Visual mapping method in low-light environment

Abstract

The application provides a visual mapping method in a low-light environment, and relates to the technical field of visual mapping. In the application, a brightness channel is separated from an input image and is divided into a plurality of tiles, local statistics of each tile are calculated to adaptively generate a contrast limiting parameter and an enhancement parameter vector, the brightness channel is enhanced based on the parameter, candidate points corresponding to feature points of an existing map are extracted after reconstruction of the image, a comprehensive confidence score is calculated, whether the candidate points and the corresponding map points are below a dynamic threshold in terms of the distance of the enhancement parameter vector is judged, and the confidence score is combined to determine whether the candidate points are adopted as new map points. Finally, the system dynamically optimizes the enhancement parameter according to the matching survival rate of feature points of each tile, continuously improves the mapping quality and stability through iterative updating, and realizes robust visual mapping in a low-light environment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of visual mapping, in particular to a visual mapping method in low-light environment. BACKGROUND

[0002] In key applications such as autonomous navigation of unmanned systems, underground space exploration, and night security monitoring, visual mapping technology often needs to operate continuously and stably in low-illumination environments. In such environments, the light intensity is significantly insufficient, resulting in a serious decline in the signal-to-noise ratio of the data captured by the image sensor. The feature point detection module, which is relied on by traditional visual mapping methods, has a sharp performance degradation due to low overall image contrast, blurred texture features, and intensified noise interference, which further leads to drift in pose estimation and even interruption of the mapping process. Therefore, it is of great practical significance to realize a technology that can stably and robustly complete mapping in low-light environments, in order to improve the environmental perception and autonomous operation capabilities of intelligent agents in complex lighting conditions.

[0003] Currently, to solve the problem of reduced reliability of visual mapping in low-light environments, existing technologies mainly focus on using image enhancement algorithms as the visual front-end preprocessing link. One common approach is to use limited contrast adaptive histogram equalization or its improved form, which performs equalization operations within each block by blocking the image and simultaneously constraining the upper limit of contrast enhancement, in order to balance the enhancement of image local details and the suppression of noise, thereby improving the extractability of feature points. Another method attempts to use deep learning models to implement end-to-end enhancement or directly extract features that are not sensitive to light changes, in order to improve the performance of the system in poor lighting conditions.

[0004] However, these existing methods still have obvious limitations. First, they usually rely on preset, globally uniform enhancement parameters, or although they are blocked, they do not dynamically adjust according to the local light and noise characteristics of each region, resulting in a lack of adaptability to different regions and difficulty in achieving balanced and effective enhancement in the entire image. Some regions still cannot extract features due to insufficient enhancement, while another part introduces a large amount of noise due to excessive enhancement. More fundamentally, the existing technology separates image enhancement from subsequent visual mapping tasks, forming an open-loop system. The enhancement module only optimizes the intermediate image quality or the number of feature points, without considering the impact of its output on the final mapping accuracy and stability of the SLAM system. For example, some enhancement operations may improve the subjective image quality or the number of feature points, but may also reduce the feature matching success rate and geometric consistency, thereby damaging the map quality. Existing methods cannot provide real-time feedback and closed-loop optimization of the front-end enhancement strategy according to the actual matching performance and geometric consistency indicators during the mapping process, so they cannot continuously provide reliable mapping capabilities in complex low-light environments. Figure One

[0005] ​The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide a visual mapping method in low-light environments to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A visual mapping method for low-light environments, comprising the following steps:

[0009] Step 1: Separate the luminance channel from the current input image in the HSV color space, divide the luminance channel into several tiles and calculate its local statistics, and generate the contrast limiting parameter for each tile based on the local statistics;

[0010] Step 2: Apply the CLAHE algorithm based on the contrast limiting parameter to obtain the enhanced luminance channel, and combine it with the original chrominance channel to reconstruct the enhanced image. Select candidate feature points from the enhanced image that correspond one-to-one with the feature points in the visual map. Combine the contrast limiting parameter and local statistics to generate the enhancement parameter vector of the tile where the candidate feature points are located, and calculate the comprehensive confidence score of each candidate feature point.

[0011] Step 3: If any candidate feature point and its corresponding feature point meet the matching conditions, then it is used as a map point. The matching conditions include that the vector distance between the enhanced parameter vector of the tile where the candidate feature point is located and the tile where the corresponding feature point is located is lower than the dynamic threshold, and the comprehensive confidence score of the candidate feature point exceeds the preset score threshold.

[0012] Step 4: Generate the matching survival rate of each tile based on map points and candidate feature points, and optimize the brightness channel of the current input image until the matching survival rate exceeds the survival rate threshold. Replace the corresponding feature points in the visual map with the optimized candidate feature points to obtain the final visual map.

[0013] Furthermore, the logic for dividing the luminance channel into several tiles and calculating its local statistics is as follows:

[0014] The luminance channel image is divided into M×N rectangular regions of the same size as tiles, with 0% overlap between each tile.

[0015] For each tile obtained from the division, its local statistics are calculated. These local statistics include the mean and standard deviation of the gray values ​​of all pixels in the tile, which are used to characterize the brightness level and contrast fluctuation of the region.

