Robot vision SLAM method and system based on multi-scale residual network

Abstract

The invention discloses a robot vision SLAM method based on a multi-scale residual network, and the method comprises the following steps: obtaining a working environment image of a robot, carrying out the equal cutting, and drawing a brightness gray-scale map; the working environment image is an RGB image; based on the brightness grey-scale map, identifying a high-brightness image and a low-brightness image in the working environment image; enhancing the low-brightness image by using a multi-scale residual network to obtain an enhanced low-brightness image; and based on the high-brightness image and the enhanced low-brightness image, performing tracking, local mapping and closed-loop detection in sequence to obtain an optimized robot pose and global map. According to the method, images under different illumination conditions are intelligently processed, the positioning and map building capability of the robot in a complex environment is remarkably improved, and the method is particularly suitable for various application scenes such as indoor navigation, unmanned warehouse management and search and rescue.

Description

Technical Field

[0001] The present invention relates to the field of intelligent robots, and more specifically to a robot vision SLAM method and system based on a multi-scale residual network. Background Art

[0002] With the development of robotics technology, more and more robots are being used in various scenarios, such as smart homes, industrial automation, logistics distribution, and emergency rescue. In these application scenarios, robots need to have the ability to navigate autonomously, that is, to perform self-localization and mapping (SLAM) in unknown environments. Visual SLAM technology plays a vital role in robot navigation due to its high cost-effectiveness and rich information.

[0003] However, in practical applications, robots often encounter environments with extremely variable lighting conditions, such as insufficient indoor lighting or strong outdoor light. These changes in lighting conditions will have a serious impact on the performance of the visual SLAM system, especially in low-light environments, where the accuracy of feature matching will decrease due to problems such as reduced image contrast and difficulty in detecting feature points, thus affecting the robot's positioning accuracy and map construction quality.

[0004] Existing visual SLAM methods perform poorly when dealing with environments with large lighting changes, especially in low-light conditions, where it is difficult to accurately detect and match feature points, resulting in increased positioning errors. Changes in lighting conditions can reduce image contrast, making it difficult to accurately detect and match feature points, which in turn affects the robustness and accuracy of the SLAM system.

[0005] At the same time, a large amount of computing resources are required to process image data, especially when processing images under different lighting conditions. Due to feature matching problems caused by changing lighting conditions, there are limitations in positioning accuracy, especially in complex and changing environments.

[0006] Therefore, how to design a robot vision SLAM method based on a multi-scale residual network that can effectively solve the image processing problem under low-light conditions and improve the robot's positioning accuracy and map construction quality in complex lighting environments is an urgent problem that technical personnel in this field need to solve. Summary of the invention

[0007] In view of this, the present invention provides a robot vision SLAM method based on a multi-scale residual network, which can better adapt to positioning and map construction in a working environment with uneven lighting or dim light by identifying and enhancing low-brightness images, and uses a multi-scale residual network to enhance the image, thereby improving the clarity of image details and maintaining good performance even in poor lighting conditions.

[0008] In order to achieve the above object, the present invention adopts the following technical solution:

[0009] In a first aspect, the present invention provides a robot vision SLAM method based on a multi-scale residual network, comprising the following steps:

[0010] S1. Obtain the robot's working environment image and cut it into equal parts to draw a brightness grayscale image; the working environment image is an RGB image;

[0011] S2. Based on the brightness grayscale image, identify high brightness images and low brightness images in the working environment image;

[0012] S3, using a multi-scale residual network to enhance the low-brightness image to obtain an enhanced low-brightness image; the multi-scale residual network includes: multiple convolutional layers conv, multiple dual attention units DAU, multiple selective weighted kernel units SWK and multiple selective kernel feature fusion units SKFF;

[0013] S4. Based on the high-brightness image and the enhanced low-brightness image, tracking, local mapping and closed-loop detection are performed in sequence to obtain an optimized robot posture and a global map.

[0014] Preferably, in S1, drawing a brightness grayscale map includes:

[0015] Each RGB image after equal cutting is converted into a grayscale image using a linear weighted average algorithm:

[0016] gary value =a1r+a2g+a3b

[0017] Among them, Gary value represents the gray value after conversion, r, g, b represent the red, green and blue pixel values ​​respectively, and a1, a2, a3 represent the weight coefficients.

