obtain the masks of all dynamic objects in each frame image, and then obtains multiple dynamic object masks, thereby improving the accuracy and robustness of segmentation. Figure 3 As shown in the results, part (a) is the image to be segmented, and part (b) is the segmentation result: the balloon is segmented first, then the human body is segmented, and the masks in each frame are fused to obtain the final accurate mask containing all dynamic objects.

[0039] Step S3, feature extraction and dynamic elimination: extract point features and line features from the preprocessed image sequence, and use multiple dynamic object masks to eliminate the corresponding dynamic features to obtain static features.

[0040] The corresponding dynamic features are eliminated using the multi-dynamic object mask, that is, the corresponding dynamic area is determined according to the multi-dynamic object mask, and the features in the corresponding dynamic area are deleted according to the dynamic mask, so as to filter the points and lines in the static area.

[0041] Optionally, this step uses the ORB detection method to extract point features, that is, uses the ORB feature points to obtain point features, and uses the LSD line detection method to extract line features, that is, uses the LSD line features to extract line features of the scene. After extracting the point features and line features, the features that fall in the multi-object dynamic mask area are removed, that is, the external points are deleted, that is, the Mask label of the pixel position corresponding to each feature is determined. If the pixel belongs to a dynamic object, the feature is removed, thereby retaining the static features.

[0042] Steps S2 to S3 provide the following ideas for solving the SLAM problem of complex scenes containing multiple types of dynamic objects and accompanied by weak texture areas: First, use the semantic segmentation network to identify and detect multiple types of dynamic objects in the scene, and remove dynamic feature points in the subsequent tracking stage (including the pose estimation process); second, in addition to point features, for weak texture areas (such as white walls, etc.), introduce more robust and easily recognizable line features to extract more complete geometric information contained in the scene. Based on this, by redesigning the error model of the SLAM method, more constraints are placed on pose estimation during the iterative optimization process.

[0043] Step S4, pose estimation: Based on point features and line features, feature matching is performed on static features, the transformation matrix is ​​solved, and the estimated pose is obtained by optimizing the reprojection error function of the joint point and line.

[0044] Specifically, the obtained point features and line features are used to match the historical frames, and the key frame is selected after finding the best match. The corresponding features are triangulated to solve the transformation matrix, and the factor graph is constructed by optimizing the common view point features, line features and key frames. Among them, the key frame is selected according to the current frame information, which includes the number of matches between point features and line features and the feature tracking status.

[0045] The estimated pose is obtained by optimizing the joint reprojection error function of points and lines, which specifically includes: first obtaining the error function of the camera pose, improving the error function by combining point features and line features, obtaining the joint reprojection error function of points and lines, and obtaining the estimated pose by optimizing the joint error function. The joint reprojection error function of points and lines can be expressed as:

[0046]

[0047] In the formula, the kth key frame pose is T cw,k = [R, t], the observed i-th spatial point p I The reprojection error is Ep k,i , the reprojection error of the jth spatial line observed in the kth key frame is El k,i , and Represent different observation covariances, ρp and ρ l They represent different robust kernel functions respectively.

[0048] The feature matching in this step refers to matching the 2D and 3D features of the retained static features, estimating the transformation matrix, and initializing the hyper-primitive map through triangulation. The hyper-primitive map contains the point cloud, rotation, scale, density, and spherical harmonic coefficients of the ORB features, and the tracking process is optimized through the factor graph solver.

[0049] Step S5, geometric mapping: obtain sparse super-primitives, identify the area to be densified according to the reprojection error of the points, and generate a new set of super-primitives in combination with unmatched feature points.

[0050] According to the definition of super primitives in Photo-SLAM, it is composed of a Body density Spherical harmonic coefficients And a set of point clouds with ORB feature information. The reprojection error function of the points is used to determine which areas need to be densified.

[0051] A sparse hyper-primitive map, consisting of a collection of point clouds storing ORB features, rotation, scale, density, and spherical harmonic coefficients, identifies areas that need to be densified based on the reprojection error of the points.

[0052] On this basis, for 2D feature points that cannot match corresponding 3D points, the depth of the surrounding neighboring feature points is used to infer their depth, which is projected into 3D space to generate a new set of super primitives. Sparse super primitives are created using inactive 2D feature points. Specifically, the nearest activated point among the inactive 2D feature points is found, the depth of the activated point is interpolated to estimate the depth of the inactive point, and the 2D features are projected into 3D space to create super primitives.

[0053] Step S6, scene mapping: obtain the rendered image, perform multi-scale Gaussian pyramid supervised training on the rendered image, and obtain the final global map. Specifically including:

[0054] First, the input image is Gaussian smoothed and downsampled to obtain a rendered image. Second, a pyramid is constructed to perform supervision at different levels of image resolution. Finally, photometric loss and SSIM loss are constructed to train and optimize the representation of the rendered image to obtain the final global map representation.