[0016] Furthermore, the logic for applying the CLAHE algorithm based on the contrast limiting parameter to obtain the enhanced luminance channel is as follows:

[0017] The contrast limiting parameter is generated based on the local noise level calculated from local statistics. For the k-th tile, its noise level is defined as follows: The contrast limiting parameter is defined as follows: The specific calculation formula is as follows:

[0018] ;

[0019] in, and Let be the mean and standard deviation of the pixel value of the k-th tile, respectively. It is a positive correction constant. Let k be the noise level of the k-th tile. Let k be the contrast limiting parameter for CLAHE. This is a limiting function, indicating that... Limited to the range Even if inside, The range of values ​​is within Inside, k is the index of the tile.

[0020] Furthermore, the logic for selecting candidate feature points from the enhanced image that correspond one-to-one with the feature points in the visual reconstruction is as follows:

[0021] Contrast limiting parameters based on each tile The CLAHE algorithm is applied to enhance the luminance channel, and the enhanced luminance channel is combined with the original chrominance channel to reconstruct the enhanced image. The FAST feature point detection algorithm is used to extract feature points from the original image in the visual mapping to obtain a feature point set. The same FAST feature point detection algorithm is used to extract feature points from the enhanced image to obtain a candidate feature point set.

[0022] For each feature point in the visual mapping, based on its predicted projection position in the image, a subset of spatially adjacent candidate feature points is selected in the enhanced image within a circular local search window with a preset radius. The descriptor distance between the feature point and each feature point in the subset of candidate feature points is calculated. The nearest neighbor to second nearest neighbor ratio method is used for preliminary matching and screening. The RANSAC-based geometric verification of the fundamental matrix is ​​performed on the descriptor matching pairs.

[0023] A one-to-one correspondence with the original feature points is established only when the candidate feature points pass the position constraints, descriptor matching and geometric verification simultaneously.

[0024] Furthermore, the logic for calculating the comprehensive confidence score for each candidate feature point is as follows:

[0025] For a candidate feature point p, calculate its comprehensive confidence score. The score is derived from a weighted combination of the normalized FAST corner response value and the normalized gradient magnitude. The specific calculation formula is as follows:

[0026] ;

[0027] in, The value is the normalized FAST corner response value at candidate feature point p, which is obtained by calculating the maximum grayscale difference between candidate feature point p and surrounding pixels. The gradient magnitude at candidate feature point p is the predefined normalized value, which is synthesized from the x and y gradient components calculated by the Sobel operator. This is a weighting coefficient used to adjust the contribution ratio of the FAST response value and gradient magnitude to the overall score; its value range is... p is the index of the candidate feature point;

[0028] The contrast limiting parameter and local statistics are combined to generate the enhancement parameter vector of the tile containing the candidate feature point, i.e., the enhancement parameter vector of the k-th tile. The contrast limiting parameter of the k-th tile and its local statistics are defined as follows: .

[0029] Furthermore, if any candidate feature point and its corresponding feature point satisfy the matching condition, the logic for using it as a map point is as follows:

[0030] If any candidate feature point meets the matching condition and is adopted as a map point, the augmentation parameter vector of its corresponding tile is... and the overall confidence score of the candidate feature point This is stored as metadata bound to the map point;

[0031] The matching condition includes that the vector distance between the tile containing the candidate feature point and the enhanced parameter vector of the corresponding feature point is less than a dynamic threshold. The specific logic is as follows:

[0032] Calculate the weighted Euclidean distance between the augmented parameter vector of the tile containing the current candidate feature point and the historical augmented parameter vector stored in the metadata of the map point to which the corresponding original feature point is bound. And compare this distance with a dynamic threshold. Satisfaction is determined by comparison, as defined below:

[0033] ;

[0034] in, Let be the enhancement parameter vector of the k-th tile where the current candidate feature point is located. This is the enhancement parameter vector for the tile containing the map point corresponding to the candidate feature point. It is a diagonal weighted matrix. As the baseline threshold, This is the scaling factor. The degree of fluctuation of the parameter vector within the spatial neighborhood of the tile:

[0035] ;

[0036] in, This is the set of spatial neighborhood tile indices centered at the current k-th tile. The number of tiles in the spatial neighborhood tile index set. This is the arithmetic mean of all augmentation parameter vectors within the spatial neighborhood tile index set. Let j be the enhancement parameter vector of the j-th tile in the spatial neighborhood tile index set, where j is the index of the spatial neighborhood tile index set;

[0037] When the vector distance criterion is Furthermore, when the overall confidence score associated with the map point exceeds the preset score threshold, and both of these criteria are met simultaneously, the candidate feature point is used as the map point.

[0038] Furthermore, the logic for generating the matching survival rate of each tile based on map points and candidate feature points is as follows:

[0039] definition To increase the total number of candidate feature points generated by tile k in the image, The match survival rate is the number of candidate feature points that are successfully matched and adopted as map points. Defined as:

[0040] ;

[0041] when Below the survival threshold If this occurs, the contrast limiting parameter of the tile is updated in a decay manner:

[0042] ;

[0043] in, As the attenuation factor, The optimized contrast limiting parameters, The contrast limiting parameters before optimization;

[0044] Based on optimized contrast limiting parameters Then, re-execute the CLAHE algorithm to enhance the brightness channels of the current and subsequent input images, and re-execute steps 2, 3, and 4 until... This generates a new set of candidate feature points;

[0045] The optimized candidate feature points replace the corresponding feature points in the visual map. The method is as follows: for matches where the survival rate has been improved and exceeds the survival rate threshold... The newly generated candidate feature points that meet the matching conditions are replaced with the corresponding old feature points in the visual map, and the descriptors and 3D position information of the relevant map points are updated, thereby completing the iterative optimization of the visual map.