[0018] Preferably, said S2 comprises:

[0019] Set the gray threshold λ1, when the gray value of a pixel is gary value When it is lower than the threshold λ1, it is defined as a darker pixel, and the proportion of darker pixels in each complete image is calculated;

[0020] A threshold λ2 is set. When the proportion of darker pixels in a complete image is higher than the threshold λ2, the complete image is defined as an initial low-brightness image, and the proportion of the initial low-brightness image in each equally divided image is calculated.

[0021] A threshold λ3 is set. When the proportion of the initial low-brightness image in a certain equally-divided image is higher than the threshold λ3, the equally-divided image is defined as a low-brightness image.

[0022] Preferably, the S3 comprises:

[0023] Based on the low brightness image Y, the convolutional layer conv is used to extract the shallow features y;

[0024] The shallow feature y is downsampled by ×1, ×2, and ×4 to obtain resolutions of h*w*c, Data flow;

[0025] The dual attention unit DAU is used to extract features from data streams at different resolutions;

[0026] The resolution is 10000000 by selective weighted kernel unit SWK. and resolution is The features of are fused to obtain feature h1, and the resolution is h*w*c and the resolution is The features of are fused to obtain feature h2;

[0027] The features h1 and h2 are input into the dual attention unit DAU and combined with the resolution The features are fused using the selective kernel feature fusion unit SKFF to obtain feature h3;

[0028] Input feature h3 into the convolutional layer conv and combine it with the shallow feature y to obtain the deep feature y d ;

[0029] The depth feature y d Input the convolution layer conv to obtain the residual image Combined with the low-brightness image Y, the enhanced low-brightness image is obtained

[0030] Preferably, the dual attention unit DAU is combined with a channel attention mechanism, expressed as:

[0031]

[0032] Among them, M c (f c ) represents the channel attention feature map, σ(·) represents the Sigmoid gating function, W1 and W2 are downsampling channels, represents the feature vector generated by global average pooling of shallow features, f c and They represent input features and output features respectively.

[0033] Preferably, the dual attention unit DAU is also combined with a spatial attention mechanism, which is expressed as:

[0034]

[0035] Among them, M s (f s ) represents the spatial attention feature map, F 5×5 represents a convolution operation with a filter size of 5×5, Represents the feature vector generated by global average pooling of shallow features, represents the feature vector generated by global maximum pooling of shallow features, f s and They represent input features and output features respectively.

[0036] Preferably, the selective weighted kernel unit SWK combines the self-attention mechanism to perform feature fusion at two different resolutions, and the fusion process is expressed as:

[0037]

[0038] in, represents the output feature map, Representation feature map The soft attention vector of Representation feature map The soft attention vector of represents the nth row vector of feature map L1, Represents the nth row vector of feature map L2, and z is the scaling factor.

[0039] Preferably, the selective kernel feature fusion unit SKFF performs feature fusion at three different resolutions, and the fusion process is expressed as:

[0040]

[0041] in, represents the output feature map, Representation feature map The soft attention vector of Representation feature map The soft attention vector of Representation feature map The soft attention vector of represents the nth row vector of feature map L3, represents the nth row vector of feature map L4, Represents the nth row vector of feature map L5.

[0042] Preferably, the multi-scale residual network is optimized using the Charbonnier loss function, and the optimization process is expressed as:

[0043]

[0044] in, represents the high brightness image and the enhanced low brightness image, X represents the real image, and ε is a constant.

[0045] In a second aspect, the present invention provides a robot vision SLAM system based on a multi-scale residual network, comprising:

[0046] Data acquisition and preprocessing module: used to acquire the robot's working environment image and cut it into equal parts, and draw a brightness grayscale image; the working environment image is an RGB image;

[0047] Image recognition module: used for identifying high-brightness images and low-brightness images in the working environment image based on the brightness grayscale image;

[0048] Image enhancement module: used for enhancing the low-brightness image by using a multi-scale residual network to obtain an enhanced low-brightness image;

[0049] The pose estimation and map construction module is used to perform tracking, local mapping and closed-loop detection in sequence based on the high-brightness image and the enhanced low-brightness image to obtain the optimized robot pose and global map.

[0050] It can be seen from the above technical solution that compared with the prior art, the technical solution of the present invention has the following advantages:

[0051] Beneficial effects:

[0052] 1. By converting RGB images into grayscale images, the computational complexity can be effectively reduced while retaining the basic structural information of the image. It can better reflect the brightness distribution of the image.