[0055] Among them, the photometric loss and SSIM loss can be expressed as:

[0056]

[0057] In the formula, represents the photometric loss and SSIM loss, λ is the preset weight parameter, I r is the rendered image, I gt is the true value of the input image. During training, the hyper-primitive map is optimized by minimizing the loss function.

[0058] This step uses neural rendering to complete the mapping process, completing the mapping from 3D spatial features to 2D plane pixels, obtaining a spatial model close to the real scene representation and an image representation of a new perspective in 2D space, which can effectively enrich the map representation information.

[0059] Step S7, loop detection: Use the preset bag-of-words model to perform loop detection. When a loop is detected, use similar changes to adjust the current frame and super primitives to correct the posture and correct the accumulated error.

[0060] This step aims to further optimize the map representation and reduce ghosting and blurring in the rendered results.

[0061] Next, the algorithm is verified on multiple data sets to prove the effectiveness and practicality of this method. Specifically:

[0062] As shown in Table 1, on the TUM and BONN datasets, scene sequences such as freiburg3_walking_xyz (fr3_w_xyz) are selected for algorithm testing, and the results of absolute trajectory error (ATE) are recorded for comparison with existing methods Dyna-SLAM, DS-SLAM, and ORB-SLAM3. By comparison, it can be seen that compared with the existing positioning algorithms, the ATE index of this method is lower, and the positioning accuracy is greatly improved.

[0063] In addition, in the evaluation of the mapping results, the peak signal-to-noise ratio (PSNR) is calculated by statistically analyzing static pixels to reflect the reconstruction quality. Figure 4 As shown, Figure 4 Part (a) is the mapping result of the DDN-SLAM method. Figure 4 Part (b) is the mapping result of the ESLAM method. Figure 4 Part (c) is the mapping result of the LC_CRFSLAM method. Figure 4 Part (d) is the mapping result of the Orbeez-SLAM method. Figure 4 Part (e) of the figure is the mapping result of this method. By comparison, it can be seen that this method has a significant improvement in mapping quality compared to existing mapping algorithms.

[0064] Table 1 ATE comparison results of various SLAM methods on TUM and BONN datasets

[0065] Sequence Dyna DS ORB3 This method TUM fr3_w_xyz 0.0214 0.0248 0.4622 0.0169 fr3_w_static 0.0092 0.0087 0.3243 0.0069 fr3_w_rpy 0.0353 0.4426 0.1683 0.0310 fr3_w_halfsphere 0.0262 0.0343 0.6531 0.0313 fr3_s_xyz 0.0152 0.0017 0.0248 0.0134 fr3_s_static 0.0067 0.0065 0.0146 0.0066 BONN ballon 0.0311 0.0323 0.0921 0.0292 balloon2 0.0297 0.0388 0.2163 0.0299 ballon_tracking 0.0326 0.0341 0.1811 0.0313 crowd 0.0183 0.3436 0.4742 0.0181 crowd2 0.0199 0.1784 0.6617 0.0184 crowd3 0.0254 0.1622 0.4536 0.0311 synchronous 0.0122 0.1136 0.6178 0.0092

[0066] Furthermore, the present embodiment also provides a monocular neural rendering SLAM system in a complex environment, including a data acquisition and preprocessing module, a dynamic object mask acquisition module, a feature extraction and dynamic culling module, a pose estimation module, a geometric mapping module, and a scene mapping module. Among them: the data acquisition and preprocessing module is used to acquire and preprocess the initial image sequence in a complex environment; the dynamic object mask acquisition module is used to input the preprocessed image sequence into a pre-trained semantic segmentation model to obtain a multi-dynamic object mask; the feature extraction and dynamic culling module is used to extract point features and line features from the preprocessed image sequence, and use the multi-dynamic object mask to eliminate the corresponding dynamic features to obtain static features; the pose estimation module is used to match the static features based on the point features and line features, solve the transformation matrix, and obtain the estimated pose by optimizing the reprojection error function of the point-line joint; the geometric mapping module is used to acquire sparse super primitives, identify the area to be densified according to the reprojection error of the point, and generate a new super primitive set in combination with the unmatched feature points; the scene mapping module is used to acquire the rendered image, and perform multi-scale Gaussian pyramid supervised training on the rendered image to obtain the final global map. The specific execution steps of each module are basically the same as the execution steps of the aforementioned method and will not be repeated here.

