A positioning method and system supporting online learning

Abstract

The application discloses a positioning method and system supporting online learning, relates to the field of computer vision, and utilizes a depth estimation algorithm based on a low-rank adapter in a depth estimation module to estimate a dense depth map of an image, utilizes a self-supervision mechanism to realize online learning of target domain depth estimation, utilizes a dense depth map and pose information in an online learning module to calculate a ranking loss function and an uncertainty loss function to supervise training of a depth estimation network, extracts ORB features and descriptors according to an image in a feature selection module, utilizes estimated uncertainty and the dense depth map to calculate pose information, utilizes alignment of the dense depth map and a sparse depth map in a pose estimation module to calculate scale information of image pose to ensure scale consistency of the system in a positioning process. The application retains a basic structure of the depth estimation network by designing the low-rank adapter, can more quickly adapt to new environmental features, improves sensitivity to environmental changes, and reduces dependence on a large amount of new data.

Description

Technical Field

[0001] This invention relates to the field of computer vision, and more particularly to a localization method and system that supports online learning. Background Technology

[0002] With the rapid development of computer vision technology, visual localization is becoming increasingly widely used in fields such as autonomous driving, robot navigation, and augmented reality. However, traditional visual localization techniques have limitations when facing complex, dynamic, and unknown environments. In particular, while deep learning-based deep estimation networks perform well on training datasets, their performance often declines when encountering new scenarios in real-world applications because these networks often fail to adequately cover all possible scene variations during training.

[0003] Traditional visual localization techniques typically employ pre-trained depth estimation networks, which are trained on fixed datasets and lack adaptability to real-time environmental changes. Because they fail to capture features of new environments in a timely manner, these networks may produce inaccurate depth estimates in unfamiliar scenes, leading to unstable and imprecise localization.

[0004] In attempts to address these issues, some methods have adapted to new environments by fine-tuning the depth estimation network. However, this approach often requires large amounts of newly labeled data, and the fine-tuning process is time-consuming. Therefore, there is an urgent need for a more flexible, real-time, and efficient visual localization technique that can dynamically adapt to changes in the environment during practical applications.

[0005] Therefore, those skilled in the art are dedicated to developing a positioning method and system that supports online learning. Summary of the Invention

[0006] In view of the above-mentioned deficiencies of the prior art, the technical problem to be solved by the present invention is that traditional visual positioning technology usually uses pre-trained depth estimation networks, which are trained on fixed datasets and lack adaptability to real-time environmental changes.

[0007] To achieve the above objectives, the present invention provides a localization system supporting online learning, characterized in that it includes a depth estimation module, an online learning module, a feature selection module, and a pose estimation module, wherein...

[0008] The depth estimation module uses a depth estimation algorithm based on a low-rank adapter to estimate the dense depth map of the RGB image. The depth estimation algorithm uses a self-supervised mechanism to achieve online learning of the target domain depth estimation.

[0009] The online learning module uses the dense depth map and pose information of the RGB image to calculate the ranking loss function and the uncertainty loss function, which are used to supervise the training of the depth estimation network.

[0010] The feature selection module extracts ORB features and descriptors from the RGB image, and selects appropriate feature points to solve the pose information using the estimated uncertainty and the dense depth map.

[0011] The pose estimation module uses the alignment of the dense depth map and the sparse depth map to calculate the scale information of the RGB image pose, ensuring the scale consistency of the system during the localization process.

[0012] Furthermore, the depth estimation module uses the depth estimation network to implement the depth estimation algorithm. The depth estimation network adopts a U-Net encoder-decoder model structure. The encoder is set to a ResNet model pre-trained on the DDAD dataset, and the decoder converts the output into depth values. The low-rank adapter is inserted into all convolutional layers in the encoder and the decoder and is connected in parallel with the convolutional layers. By freezing the pre-training parameters of the depth estimation network, the parameters of the low-rank adapter are learned and adjusted online only. The depth estimation network with the low-rank adapter module is used to perform depth estimation on keyframes of each RGB image to obtain the dense depth map.