[0046] Compared with the prior art, the beneficial effects of the present invention are:

[0047] This invention addresses the problem of uneven distribution of illumination and noise in different regions of an image. In step one, it dynamically generates a contrast limiting parameter and enhancement parameter vector that are adapted to the mean and standard deviation of brightness within each tile, thereby achieving differentiated enhancement processing for each region of the image: conservative enhancement is used in dark noise regions to suppress noise, while stronger contrast enhancement is implemented in regions with moderate brightness and rich texture to highlight feature structures, thus significantly improving the overall quality of the enhanced image and the usability of subsequent feature extraction.

[0048] This invention determines the consistency of the augmentation environment by calculating the distance between the enhancement parameter vectors of the candidate feature point and the tile containing the existing map point, and combines this with the comprehensive confidence score of the feature point to jointly decide whether to accept the point as a reliable map point. Furthermore, the system automatically evaluates the actual effect of the current enhancement parameters by continuously calculating the matching survival rate of feature points in each tile. If the survival rate of a certain area is low, its contrast limiting parameter is adaptively reduced, and the enhancement is re-enhanced and iteratively optimized until the feature points generated in that area can be stably matched.

[0049] This invention, based on a feedback adjustment mechanism for map point survival rate, enables visual mapping systems to extract richer and more stable feature points in low-light environments. Furthermore, it allows for continuous self-optimization through the feedback mechanism, significantly improving pose estimation accuracy and other performance characteristics. Figure One Its performance in terms of consistency and system robustness provides key technical support for achieving reliable autonomous navigation and map building in all weather conditions. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the overall method flow of the present invention;

[0051] Figure 2 This is a curve showing the fluctuation level versus the dynamic threshold fitting of the present invention.

[0052] Figure 3 This is a line graph showing the baseline threshold, scaling factor, fluctuation level, and dynamic threshold of this invention.

[0053] Figure 4 This is a bubble chart of the baseline threshold and dynamic threshold of the present invention;

[0054] Figure 5 This is a dual Y-axis graph of the baseline threshold and dynamic threshold of this invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0056] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0057] Example:

[0058] Please see Figures 1-5 The present invention provides a technical solution:

[0059] A visual mapping method for low-light environments, comprising the following steps:

[0060] Step 1: Separate the luminance channel from the current input image in the HSV color space, divide the luminance channel into several tiles and calculate its local statistics, and generate the contrast limiting parameter for each tile based on the local statistics;

[0061] The logic for dividing the luminance channel into several tiles and calculating its local statistics is as follows:

[0062] First, the image is converted from the conventional RGB color space to the HSV color space. The advantage of doing this is that it separates the brightness information from the chroma and saturation information. When enhancing an image in a low-light environment, the main concern is to improve the visibility of the brightness information, rather than to distort the color information. Separating the brightness channel for separate processing and then merging it with the original chroma information can preserve the original color of the object to the greatest extent.

[0063] The luminance channel image is divided into M×N rectangular regions of the same size as tiles, with 0% overlap between each tile.

[0064] The entire luminance channel image is divided into several non-overlapping small rectangular regions (tiles). This is because the lighting conditions of an image are often uneven. Global enhancement strategies, such as using a fixed parameter to perform histogram equalization on the entire image, will lead to over-enhancement in some areas and under-enhancement in others. Tile processing allows the algorithm to adaptively adjust the enhancement parameters according to the local characteristics (brightness, contrast, noise level) of each small region, thereby achieving a more balanced and natural enhancement effect.

[0065] For each tile obtained from the segmentation, its local statistics are calculated. These local statistics include the mean and standard deviation of the gray values ​​of all pixels within the tile, which are used to characterize the brightness level and contrast fluctuation of the region. Calculating local statistics is an adaptively enhanced perception step. By calculating the statistical characteristics (mean and standard deviation) of the pixels within each tile, the attributes of the region are quantitatively described. The mean senses the overall brightness level of the region, while the standard deviation senses the contrast or fluctuation of the region. In low-light images, a high standard deviation may originate from realistic texture details or from a large amount of noise.

[0066] The logic for obtaining the enhanced luminance channel by applying the CLAHE algorithm based on the contrast limiting parameter is as follows:

[0067] The contrast limiting parameter is generated based on the local noise level calculated from local statistics. For the k-th tile, its noise level is defined as follows: The contrast limiting parameter is defined as follows: The specific calculation formula is as follows:

[0068] ;

[0069] in, and Let be the mean and standard deviation of the pixel value of the k-th tile, respectively. This is a positive correction constant used to prevent the denominator from being zero. Let k be the noise level of the k-th tile. Let k be the contrast limiting parameter for CLAHE. This is a limiting function, indicating that... Limited to the range Even if inside, The range of values ​​is within Inside, k is the index of the tile;

[0070] It is a dimensionless quantity designed to comprehensively measure the relative significance of noise within a tile. The higher the value, the more severe the noise in the area. Usually appears in very dark ( Small) and large pixel value fluctuations ( Large areas are dangerous areas for enhancement algorithms in low-light images. If they are over-enhanced, they will amplify the noise as well, producing unsightly graininess or artifacts, which will reduce the image quality.