[0053] 2. By setting different thresholds, high-brightness and low-brightness areas can be automatically identified according to the actual scene, which helps to carry out targeted processing of images under different brightness conditions in subsequent processing. It can also cope with uneven lighting conditions well, thereby improving the robustness of the SLAM system under different lighting conditions.

[0054] 3. Using a multi-scale residual network, features can be extracted from images of different resolutions. The dual attention unit (DAU) enhances the network's attention to important features through channel attention and spatial attention, thereby improving the quality of feature expression. Feature fusion through the selective weighted kernel unit (SWK) and the selective kernel feature fusion unit (SKFF) can more efficiently utilize multi-scale features.

[0055] 4. By optimizing image quality, feature matching and map construction can be performed more accurately, thereby improving the accuracy of robot positioning and map construction. The Charbonnier loss function is more resistant to noise than the traditional mean square error loss function, which helps to converge to the optimal solution faster. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0057] Figure 1 A flow chart of a robot vision SLAM method based on a multi-scale residual network provided in an embodiment of the present invention;

[0058] Figure 2 A schematic diagram of a process of enhancing the low-brightness image using a multi-scale residual network provided in an embodiment of the present invention;

[0059] Figure 3 A schematic diagram of a low-brightness image provided by an embodiment of the present invention;

[0060] Figure 4 A schematic diagram of a low-brightness image after enhancement processing provided by an embodiment of the present invention;

[0061] Figure 5 A framework diagram of a robot vision SLAM system based on a multi-scale residual network provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0062] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0063] Embodiment 1;

[0064] like Figure 1 As shown, this embodiment provides a robot vision SLAM method based on a multi-scale residual network, comprising the following steps:

[0065] S1. Obtain the robot's working environment image and cut it into equal parts to draw a brightness grayscale image; the working environment image is an RGB image;

[0066] S2. Based on the brightness grayscale image, identify high brightness images and low brightness images in the working environment image;

[0067] S3, using a multi-scale residual network to enhance the low-brightness image to obtain an enhanced low-brightness image; the multi-scale residual network includes: multiple convolutional layers conv, multiple dual attention units DAU, multiple selective weighted kernel units SWK and multiple selective kernel feature fusion units SKFF;

[0068] S4. Based on the high-brightness image and the enhanced low-brightness image, tracking, local mapping and closed-loop detection are performed in sequence to obtain an optimized robot posture and a global map.

[0069] This method divides the RGB image of the robot's working environment into equal parts and draws a brightness grayscale map, thereby identifying high-brightness and low-brightness images, enhancing the low-brightness image using a multi-scale residual network, and combining it with the high-brightness image for tracking, local mapping, and closed-loop detection, thereby optimizing the robot's pose estimation and global map construction.

[0070] It significantly improves the robot's positioning and mapping capabilities in complex environments, and is particularly suitable for various application scenarios such as indoor navigation, unmanned warehouse management, and search and rescue.

[0071] The above steps and related technical features are further described in detail below:

[0072] In the present embodiment S1, the image captured by the visual sensor may have a part with darker brightness. It is impossible to accurately judge whether the image has a part with darker brightness directly based on the grayscale image of the brightness of the entire image.

[0073] Based on this, the robot's working environment image is obtained and cut equally, and the image is divided into multiple sub-areas. This can be implemented by an image processing library in a programming language, such as Python's OpenCV or PIL (Pillow). And determine the number and shape of the segmentation (for example, divide the image into N×N blocks).

[0074] Draw a brightness grayscale image, including:

[0075] Each RGB image after equal cutting is converted into a grayscale image using a linear weighted average algorithm:

[0076] gary value =a1r+a2g+a3b

[0077] Among them, Gary valuerepresents the gray value after conversion, r, g, b represent the red, green and blue pixel values ​​respectively, and a1, a2, a3 represent the weight coefficients.

[0078] This step clarifies the method of converting the segmented RGB image into a grayscale image. By using a linear weighted average algorithm, the key information of the image can be effectively retained, while simplifying the subsequent processing flow and improving computational efficiency.

[0079] This embodiment S2 includes:

[0080] Set the gray threshold λ1, when the gray value of a pixel is gary value When it is lower than the threshold λ1, it is defined as a darker pixel, and the proportion of darker pixels in each complete image is calculated;

[0081] A threshold λ2 is set. When the proportion of darker pixels in a complete image is higher than the threshold λ2, the complete image is defined as an initial low-brightness image, and the proportion of the initial low-brightness image in each equally divided image is calculated.