[0013] Furthermore, the low-rank adapter structure employs series-attention and a pyramid structure. The low-rank adapter captures new information and knowledge that has not been learned in the image and improves the performance and adaptability of the deep estimation network in unfamiliar environments by adjusting its own learnable parameters.

[0014] Furthermore, the online learning module supervises the learning of the deep estimation network by calculating a loss function, and the pose estimation module optimizes the input image. With images relative pose Then, the photometric loss function, image smoothing loss function, geometric consistency loss function, feature point ranking loss function, and uncertainty loss function required for the depth estimation network to learn are constructed.

[0015] Furthermore, the overall loss of the online learning module is:

[0016] ,

[0017] in, , , and For loss weighting, For luminance loss, For image smoothing loss, For geometric consistency loss, For feature point ranking loss, Losses due to uncertainty;

[0018] The photometric loss function is:

[0019] ,

[0020] The image smoothing loss function is:

[0021] ,

[0022] The geometric consistency loss function is:

[0023]

[0024] The feature point ranking loss function is:

[0025] ,

[0026] ,

[0027]

[0028] The uncertainty loss function is:

[0029] ,

[0030] in, For coefficients, For the input image, To utilize image relative pose and predicted depth value Image synthesized through back projection transformation , , To predict depth values, For the input image and Relative pose, For camera parameters, To measure using the SSIM function and The similarity between them yes The synthesis depth is determined by and Generated through a rigid transformation at the underlying level. To predict depth maps Point pairs in the middle, It is the tolerance threshold. The depth values ​​in the sparse depth map. The true ordinal labels are derived from the sparse depth map, where N is the number of points in the sparse depth map. For image Depth estimation map The dense depth map is obtained by flipping it horizontally again. , To predict depth maps The point in the middle.

[0031] Furthermore, the feature selection module utilizes depth estimation uncertainty information as a trigger flag for feature selection based on dense depth maps. The depth estimation uncertainty is set as the uncertainty result calculated by the depth estimation network based on the RGB image, and the uncertainty result is calculated in the following manner:

[0032] ,

[0033] in, For uncertainty results, RGB image The depth estimate is obtained through a depth estimation network. RGB image flipped horizontally The depth estimate is obtained through a depth estimation network. Depend on The depth estimate obtained by horizontally flipping again, where N is the predicted depth map. The number of points in the graph.

[0034] Furthermore, when the uncertainty result is less than a preset threshold, the feature selection module is triggered to select appropriate feature points using a dense depth map for pose estimation.

[0035] Furthermore, the pose estimation module calculates the scale information of the pose using the dense depth map and the sparse depth map, and outputs the final pose information. The scale information of the pose is calculated in the following manner:

[0036]

[0037] in, For dense depth maps, This is a sparse depth map. The final output pose information , The in-situ pose is scale-free, and N is the sparse depth map. The number of midpoints.

[0038] On the other hand, the present invention provides a localization method that supports online learning. This method uses the online learning-supporting localization system provided by the present invention, employs a depth estimation algorithm based on a low-rank adapter to estimate a dense depth map of an RGB image, extracts ORB features and descriptors from the RGB image, selects appropriate feature points using the estimation uncertainty and the dense depth map to calculate pose information, and calculates a ranking loss function and an uncertainty loss function using the dense depth map and the pose information to supervise the training of the depth estimation network. The depth estimation algorithm utilizes a self-supervised mechanism to achieve online learning of target domain depth estimation.

[0039] Furthermore, the positioning method includes the following steps:

[0040] S101: Using the depth estimation module, estimate the dense depth map of the RGB image;

[0041] S103: The online learning module uses the dense depth map and the pose information to calculate the ranking loss function and the uncertainty loss function to supervise the training of the depth estimation network;

[0042] S105: The feature selection module extracts ORB features and descriptors from the RGB image, and selects appropriate feature points for solving the pose information using the estimated uncertainty and dense depth map;

[0043] S107: The pose estimation module uses the alignment of the dense depth map and the sparse depth map to calculate the scale information of the pose, so as to ensure the scale consistency of the system during the localization process.