[0071] The larger, The larger the area, the higher its potential noise level, because the greater the fluctuation in the region. The larger the area, the brighter it is. The smaller the value, the more stable the fluctuation. In bright areas, noise is less noticeable than in dark areas, and it is less likely to be detected. It is a very small constant that prevents the denominator from being zero, ensuring that the formula is also defined in the pure black region;

[0072] These are the parameters directly input into the CLAHE algorithm. They determine the strictness of histogram cropping. They limit the maximum number of pixels that any bin (grayscale level) in the histogram can have. Pixels exceeding this limit will be cropped and redistributed evenly throughout the histogram.

[0073] The larger the value, the more aggressive the allowed contrast enhancement. The histogram is cropped less, resulting in very strong local contrast and more prominent details in the equalized image. However, this also dramatically increases the risk of amplifying noise. The smaller the value, the more conservative the enhancement strategy, the more the histogram is clipped, the smoother the equalization effect, and the more effectively the noise amplification is suppressed;

[0074] If the noise level of a region is estimated If it is very high (dark and highly volatile), then a relatively large [value] will be generated. CLAHE parameters The larger the value, the weaker the equalization effect, and the closer it is to the original histogram. Therefore, in areas with high noise, the algorithm is actually actively reducing the enhancement intensity to prevent the noise from being over-amplified; conversely, in areas with lower noise levels... Very low (brighter or very smooth) area, The smaller the contrast, the more aggressive the contrast enhancement, allowing CLAHE to extract details to the maximum extent.

[0075] The calculation results are limited to a reasonable empirical range. A lower limit of 0.5 ensures that a basic enhancement level is maintained even in the most ideal regions, while an upper limit of 40 ensures that the enhancement operation is not completely turned off even in the noisiest regions. When the value is infinite, it is equivalent to not performing histogram equalization, and still retains a certain enhancement effect;

[0076] Generating contrast limiting parameters is a decision-making step in adaptive enhancement. Based on statistics, parameters unique to each tile are calculated to control the core operation of the CLAHE algorithm—histogram clipping. Its purpose is to achieve a balance between dark areas with low contrast (which need to be enhanced) and areas where noise may be amplified (which need to be suppressed).

[0077] Step 2: Apply the CLAHE algorithm based on the contrast limiting parameter to obtain the enhanced luminance channel, and combine it with the original chrominance channel to reconstruct the enhanced image. Select candidate feature points from the enhanced image that correspond one-to-one with the feature points in the visual map. Combine the contrast limiting parameter and local statistics to generate the enhancement parameter vector of the tile where the candidate feature points are located, and calculate the comprehensive confidence score of each candidate feature point.

[0078] The logic for selecting candidate feature points from the enhanced image that correspond one-to-one with feature points in the visual reconstruction is as follows:

[0079] Contrast limiting parameters based on each tile The CLAHE algorithm is applied to enhance the luminance channel, and the enhanced luminance channel is combined with the original chrominance channel to reconstruct the enhanced image. The FAST feature point detection algorithm is used to extract feature points from the original image in the visual mapping to obtain a feature point set. The same FAST feature point detection algorithm is used to extract feature points from the enhanced image to obtain a candidate feature point set.

[0080] Applying CLAHE and image reconstruction is the core image preprocessing step. In low light environments, the overall image contrast is low and the noise is high, resulting in a small number and poor quality of directly extracted feature points. CLAHE adaptively enhances the contrast of each local tile, making the texture of the originally dark areas clearly visible, thereby greatly increasing the number of stable and repeatable feature points.

[0081] The detailed steps of the CLAHE algorithm to enhance the luminance channel are as follows:

[0082] The input full luminance channel image is divided into multiple rectangular regions of the same size, called tiles. For each tile, its gray-level histogram is calculated independently. This histogram counts the distribution of pixel luminance values ​​within the tile, showing how many pixels are in each gray level from the darkest to the brightest. The CLAHE algorithm examines the histogram of the tile to find the highest "peak", that is, the gray level with the most pixels.

[0083] A threshold is set, stipulating that the number of pixels at any gray level in the histogram cannot exceed this value. For example, if... If the value is set to 40, then the maximum number of pixels for each gray level in the histogram can be 40. Iterating through each gray level of the histogram, if the number of pixels for a certain gray level exceeds... The excess portion is then "trimmed off," a process that can be imagined as "peak shaving," flattening out excessively high peaks. Height;

[0084] The previous step removed many "redundant" pixels. These pixels are not discarded directly; the CLAHE algorithm redistributes these cropped pixels evenly across all gray levels of the histogram (some implementations only distribute them to those with less than one pixel). (On the gray levels), after "peak clipping", the histogram shape becomes flatter, avoiding excessive concentration of certain gray levels; the redistribution operation "fills the valleys", making the overall histogram more uniform. This operation ensures that the contrast enhancement is not too drastic during subsequent equalization, thus effectively suppressing the amplification of noise.