[0082] A threshold λ3 is set. When the proportion of the initial low-brightness image in a certain equally-divided image is higher than the threshold λ3, the equally-divided image is defined as a low-brightness image.

[0083] In this step, the grayscale threshold is set to identify darker pixels, and the proportion of these darker pixels in the entire image is further calculated to distinguish low-brightness images. It provides an objective and quantifiable standard to automatically identify and filter low-brightness images without manual intervention. By gradually analyzing the brightness distribution from a single pixel to the entire image, the image parts that need to be enhanced can be effectively located, thereby ensuring that the multi-scale residual network can enhance these low-brightness images in a targeted manner and improve the robustness and accuracy of the overall visual SLAM system.

[0084] In addition, the low-brightness images screened out in this way can reduce unnecessary consumption of computing resources and improve the processing efficiency and real-time performance of the system.

[0085] like Figure 2 As shown, this embodiment S3 includes:

[0086] like Figure 3 As shown, based on the low brightness image Y, the convolution layer conv is used to extract the shallow feature y;

[0087] The shallow feature y is downsampled by ×1, ×2, and ×4 to obtain resolutions of h*w*c, Data flow;

[0088] The dual attention unit DAU is used to extract features from data streams at different resolutions;

[0089] The resolution is 10000000 by selective weighted kernel unit SWK. and resolution is The features of are fused to obtain feature h1, and the resolution is h*w*c and the resolution is The features of are fused to obtain feature h2;

[0090] The features h1 and h2 are input into the dual attention unit DAU and combined with the resolution The features are fused using the selective kernel feature fusion unit SKFF to obtain feature h3;

[0091] Input feature h3 into the convolutional layer conv and combine it with the shallow feature y to obtain the deep feature y d ;

[0092] like Figure 4 As shown, the depth feature y d Input the convolution layer conv to obtain the residual image Combined with the low-brightness image Y, the enhanced low-brightness image is obtained

[0093] In this step, the dual attention unit DAU is combined with the channel attention mechanism, which can dynamically adjust the attention distribution according to different channel characteristics, so that the network can pay more attention to important feature information, expressed as:

[0094]

[0095] Among them, M c (f c ) represents the channel attention feature map, σ(·) represents the Sigmoid gating function, W1 and W2 are downsampling channels, represents the feature vector generated by global average pooling of shallow features, f c and They represent input features and output features respectively.

[0096] The dual attention unit DAU is also combined with a spatial attention mechanism, which can further help the network focus on the key areas in the image, improve the accuracy of feature extraction, and thus enhance the model's ability to understand local details, expressed as:

[0097]

[0098] Among them, M s (f s ) represents the spatial attention feature map, F 5×5represents a convolution operation with a filter size of 5×5, Represents the feature vector generated by global average pooling of shallow features, represents the feature vector generated by global maximum pooling of shallow features, f s and They represent input features and output features respectively.

[0099] The selective weighted kernel unit SWK combines the self-attention mechanism to fuse features at two different resolutions, which can better capture the multi-level information in the image and enhance the model's comprehensive processing ability for features of different scales. The fusion process is expressed as:

[0100]

[0101] in, represents the output feature map, Representation feature map The soft attention vector of Representation feature map The soft attention vector of represents the nth row vector of feature map L1, Represents the nth row vector of feature map L2, and z is the scaling factor.

[0102] The selective kernel feature fusion unit SKFF performs feature fusion at three different resolutions. The multi-scale fusion method is beneficial to improving the effect of image enhancement. The fusion process is expressed as:

[0103]

[0104] in, represents the output feature map, Representation feature map The soft attention vector of Representation feature map The soft attention vector of Representation feature map The soft attention vector of represents the nth row vector of feature map L3, represents the nth row vector of feature map L4, Represents the nth row vector of feature map L5.

[0105] In this step, a multi-scale residual network architecture is used to enhance low-brightness images. First, shallow features are extracted through convolutional layers, and then data streams of different resolutions are used to capture multi-level information of the image. Through the dual attention unit (DAU) combined with the channel attention and spatial attention mechanisms, the network can more accurately focus on key areas and suppress unimportant background information. The channel attention mechanism captures global context information through global average pooling, while the spatial attention mechanism focuses on specific spatial locations through convolution operations. The selective weighted kernel unit (SWK) is combined with the self-attention mechanism to achieve effective fusion of features of different resolutions and enhance the correlation between features. Finally, the selective kernel feature fusion unit (SKFF) further integrates information of different scales, making the final output feature map richer and more hierarchical.