[0044] In a preferred embodiment of the present invention, compared with the prior art, the present invention has the following beneficial effects:

[0045] 1. This invention designs a low-rank adapter, enabling the depth estimation network to adapt to the characteristics of new environments more quickly while retaining its basic structure, thereby improving its sensitivity to environmental changes and real-time adaptability.

[0046] 2. This invention enables deep estimation networks to adapt to new environments in real time through low-rank adapters and self-supervised learning mechanisms, reducing reliance on large amounts of new data.

[0047] 3. This invention, through the ingenious design of a low-rank adapter, achieves modulation of the module's parameters, enabling it to quickly adapt to the characteristics of new environments. This flexible modulation mechanism preserves the basic parameters and structure of the deep estimation network while making it more adaptable, allowing it to rapidly capture changes in new environments in practical applications.

[0048] The following will further explain the concept, specific structure, and technical effects of the present invention in conjunction with the accompanying drawings, so as to fully understand the purpose, features, and effects of the present invention. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the system modules according to a preferred embodiment of the present invention;

[0050] Figure 2 This is a schematic diagram of a depth estimation module based on a low-rank adapter according to a preferred embodiment of the present invention;

[0051] Figure 3 This is a schematic diagram of a preferred embodiment of the invention based on a low-rank adapter;

[0052] Figure 4 This is a flowchart illustrating a preferred embodiment of the positioning method of the present invention. Detailed Implementation

[0053] The following description, with reference to the accompanying drawings, illustrates several preferred embodiments of the present invention to make its technical content clearer and easier to understand. The present invention can be embodied in many different forms, and the scope of protection of the present invention is not limited to the embodiments mentioned herein.

[0054] In the accompanying drawings, components with the same structure are indicated by the same numerical designation, and components with similar structures or functions are indicated by similar numerical designations. The dimensions and thicknesses of each component shown in the drawings are arbitrary, and the present invention does not limit the dimensions and thicknesses of each component. To make the illustrations clearer, the thickness of some components has been appropriately exaggerated in the drawings.

[0055] like Figure 1 As shown, the localization system supporting online learning provided in this embodiment of the invention introduces a low-rank adapter and an online learning mechanism to improve the adaptability of the deep estimation network to new environments. The localization system includes a depth estimation module, an online learning module, a feature selection module, and a pose estimation module.

[0056] The depth estimation module uses a low-rank adapter-based depth estimation algorithm to estimate the dense depth map of RGB images. The depth estimation algorithm utilizes a self-supervised mechanism to achieve online learning for target domain depth estimation.

[0057] The online learning module uses the dense depth map and pose information of RGB images to calculate the ranking loss function and uncertainty loss function, which are used to supervise the training of the depth estimation network.

[0058] The feature selection module extracts ORB features and descriptors from the RGB image, and selects appropriate feature points for solving pose information by utilizing the estimated uncertainty and dense depth map.

[0059] The pose estimation module uses the alignment of dense and sparse depth maps to calculate the scale information of the RGB image pose, ensuring the scale consistency of the system during the localization process.

[0060] Specifically, the depth estimation module uses a depth estimation network to implement the depth estimation algorithm. This network employs a U-Net encoder-decoder model structure. The encoder is a ResNet model pre-trained on the DDAD (Dense Depth for Autonomous Driving) dataset, and the decoder converts the output into depth values. A low-rank adapter is inserted into all convolutional layers in both the encoder and decoder, and is connected in parallel with the convolutional layers. By freezing the pre-trained parameters of the depth estimation network, the parameters of the low-rank adapter are learned and adjusted online only. The depth estimation network with the low-rank adapter module performs depth estimation on each frame of the RGB image keyframes, resulting in a dense depth map. The low-rank adapter structure uses series-attention and a pyramid structure. This low-rank adapter captures new, unlearned information and knowledge in the image, and by adjusting its learnable parameters, improves the performance and adaptability of the depth estimation network in unfamiliar environments.