[0085] After cropping and redistribution, a new, modified histogram is obtained for the tile. Based on this new histogram, a histogram equalization transformation function is calculated separately for this tile. This function is a mapping table that defines which new gray level each pixel in the original image should become. By applying this transformation function, the brightness value of each pixel in the tile is mapped to a new value, thereby significantly improving the contrast within the tile.

[0086] If each tile is equalized independently, the final synthesized image will produce obvious and discontinuous "blocking effect" at the tile boundaries. To eliminate this effect, CLAHE uses bilinear interpolation to calculate the final value of each pixel. For any pixel in the image, it finds its four neighboring tiles, usually the top left, top right, bottom left, and bottom right tiles. Using the transformation function of each of these four tiles, it calculates four mapping values ​​for the pixel. Based on the distance of the pixel relative to the center point of these four tiles, it performs a weighted average of the four mapping values ​​to obtain the final gray value of the pixel. The interpolation process ensures that the pixel values ​​are smoothly transitioned, thus generating a visually natural and seamless enhanced image.

[0087] The magnitude of contrast enhancement is dynamically limited based on the local statistics (noise level) of each tile. Stronger enhancement is allowed in flat (low noise) areas to highlight details, while the enhancement magnitude is limited in textured (high noise) areas to prevent noise from being over-amplified. This adaptability is the core of ensuring the robustness of the enhancement effect.

[0088] After processing the luminance channel, it is merged with the original hue (H) and saturation (S) channels to reconstruct a complete RGB image. This ensures that the enhancement process only changes the luminance information, preserves the original color of the object to the maximum extent, and avoids the interference that color distortion may cause to subsequent feature recognition.

[0089] For each feature point in the visual mapping, based on its predicted projection position in the image, a subset of spatially adjacent candidate feature points is selected in the enhanced image within a circular local search window with a preset radius. The descriptor distance between the feature point and each feature point in the subset of candidate feature points is calculated. The nearest neighbor to second nearest neighbor ratio method is used for preliminary matching and screening. The RANSAC-based geometric verification of the fundamental matrix is ​​performed on the descriptor matching pairs.

[0090] Performing global feature matching directly on two vastly different images—the original low-light image and the enhanced image—is computationally intensive and unreliable. Predicting the projection location provides powerful spatial prior knowledge; based on previous map and camera motion estimation, it can predict where a map point is most likely to appear in the new image. The local search window narrows the matching search range from the entire image to a small neighborhood, greatly reducing computation and eliminating a large number of irrelevant erroneous candidate points.

[0091] The location of feature points is only a rough indication; it is also necessary to see if they look similar. Descriptors, such as ORB and BRIEF, are high-dimensional vectors that digitally represent the visual appearance around the feature points. The nearest neighbor distance (d1) is the distance of the best-matching descriptor, and the second nearest neighbor distance (d2) is the distance of the second-best-matching descriptor. The ratio (d1 / d2): if the best match is very unique, then d1 will be much smaller than d2, and the ratio will be very small, such as less than 0.8. If there are two candidates that look very similar, d1 and d2 will be very close, and the ratio will be large, such as greater than 0.9. This indicates high uncertainty in the match, and the match is rejected. The smaller the ratio, the higher the confidence of the match.

[0092] Geometric verification is the most robust verification checkpoint. Even if two points are similar in appearance and close in position, they must still conform to the geometric constraints of the entire camera motion. The fundamental matrix describes the geometric relationship between two viewpoints. The RANSAC algorithm is used to robustly estimate this geometric model in a set of matching pairs that include outliers (mismatches). The process is as follows: a small number of matching pairs are randomly selected to calculate a fundamental matrix F. Then, all matching pairs are checked to see if they conform to F, that is, whether one point is on the epipolar line corresponding to another point. Those that conform are called inliers, and those that do not are called outliers. This process is iterated multiple times, and finally the model with the most inliers is selected.

[0093] Only when a candidate feature point passes positional constraints, descriptor matching, and geometric verification simultaneously is a one-to-one correspondence with the original feature point established. This screening process is a funnel-shaped design from coarse to fine. Spatial position is used for extremely coarse screening, appearance descriptor is used for fine screening, and geometric consistency is used for final fine screening. The three-layer filtering ensures that the final one-to-one correspondence has extremely high accuracy and reliability.

[0094] The logic for calculating the overall confidence score for each candidate feature point is as follows:

[0095] For a candidate feature point p, calculate its comprehensive confidence score. The score is derived from a weighted combination of the normalized FAST corner response value and the normalized gradient magnitude. The specific calculation formula is as follows:

[0096] ;

[0097] in, The value is the normalized FAST corner response value at candidate feature point p, which is obtained by calculating the maximum grayscale difference between candidate feature point p and surrounding pixels. The gradient magnitude at candidate feature point p is the predefined normalized value, which is synthesized from the x and y gradient components calculated by the Sobel operator. This is a weighting coefficient used to adjust the contribution ratio of the FAST response value and gradient magnitude to the overall score; its value range is... p is the index of the candidate feature point;

[0098] It reflects the overall quality and stability of the candidate feature point p. A high score means that the point has both significant feature appearance and strong local texture support. Therefore, it is more likely to be a stable, repeatable feature point that contributes greatly to matching and localization.