[0106] The method of combining multi-scale and attention mechanisms in the multi-scale residual network not only improves the generalization ability of the model, but also enhances the restoration effect of low-brightness image details, thereby significantly improving the quality and efficiency of image enhancement.

[0107] In this embodiment, the multi-scale residual network is optimized using the Charbonnier loss function, and the optimization process is expressed as:

[0108]

[0109] in, represents the high brightness image and the enhanced low brightness image, X represents the real image, and ε is a constant.

[0110] It has better numerical stability, can train the network more efficiently, and ensures the accuracy and robustness of the model when processing low-brightness images.

[0111] In this embodiment S4, based on the high-brightness image and the enhanced low-brightness image, tracking, local mapping and closed-loop detection are performed in sequence to obtain an optimized robot posture and a global map.

[0112] In this step, image tracking is performed, including: using a robust feature detector to extract features; using feature descriptors to perform feature matching, and combining the information of the image brightness grayscale map to improve the matching accuracy; using an optical flow algorithm to estimate the motion between consecutive frames; estimating the relative position change of the camera through matching point pairs; using the RANSAC algorithm to eliminate erroneous matching points; and adjusting exposure parameters according to changes in image brightness to improve the matching success rate.

[0113] Local mapping is performed based on the tracking results, including: adding new keyframes when the camera pose changes significantly; using a multi-scale residual network to assist in determining which frames contain enough new information; using triangulation methods to construct a 3D point cloud from feature matching between keyframes; and applying pose graph optimization techniques, such as bundle adjustment-based optimization, to reduce cumulative errors.

[0114] Perform loop closure detection, including: using the bag-of-words model to detect loop closure candidates; using the outputs of the multi-scale residual network at different scales to evaluate the possibility of loop closure; using the graph optimization method to update the camera pose and map. It also includes adding the output of the multi-scale residual network as an additional constraint to the graph optimization process of loop closure detection.

[0115] Finally, the global map is constructed: all local maps are merged into a global map; the information obtained from closed-loop detection is used to optimize the global map; the map is visualized in real time, and nonlinear optimization technology is used to further improve the map accuracy.

[0116] Embodiment 2;

[0117] like Figure 5 As shown, this embodiment provides a robot vision SLAM system based on a multi-scale residual network, including:

[0118] Data acquisition and preprocessing module: used to acquire the robot's working environment image and cut it into equal parts, and draw a brightness grayscale image; the working environment image is an RGB image;

[0119] Image recognition module: used for identifying high-brightness images and low-brightness images in the working environment image based on the brightness grayscale image;

[0120] Image enhancement module: used for enhancing the low-brightness image by using a multi-scale residual network to obtain an enhanced low-brightness image;

[0121] The pose estimation and map construction module is used to perform tracking, local mapping and closed-loop detection in sequence based on the high-brightness image and the enhanced low-brightness image to obtain the optimized robot pose and global map.

[0122] The system acquires RGB images through the data acquisition and preprocessing module and converts them into brightness grayscale images, and then uses the image recognition module to identify high-brightness and low-brightness areas; for low-brightness images, the image enhancement module uses a multi-scale residual network for enhancement processing; finally, the pose estimation and map construction module performs tracking, local mapping and closed-loop detection based on the processed images to achieve precise positioning of the robot and construction of a global map of the environment.

[0123] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0124] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A robot vision SLAM method based on a multi-scale residual network, characterized in that: The following steps are involved: S1. Obtain the robot's working environment image and cut it into equal parts to draw a brightness grayscale image; the working environment image is an RGB image; S2. Based on the brightness grayscale image, identify high brightness images and low brightness images in the working environment image; S3, using a multi-scale residual network to enhance the low-brightness image to obtain an enhanced low-brightness image; The multi-scale residual network includes: multiple convolutional layers conv, multiple dual attention units DAU, multiple selective weighted kernel units SWK and multiple selective kernel feature fusion units SKFF; S4. Based on the high-brightness image and the enhanced low-brightness image, tracking, local mapping and closed-loop detection are performed in sequence to obtain an optimized robot posture and a global map.

2. A robot vision SLAM method based on a multi-scale residual network according to claim 1, characterized in that: In S1, drawing a brightness grayscale image includes: Each RGB image after equal cutting is converted into a grayscale image using a linear weighted average algorithm: gary value =a1r+a2g+a3b Among them, Gary value represents the gray value after conversion, r, g, b represent the red, green and blue pixel values ​​respectively, and a1, a2, a3 represent the weight coefficients.