[0061] The online learning module supervises the learning of the deep estimation network by calculating the loss function, and the pose estimation module optimizes the input image. With images relative pose Then, we construct the photometric loss function, image smoothing loss function, geometric consistency loss function, feature point ranking loss function, and uncertainty loss function required for deep estimation network learning.

[0062] The overall loss for this online learning module is:

[0063] ,

[0064] in, , , and For loss weighting, For luminance loss, For image smoothing loss, For geometric consistency loss, For feature point ranking loss, The uncertainty loss is calculated using the following functions: photometric loss, image smoothing loss, geometric consistency loss, feature point ranking loss, and uncertainty loss.

[0065] Specifically, the photometric loss function is:

[0066] ,

[0067] The image smoothing loss function is:

[0068] ,

[0069] The geometric consistency loss function is:

[0070]

[0071] The feature point ranking loss function is:

[0072] ,

[0073] ,

[0074]

[0075] The uncertainty loss function is:

[0076] ,

[0077] in, For coefficients, For the input image, To utilize image relative pose and predicted depth value Image synthesized through back projection transformation , , To predict depth values, For the input image and Relative pose, For camera parameters, To measure using the SSIM function and The similarity between them yes The synthesis depth is determined by and Generated through a rigid transformation at the underlying level. To predict depth maps Point pairs in the middle, It is the tolerance threshold. The depth values ​​in the sparse depth map. The true ordinal labels are derived from the sparse depth map, where N is the number of points in the sparse depth map. For image Depth estimation map The dense depth map is obtained by flipping it horizontally again. , To predict depth maps The point in the middle.

[0078] The feature selection module uses depth estimation uncertainty information as a trigger flag for feature selection based on dense depth maps. The depth estimation uncertainty is set as the uncertainty result calculated by the depth estimation network based on the RGB image. The uncertainty result is calculated in the following way:

[0079] ,

[0080] in, For uncertainty results, RGB image The predicted depth map obtained through the depth estimation network, I' is the RGB image of image I after horizontal flipping. For image The predicted depth map obtained through a depth estimation network Depend on The depth map is obtained by flipping it horizontally again, where N is the predicted depth map. The number of pixels.

[0081] The pose estimation module uses dense and sparse depth maps to calculate the scale information of the pose and outputs the final pose information. The scale information of the pose is calculated in the following way:

[0082]

[0083] in, For dense depth maps, This is a sparse depth map. The final output pose information , The in-situ pose is scale-free, and N is the sparse depth map. The number of midpoints.

[0084] like Figure 4 As shown, this embodiment of the invention also provides a localization method that supports online learning. This localization method uses the localization system that supports online learning as described in this embodiment of the invention. It utilizes a depth estimation algorithm based on a low-rank adapter to estimate a dense depth map of an RGB image, extracts ORB features and descriptors from the RGB image, selects appropriate feature points using the estimated uncertainty and dense depth map to solve for pose information, and calculates a ranking loss function and an uncertainty loss function using the dense depth map and pose information to supervise the training of the depth estimation network. The depth estimation algorithm utilizes a self-supervised mechanism to achieve online learning of target domain depth estimation.

[0085] The positioning method provided in this embodiment of the invention includes the following steps:

[0086] S101: Utilize the depth estimation module to estimate the dense depth map of the RGB image;

[0087] S103: Online learning module, which uses dense depth maps and pose information to calculate ranking loss function and uncertainty loss function to supervise the training of depth estimation network;

[0088] S105: Feature selection module, extracts ORB features and descriptors from RGB image, and selects appropriate feature points for solving pose information using the estimated uncertainty and dense depth map;

[0089] S107: Pose estimation module, which uses the alignment of dense and sparse depth maps to calculate the scale information of the pose to ensure the scale consistency of the system during the localization process.