[0099] The larger the value, the higher the overall quality of the candidate feature point p, and the more reliable it is. It should be given higher weight and trust in subsequent matching, map point generation and optimization processes. It is a weighted linear combination of two independent variables, the two independent variables, and The larger the value, the better. The larger the value, the better;

[0100] The FAST algorithm determines whether a point is a corner point by comparing the grayscale difference between a point and the surrounding pixels. The maximum value of this difference is its response value. The larger the response value, the greater the difference between the candidate feature point p and the surrounding environment, and the more obvious the characteristics of a corner point. The larger the value, the more obvious and stable the feature of the point as a corner point. A very prominent corner point, such as a window corner, is more likely to be repeatedly detected under different lighting and viewing angles than a blurry corner point, such as a worn table corner.

[0101] The original response value The result after min-max normalization, the original value The larger the value, the more prominent the corner point. The larger the value, the better. Normalization aims to eliminate the scaling effect caused by differences in overall contrast between different images, compressing its value to a smaller scale. The interval facilitates weighted combination with gradient terms;

[0102] The gradient magnitude reflects the local texture richness and edge intensity at the location of the candidate feature point p. It calculates the rate of change of the image at the candidate feature point along the x and y directions. The gradient magnitude is calculated using the Sobel operator; a larger magnitude indicates a more drastic change in the image at that point. The larger the value, the more clearly the point is located in a region with a clear texture, rather than a smooth or blurry region. Points with high gradient magnitudes have a more distinctive appearance, which is beneficial for generating high-quality feature descriptors and thus improving the matching success rate.

[0103] The larger the value, the richer the local image structure information around the point. Rich texture means that the descriptor will be more unique and the matching accuracy will be higher when performing feature matching. At the same time, in high gradient regions, the localization accuracy of feature points, such as obtaining sub-pixel coordinates through Gaussian fitting, will also be higher. It is the result of the original gradient magnitude after min-max normalization. The larger the original gradient magnitude, the stronger the texture. The larger it is, the more normalization is used for the same purpose as above: to achieve a uniform scale.

[0104] This reflects the relative importance of corner saliency (FAST response) and local texture intensity (gradient magnitude) in the overall score; This range indicates that, in this technique, corner response values ​​are considered slightly more important than gradient magnitudes. This is because the FAST response directly determines whether a point feature exists, while the gradient magnitude serves as supplementary verification of its quality; a point must first be a good corner. High weight), and on this basis, if it also has rich texture ( If the weight is high, then it is an excellent feature point. If the value is too low, such as 0.3, some edge points with strong textures but weak corner properties may get high scores, which may not be the desired result.

[0105] The contrast limiting parameter and local statistics are combined to generate the enhancement parameter vector of the tile containing the candidate feature point, i.e., the enhancement parameter vector of the k-th tile. The contrast limiting parameter of the k-th tile and its local statistics are defined as follows: ;

[0106] Constructing the enhancement parameter vector is the information encapsulation step, which encapsulates three key parameters representing the tile's enhancement environment and enhancement behavior. These are combined into a vector, which plays two key roles in subsequent steps:

[0107] Environment descriptor for subsequent matching: When determining whether two feature points are likely to match, in addition to appearance, their enhanced environments at the time of generation are compared, i.e., whether the vectors are similar. Points with similar environments have a higher probability of successful matching. State record for loop closure optimization: When the matching survival rate of a tile generation is low, this vector records the current state, providing guidance on how to adjust. The optimization results provided the basis for this.

[0108] The core significance lies in providing crucial contextual information for subsequent feature matching. In low-light visual SLAM, the appearance of the same physical point can change drastically at different times, under different lighting conditions, and with different enhancement parameters, causing traditional feature descriptor matching to fail. Vectors provide an additional, powerful constraint: only those vectors that, under similar augmented environments, are considered valid. Feature points generated by vector similarity are more likely to be the same physical point. For example, a point augmented with conservative parameters in a very dark environment and a point augmented with aggressive parameters in a brighter environment, even if they look somewhat similar, are very likely not the same point because their vector similarity is different. The vectors can differ greatly.

[0109] Step 3: If any candidate feature point and its corresponding feature point meet the matching conditions, then it is used as a map point. The matching conditions include that the vector distance between the enhanced parameter vector of the tile where the candidate feature point is located and the tile where the corresponding feature point is located is lower than the dynamic threshold, and the comprehensive confidence score of the candidate feature point exceeds the preset score threshold.

[0110] If any candidate feature point and its corresponding feature point satisfy the matching condition, then the logic for treating it as a map point is as follows:

[0111] If any candidate feature point meets the matching condition and is adopted as a map point, the augmentation parameter vector of its corresponding tile is... and the overall confidence score of the candidate feature point This is stored as metadata bound to the map point;

[0112] When a candidate point is successfully accepted as a map point, not only is its geometric location and appearance descriptor recorded, but its augmented parameter vector and comprehensive confidence score are also bound to it as metadata. This is equivalent to issuing a "birth certificate" to each map point, recording under what conditions and with what quality it was created.