3. A robot vision SLAM method based on a multi-scale residual network according to claim 1, characterized in that: The S2 comprises: Set the grayscale threshold λ1, when the grayscale value of a pixel is gary value When it is lower than the threshold λ1, it is defined as a darker pixel, and the proportion of darker pixels in each complete image is calculated; A threshold λ2 is set. When the proportion of darker pixels in a complete image is higher than the threshold λ2, the complete image is defined as an initial low-brightness image, and the proportion of the initial low-brightness image in each equally divided image is calculated. A threshold λ3 is set. When the proportion of the initial low-brightness image in a certain equally-divided image is higher than the threshold λ3, the equally-divided image is defined as a low-brightness image.

4. A robot vision SLAM method based on a multi-scale residual network according to claim 1, characterized in that: The S3 includes: Based on the low brightness image Y, the convolutional layer conv is used to extract the shallow features y; The shallow feature y is downsampled by ×1, ×2, and ×4 to obtain resolutions of h*w*c, Data flow; The dual attention unit DAU is used to extract features from data streams at different resolutions; The resolution is 10000000 by selective weighted kernel unit SWK. and resolution is The features of are fused to obtain feature h1, and the resolution is h*w*c and the resolution is The features of are fused to obtain feature h2; The features h1 and h2 are input into the dual attention unit DAU and combined with the resolution The features are fused using the selective kernel feature fusion unit SKFF to obtain feature h3; Input feature h3 into the convolutional layer conv and combine it with the shallow feature y to obtain the deep feature y d ; The depth feature y d Input the convolution layer conv to obtain the residual image Combined with the low-brightness image Y, the enhanced low-brightness image is obtained 5. A robot vision SLAM method based on a multi-scale residual network according to claim 4, characterized in that: The dual attention unit DAU is combined with a channel attention mechanism, expressed as: Among them, M c (f c ) represents the channel attention feature map, σ(·) represents the Sigmoid gating function, W1 and W2 are downsampling channels, represents the feature vector generated by global average pooling of shallow features, f c and They represent input features and output features respectively.

6. A robot vision SLAM method based on a multi-scale residual network according to claim 4, characterized in that: The dual attention unit DAU is also combined with a spatial attention mechanism, expressed as: Among them, M s (f s ) represents the spatial attention feature map, F 5×5 represents a convolution operation with a filter size of 5×5, Represents the feature vector generated by global average pooling of shallow features, represents the feature vector generated by global maximum pooling of shallow features, f s and They represent input features and output features respectively.

7. A robot vision SLAM method based on a multi-scale residual network according to claim 4, characterized in that: The selective weighted kernel unit SWK combines the self-attention mechanism to perform feature fusion at two different resolutions. The fusion process is expressed as: in, represents the output feature map, Representation feature map The soft attention vector of Representation feature map The soft attention vector of represents the nth row vector of feature map L1, Represents the nth row vector of feature map L2, and z is the scaling factor.

8. A robot vision SLAM method based on a multi-scale residual network according to claim 4, characterized in that: The selective kernel feature fusion unit SKFF performs feature fusion at three different resolutions, and the fusion process is expressed as follows: in, represents the output feature map, Representation feature map The soft attention vector of Representation feature map The soft attention vector of Representation feature map The soft attention vector of represents the nth row vector of feature map L3, represents the nth row vector of feature map L4, Represents the nth row vector of feature map L5.

9. A robot vision SLAM method based on a multi-scale residual network according to claim 1, characterized in that: The multi-scale residual network is optimized using the Charbonnier loss function, and the optimization process is expressed as: in, represents the high brightness image and the enhanced low brightness image, X represents the real image, and ε is a constant.

10. A robot vision SLAM system based on a multi-scale residual network, characterized in that: include: Data acquisition and preprocessing module: used to acquire the robot's working environment image and cut it into equal parts, and draw a brightness grayscale image; The working environment image is an RGB image; Image recognition module: used for identifying high-brightness images and low-brightness images in the working environment image based on the brightness grayscale image; Image enhancement module: used for enhancing the low-brightness image by using a multi-scale residual network to obtain an enhanced low-brightness image; The pose estimation and map construction module is used to perform tracking, local mapping and closed-loop detection in sequence based on the high-brightness image and the enhanced low-brightness image to obtain the optimized robot pose and global map.

Images