[0090] Compared with the prior art, the present invention has the following beneficial technical effects:

[0091] 1. Traditional visual positioning techniques typically employ pre-trained depth estimation networks. These networks, trained on fixed datasets, lack adaptability to real-time environmental changes. This invention introduces a newly designed low-rank adapter, enabling the depth estimation network to adapt more quickly to new environmental features while retaining its basic structure, thus improving its sensitivity to environmental changes. The monocular depth estimation network SC_Depth is pre-trained on the KITTI dataset. Multiple low-rank adapter-based LoRA modules are connected in parallel to all convolutional layers of the SC_Depth network. The pre-trained parameters of the SC_Depth network are frozen, and the parameters of all LoRA modules are learned and adjusted online only. The SC_Depth modules with LoRA modules are used to perform depth estimation on keyframes of each RGB image, resulting in a dense depth map. Through the low-rank adapter and self-supervised learning mechanism, the depth estimation network gains the ability to adapt to new environments in real time, reducing its dependence on large amounts of new data.

[0092] 2. This invention, through the ingenious design of a low-rank adapter, achieves modulation of the module's parameters, enabling it to quickly adapt to the characteristics of new environments. This flexible modulation mechanism preserves the basic parameters and structure of the deep estimation network while making it more adaptable, allowing it to rapidly capture changes in new environments in practical applications.

[0093] 3. Some methods fine-tune the depth estimation network to adapt to new environments, but this often requires a large amount of newly labeled data, and the fine-tuning process is time-consuming. This invention introduces an online learning mechanism, enabling the depth estimation network to learn in real-time in practical application scenarios. By dynamically adjusting parameters through real-time data feedback, the network performance is improved, and it adapts to environmental changes. A self-supervised learning mechanism is introduced using a ranking loss function and a depth uncertainty loss function based on sparse depth maps. Real-time pose data and sparse depth map data feedback during system operation enable online learning of the depth estimation module parameters. This invention, by introducing a self-supervised learning mechanism and real-time data feedback, allows the network to learn online during runtime, avoiding the time-consuming offline fine-tuning process of traditional methods. By introducing a self-supervised learning mechanism, the depth estimation network can learn online during runtime through real-time data feedback, achieving real-time adaptation to environmental changes. This not only avoids the time-consuming offline fine-tuning process of traditional methods but also gives the system efficient online learning capabilities, thereby improving the network's performance in new environments.

[0094] The present invention will now be described in detail with reference to preferred embodiments.

[0095] like Figure 1 As shown, a preferred embodiment of the present invention provides a localization system supporting online learning, comprising a depth estimation module, an online learning module, a feature selection module, and a pose estimation module, wherein...

[0096] The depth estimation module is used to obtain a dense depth map for each frame of the image. This module relates to a depth estimation algorithm based on a low-rank adapter proposed in this invention. This algorithm utilizes a self-supervised mechanism to achieve online learning of target domain depth estimation, such as... Figure 2 As shown. The depth estimation network SC_Depth used in this invention is a classic U-Net encoder-decoder model structure, where the encoder is a ResNet model pre-trained on the DDAD dataset, and the decoder converts the output into depth values. A low-rank adapter is inserted into all convolutional layers in both the encoder and decoder, and is connected in parallel with the convolutional layers. The structure of the low-rank adapter is as follows: Figure 3 As shown, the low-rank adapter structure employs series-attention and a pyramid structure. This adapter structure can effectively capture new, unlearned information and knowledge in images, and improve the performance and adaptability of deep estimation networks in unfamiliar environments by adjusting its own learnable parameters.

[0097] The online learning module supervises the learning of the deep estimation network by calculating the loss function, and the pose estimation module optimizes the input image. With images relative pose Then, we construct the photometric loss function, image smoothing loss function, geometric consistency loss function, feature point ranking loss function, and uncertainty loss function required for deep estimation network learning.

[0098] The photometric loss function is expressed as:

[0099]

[0100] in, Represents the use of relative pose in images and predicted depth value The image synthesized through back projection transformation These are the camera's parameters.