[0113] The matching condition includes that the vector distance between the tile containing the candidate feature point and the enhanced parameter vector of the corresponding feature point is less than a dynamic threshold. The specific logic is as follows:

[0114] Calculate the weighted Euclidean distance between the augmented parameter vector of the tile containing the current candidate feature point and the historical augmented parameter vector stored in the metadata of the map point to which the corresponding original feature point is bound. And compare this distance with a dynamic threshold. Satisfaction is determined by comparison, as defined below:

[0115] ;

[0116] in, Let be the enhancement parameter vector of the k-th tile where the current candidate feature point is located. This is the enhancement parameter vector for the tile containing the map point corresponding to the candidate feature point. It is a diagonal weighted matrix. As the baseline threshold, This is the scaling factor. The degree of fluctuation of the parameter vector within the spatial neighborhood of the tile:

[0117] ;

[0118] in, This is the set of spatial neighborhood tile indices centered at the current k-th tile. The number of tiles in the spatial neighborhood tile index set. This is the arithmetic mean of all augmentation parameter vectors within the spatial neighborhood tile index set. Let j be the enhancement parameter vector of the j-th tile in the spatial neighborhood tile index set, where j is the index of the spatial neighborhood tile index set;

[0119] Defined as the set of indices of all tiles within a rectangular window formed by extending r tiles in each of the four directions (up, down, left, and right) from the geometric center of tile k. Here, r is a preset neighborhood radius, a positive integer that determines the size of the neighborhood. For example, when r is 1... It is a 3x3 set of tiles, containing a total of 9 tiles (including the center tile k itself). When r is 2, It is a 5x5 tile set containing 25 tiles;

[0120] This reflects the dispersion of the enhancement parameter vectors of each tile within the spatial neighborhood centered on the k-th tile. It indicates the complexity of the scene in that local area; for example, in an area containing both dark corners and bright windows, the mean brightness and contrast limiting parameters of each tile will vary significantly, leading to… It is very tall, and has a uniform wall. Then it will be very low;

[0121] A higher value indicates that the area is a scene with uneven lighting and complex textures. In such areas, enhancement parameters will naturally vary significantly. Therefore, even if the current parameters differ greatly from historical parameters, it may still be reasonable, and the threshold should be relaxed. ;

[0122] When the vector distance criterion is Furthermore, when the overall confidence score associated with the map point exceeds the preset score threshold, and both of these criteria are met simultaneously, the candidate feature point is taken as a map point. These two conditions complement each other and together ensure the high reliability and consistency of the newly added map points, laying a solid foundation for building stable and accurate maps under adverse lighting conditions.

[0123] in, It reflects the degree of difference between the enhancement parameters of the k-th tile in the current image and the historical enhancement parameters corresponding to the point when it was generated, stored in the map point metadata. It indicates whether the lighting and noise environment of a local area in the current image is similar to the environment when the map point was created in the historical environment. It measures the consistency between the two processing conditions;

[0124] The larger the value, the greater the difference between the processing parameters of the current frame and the historical parameters. This means that the lighting conditions, noise level or scene content of this area of ​​the current image may have changed significantly, resulting in the need to use a completely different enhancement strategy. This change will reduce the confidence of the match because the appearance of the features may change non-geometrically due to different processing methods.

[0125] It reflects the real-time enhancement processing strategy of the current input image in the k-th tile region, indicating the optimal enhancement intensity dynamically calculated by the system based on the average brightness, contrast and noise level of the region in the current frame; It reflects the historical enhancement processing strategy of the tile area where the map point was successfully created, indicating the environmental conditions of the area in a past frame and the optimal enhancement intensity that the system considered at that time.

[0126] This reflects the relative importance of different components in the augmentation parameter vector. Include The three components have different physical meanings, such as the contrast limiting parameter. Possibly higher than average brightness It has a greater impact on feature matching, therefore The matrix uses weights on its diagonal to scale the differences across different dimensions, resulting in a final distance. It can better reflect the actual impact of parameter differences on system performance. The greater the weight, the greater the contribution of parameter changes in this dimension to the distance value.

[0127] It reflects an allowable parameter difference boundary, which is not fixed but adaptively adjusted according to the parameter uniformity of local regions of the image. It indicates how much parameter difference is acceptable for the k-th tile in the current image. It is a flexible, context-dependent tolerance standard. The larger the value, the greater the difference between the current parameter and the historical parameter is allowed by the system. This usually occurs in complex scenarios where the parameters change drastically in local areas of the image. The system relaxes the matching requirements to avoid false rejection of true matches.

[0128] It reflects the maximum parameter difference that the system can accept under an ideal environment with uniform parameters. It is a basic tolerance value preset through experimental experience and serves as the benchmark for all dynamic adjustments. Reflects the degree of fluctuation of local parameters For the final threshold The intensity of the impact; It is an adjustment knob. The larger the value, the greater the local fluctuation. The stronger the impact on the relaxed threshold, the more the system adjusts... Values ​​are used to optimize matching sensitivity and specificity in complex scenarios.

[0129] Table 1 shows the specific data for some tile numbers and dynamic thresholds.

[0130] Table 1

[0131]

[0132] Analysis of the data reveals a clear synergistic relationship between different enhancement parameters. The data shows a stable positive correlation between the degree of fluctuation and the dynamic threshold; as the degree of fluctuation increases, the dynamic threshold also increases accordingly. For example, when the fluctuation of tile number 2 is 1.87, its dynamic threshold reaches 20.1, while the fluctuation of tile number 1 is only 0.15, and its dynamic threshold is 10.5. This indicates that in environments with high complexity in local image regions, the system needs to set a higher matching tolerance threshold to adapt to the natural fluctuations of the parameters.