[0067] In summary, the present invention proposes a monocular neural rendering SLAM method and system in a complex environment. The present invention utilizes point-line feature construction, and proposes methodological improvements based on existing dynamic SLAM and neural rendering SLAM systems, which can improve the robustness of the SLAM system in complex environments such as dynamic and weak textures, and the neural rendering method can provide rendered images close to real scenes from a new perspective, enriching the representation of the SLAM map. In addition, the present method effectively avoids the problem of misidentification and missed identification when the existing image segmentation model simultaneously segments multiple objects by utilizing the fusion of a single object mask. At the same time, the present method constructs a relationship diagram of the fusion of point-line features, constructs a joint error function to solve the rotation matrix and translation vector of the camera, and introduces line features to improve the problem of insufficient number of point features under weak texture conditions.

[0068] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A monocular neural rendering SLAM method in complex environments, characterized in that: The following steps are involved: Data acquisition and preprocessing steps: Acquire the initial image sequence in the complex environment and perform preprocessing; Dynamic object mask acquisition step: input the preprocessed image sequence into the pre-trained semantic segmentation model to obtain multiple dynamic object masks; Feature extraction and dynamic elimination step: extracting point features and line features from the preprocessed image sequence, and using the multi-dynamic object mask to eliminate the corresponding dynamic features to obtain static features; Pose estimation step: based on the point features and the line features, feature matching is performed on the static features, a transformation matrix is ​​solved, and an estimated pose is obtained by optimizing a reprojection error function of a point-line combination; Geometric mapping step: obtain sparse super-primitives, identify the area to be densified based on the reprojection error of the points, and generate a new set of super-primitives based on the unmatched feature points; Scene mapping step: obtain a rendered image, perform multi-scale Gaussian pyramid supervised training on the rendered image, and obtain the final global map.

2. The monocular neural rendering SLAM method in a complex environment according to claim 1, characterized in that: The method also includes a loop detection step: loop detection is performed using a preset bag-of-words model, and when a loop is detected, similar changes are used to adjust the current frame and super primitives to correct the posture and correct the accumulated error.

3. The monocular neural rendering SLAM method in a complex environment according to claim 1, characterized in that: The specific implementation process of the dynamic object mask acquisition step includes: Track and segment each dynamic object in turn to obtain the masks corresponding to all dynamic objects; All masks are fused on each frame to obtain masks of all dynamic objects in each frame.

4. The monocular neural rendering SLAM method in a complex environment according to claim 1, characterized in that: In the feature extraction and dynamic elimination steps, the point features are extracted using the ORB detection method, and the line features are extracted using the LSD line detection method.

5. The monocular neural rendering SLAM method in a complex environment according to claim 1, characterized in that: In the pose estimation step, triangulation is performed through feature matching to complete the initialization of the hyper-primitive map, key frames are selected according to the current frame information, and a factor graph is constructed by optimizing the common view point features, line features and key frames. The factor graph is used to optimize the tracking process.

6. The monocular neural rendering SLAM method in a complex environment according to claim 5, characterized in that: The current frame information includes the matching number of point features and line features and the feature tracking status.

7. The monocular neural rendering SLAM method in a complex environment according to claim 1, characterized in that: The reprojection error function of the point-line joint is expressed as: In the formula, the kth key frame pose is T cw,k = [R, t], the observed i-th spatial point p I The reprojection error is Ep k,i , the reprojection error of the jth spatial line observed in the kth key frame is El k,i , and Represent different observation covariances, ρ p and ρ l They represent different robust kernel functions respectively.

8. The monocular neural rendering SLAM method in a complex environment according to claim 1, characterized in that: The scene mapping step specifically includes: Perform Gaussian smoothing and downsampling on the input image to obtain a rendered image; Build a pyramid to provide supervision at different levels of image resolution; Photometric loss and SSIM loss are constructed to train and optimize the representation of the rendered image.

9. The monocular neural rendering SLAM method in a complex environment according to claim 8, characterized in that: The photometric loss and SSIM loss are expressed as: In the formula, represents the photometric loss and SSIM loss, λ is the preset weight parameter, I r is the rendered image, I gt is the true value of the input image.

10. A monocular neural rendering SLAM system in a complex environment, characterized in that: include: Data acquisition and preprocessing module, used to acquire the initial image sequence in complex environment and perform preprocessing; The dynamic object mask acquisition module is used to input the preprocessed image sequence into the pre-trained semantic segmentation model to obtain multiple dynamic object masks; A feature extraction and dynamic elimination module is used to extract point features and line features from the preprocessed image sequence, and eliminate the corresponding dynamic features using the multi-dynamic object mask to obtain static features; A pose estimation module, for performing feature matching on the static features based on the point features and the line features, solving a transformation matrix, and obtaining an estimated pose by optimizing a reprojection error function of a point-line union; The geometric mapping module is used to obtain sparse super-primitives, identify the area to be densified based on the reprojection error of the points, and generate a new set of super-primitives in combination with unmatched feature points; And a scene mapping module is used to obtain a rendered image, perform multi-scale Gaussian pyramid supervised training on the rendered image, and obtain a final global map.

Images