[0101]

[0102] in addition, This indicates that the measurement is performed using the SSIM function. and The similarity between them.

[0103] Due to light loss Low-texture regions of a scene lack information; therefore, this invention introduces smoothness before regularizing the depth map estimation. This invention employs an edge-aware image smoothing loss function:

[0104]

[0105] in The first derivative along the spatial direction represents the smoothness guided by the image edges.

[0106] To enhance the geometric consistency of depth estimation, a penalty is applied. and (pass Inconsistencies in back-projection calculations are used to constrain predictions. and They conform to the same 3D structure. Specifically, this invention proposes a differentiable depth inconsistency operation to compute pixel-level inconsistencies between two depth maps, namely the geometric consistency loss function:

[0107]

[0108] in yes The synthesis depth is determined by and It is generated through a rigid transformation at the underlying level.

[0109] The feature point ranking loss function is defined on a set of sparse point pairs with ordinal annotations. Specifically, for predicting depth maps... Point pairs in For the point pairs, the pairwise sorting loss is:

[0110]

[0111] in These are the actual ordinal labels, which can be derived from the sparse depth map:

[0112]

[0113] in It is the tolerance threshold. This represents the depth value in a sparse depth map. When these feature points are very close in depth space, i.e. This loss prompted the network to predict They must be identical; otherwise, their difference will increase. Therefore, given a sparse depth map, the feature point ranking loss function is:

[0114]

[0115] in , To predict depth maps The point in the middle.

[0116] The uncertainty loss function ensures that the network estimates the depth consistently for the same image under different flipping conditions. Therefore, its loss function is:

[0117]

[0118] in Indicates for image Image after horizontal flip Depth estimation map The dense depth map is obtained by horizontally flipping it again, where N is the predicted depth map. The number of points in the graph.

[0119] Ultimately, the overall loss of this online learning module was:

[0120]

[0121] in It is the loss weighting term.

[0122] The feature selection module uses depth estimation uncertainty information as a trigger for feature selection based on dense depth maps. The depth estimation uncertainty is generated by the RGB image... And horizontally flipped RGB images The depth estimates obtained through the improved depth estimation network are as follows: and depth estimation Then Flip it horizontally again to get and with Calculate the difference to obtain the uncertainty result of the deep estimation network after this online learning.

[0123] .

[0124] when When the value is less than a certain threshold, this module is triggered to select appropriate feature points using a dense depth map for pose estimation.

[0125] For feature point selection, after the pose estimation module is initialized, and after the system runs until the depth estimation uncertainty is less than the threshold. At that time, for points with shared viewing frames, the pose of the shared viewing frames is known. 2D image coordinates 3D world coordinates Establish the following optimization function:

[0126]

[0127]

[0128]

[0129] Derivation:

[0130]

[0131]

[0132] remember for ,

[0133] in, and The projection and back projection functions represent 3D world coordinate points.

[0134] The optimization function can then be written as:

[0135]

[0136] After optimization, use a dense depth map. Remove 3D points with large errors:

[0137]

[0138] The pose estimation module uses dense depth maps and sparse depth maps to calculate the scale information of the pose. The dense depth map... and sparse depth map Alignment method, calculating pose scale information:

[0139]

[0140] The final pose output is ,in Represents the in-situ pose without scale information, where N is the sparse depth map. The number of midpoints.

[0141] like Figure 4 As shown, a flowchart illustrating a preferred embodiment of the present invention for a localization method supporting online learning is provided, wherein:

[0142] Step 1 involves a depth estimation module, which estimates the dense depth map of the RGB image;

[0143] Step 2 involves an online learning module, which uses dense depth maps, pose information, and other information to calculate ranking loss function and uncertainty loss function to supervise the training of the depth estimation network.

[0144] Step 3 involves a feature selection module, which extracts ORB features and descriptors from the RGB image, and selects appropriate feature points for solving pose information using the estimated uncertainty and dense depth map.

[0145] Step 4 involves the pose estimation module, which uses the alignment of dense and sparse depth maps to calculate the scale information of the pose of the i-th frame, so as to ensure the scale consistency of the system during the localization process.

[0146] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A positioning system supporting online learning, characterized in that, It includes a depth estimation module, an online learning module, a feature selection module, and a pose estimation module, among which, The depth estimation module uses a depth estimation algorithm based on a low-rank adapter to estimate the dense depth map of the RGB image. The depth estimation algorithm uses a self-supervised mechanism to achieve online learning of the target domain depth estimation. The online learning module uses the dense depth map and pose information of the RGB image to calculate the ranking loss function and the uncertainty loss function, which are used to supervise the training of the depth estimation network. The feature selection module extracts ORB features and descriptors from the RGB image, and selects appropriate feature points to solve the pose information using the estimated uncertainty and the dense depth map. The pose estimation module uses the alignment of the dense depth map and the sparse depth map to calculate the scale information of the RGB image pose, ensuring the scale consistency of the system during the localization process. in, The depth estimation module uses the depth estimation network to implement the depth estimation algorithm. The depth estimation network adopts a U-Net encoder-decoder model structure. The encoder is set to a ResNet model pre-trained on the DDAD dataset. The decoder converts the output into depth values. The low-rank adapter is inserted into all convolutional layers in the encoder and the decoder and is connected in parallel with the convolutional layers. By freezing the pre-training parameters of the depth estimation network, the parameters of the low-rank adapter are learned and adjusted online only. The depth estimation network with the low-rank adapter module is used to estimate the depth of each RGB image keyframe to obtain the dense depth map. The low-rank adapter employs series-attention and a pyramid structure. It captures new information and knowledge in the image that has not been learned before, and improves the performance and adaptability of the deep estimation network in unfamiliar environments by adjusting its own learnable parameters. The online learning module supervises the learning of the deep estimation network by calculating a loss function, and the pose estimation module optimizes the input image. With images relative pose Then, construct the photometric loss function, image smoothing loss function, geometric consistency loss function, feature point ranking loss function, and uncertainty loss function required for the depth estimation network to learn; The feature selection module uses depth estimation uncertainty information as a trigger flag for feature selection based on dense depth maps. The depth estimation uncertainty is set as the uncertainty result calculated by the depth estimation network based on the RGB image, and the uncertainty result is calculated in the following way: ， in, For uncertainty results, RGB image The predicted depth map obtained through the depth estimation network, I' is the RGB image of image I after horizontal flipping. For image The predicted depth map obtained through a depth estimation network Depend on The depth map is obtained by flipping it horizontally again, where N is the predicted depth map. The number of pixels.

2. The positioning system as described in claim 1, characterized in that, When the uncertainty result is less than a preset threshold, the feature selection module is triggered to select appropriate feature points using a dense depth map for pose estimation.

3. A positioning method supporting online learning, characterized in that, The localization method uses a localization system supporting online learning as described in any one of claims 1-2. It employs a depth estimation algorithm based on a low-rank adapter to estimate a dense depth map of an RGB image. ORB features and descriptors are extracted from the RGB image. Using the estimation uncertainty and the dense depth map, suitable feature points are selected to calculate pose information. A ranking loss function and an uncertainty loss function are calculated using the dense depth map and the pose information to supervise the training of the depth estimation network. The depth estimation algorithm utilizes a self-supervised mechanism to achieve online learning of the target domain depth estimation.

4. The positioning method as described in claim 3, characterized in that, The positioning method includes the following steps: S101: Using the depth estimation module, estimate the dense depth map of the RGB image; S103: The online learning module uses the dense depth map and the pose information to calculate the ranking loss function and the uncertainty loss function to supervise the training of the depth estimation network; S105: The feature selection module extracts ORB features and descriptors from the RGB image, and selects appropriate feature points for solving the pose information using the estimated uncertainty and dense depth map; S107: The pose estimation module uses the alignment of the dense depth map and the sparse depth map to calculate the scale information of the pose, so as to ensure the scale consistency of the system during the localization process.

Images