[0133] When analyzing the relationship between the baseline threshold and the dynamic threshold, it was found that the setting of the baseline threshold directly affects the baseline level of the dynamic threshold. A higher baseline threshold will correspondingly increase the overall range of the dynamic threshold. For example, the baseline threshold for tile number 4 is 10.5, and the dynamic threshold reaches 17.3 when the fluctuation level is 1.24. On the other hand, the baseline threshold for tile number 3 is 9.5, and the dynamic threshold is 12.6 when the fluctuation level is 0.73. This shows that the baseline threshold, as a basic parameter, provides a benchmark reference for the adjustment of the dynamic threshold.

[0134] Meanwhile, the scaling factor also affects the dynamic threshold's response to fluctuations. A larger scaling factor makes the dynamic threshold more sensitive to changes in fluctuation. For example, tile number 4 has a scaling factor of 0.55, and its dynamic threshold increases more significantly with fluctuation, while tile number 3 has a scaling factor of 0.45, and its dynamic threshold increases relatively slowly. This indicates that the scaling factor determines the system's response to the complexity of the local environment; a higher scaling factor allows the system to adapt more flexibly to complex and changing image environments.

[0135] Step 4: Generate the matching survival rate of each tile based on map points and candidate feature points, and optimize the brightness channel of the current input image until the matching survival rate exceeds the survival rate threshold. Replace the corresponding feature points in the visual map with the optimized candidate feature points to obtain the final visual map.

[0136] The logic for generating the matching survival rate of each tile based on map points and candidate feature points is as follows:

[0137] definition To increase the total number of candidate feature points generated by tile k in the image, The match survival rate is the number of candidate feature points that are successfully matched and adopted as map points. Defined as:

[0138] ;

[0139] when Below the survival threshold If this occurs, the contrast limiting parameter of the tile is updated in a decay manner:

[0140] ;

[0141] in, As the attenuation factor, The optimized contrast limiting parameters, The contrast limiting parameters before optimization;

[0142] This reflects the conversion success rate of candidate feature points extracted from tile k within a specific time period or iteration round, ultimately passing all verification checkpoints and being adopted as map points. It directly indicates the image enhancement parameters currently used for that tile, especially the contrast limiting parameters. The effectiveness and reliability of the augmentation strategy are as follows: a high survival rate means that the current augmentation strength is just right, which improves feature visibility, suppresses noise, and produces a large number of stable features that can be used for mapping; a low survival rate is a strong signal that there may be a problem with the current augmentation strategy.

[0143] A higher value indicates that the tile is a high-yield and high-quality area, with reasonable enhancement parameter settings, and most of the produced feature points are effective. The smaller the value, the more inefficient the region is. The enhancement process may have introduced a large number of unstable noise points or distorted the features, causing it to fail the subsequent rigorous verification.

[0144] It reflects the number of valid outputs produced by a tile, indicating how many feature points are extracted from the tile and whose quality and uniqueness are sufficient to be accepted by the system through descriptor matching, geometric verification, and parameter consistency checks;

[0145] This reflects the amount of raw output produced by a tile, indicating how many potential corner points or features the image enhancement algorithm activated within that tile. Strong contrast enhancement will significantly increase this. ;

[0146] This reflects the new, adjusted image enhancement intensity applied to tile k with a low survival rate, indicating the system's decision to suppress noise and improve feature quality by reducing the enhancement intensity. The larger the value, the more contrast stretching the system allows for that area. The smaller the value, the more conservative the system's enhancement processing for that area; These are the enhancement parameters used in the previous iteration, and they serve as the baseline for the current adjustment.

[0147] It reflects the step size or intensity of each optimization adjustment of the system. It is a constant between 0 and 1, which determines the magnitude of parameter decay. For example, This means that the enhancement parameter will be reduced to 80% of its original value in each iteration; The larger the value, the more aggressive the optimization process, the greater the parameter reduction, and the faster the system is expected to converge to a steady state. The smaller the value, the more conservative and refined the optimization process is, avoiding missing the optimal value due to over-adjustment;

[0148] The function sets a lower limit for parameter adjustment to prevent the parameter from decaying to an invalid or negative value, thus ensuring the stability of the algorithm.

[0149] Based on optimized contrast limiting parameters Then, re-execute the CLAHE algorithm to enhance the brightness channels of the current and subsequent input images, and re-execute steps 2, 3, and 4 until... This generates a new set of candidate feature points;

[0150] In low-light environments, the main reason for low matching survival rates is usually the over-enhancement of noise, leading to instability of feature points. Therefore, the optimization direction is to suppress noise by reducing the enhancement intensity to improve the stability and repeatability of feature points. Low-light images themselves have a lot of noise, and strong contrast stretching will greatly amplify the originally inconspicuous noise, making it look like "pseudo-textures" or "pseudo-features". Over-equalization will cause unnatural color blocks, halos, or over-sharpened edges in local areas of the image, which are not real scene features. These amplified noises and distortions will be mistakenly identified as valid corner points by feature detectors, such as FAST, thus increasing the total number of candidate feature points. The number of successful matches is artificially inflated, containing a large number of low-quality, unstable noise points. The noise level is reduced because these noise points lack a real and stable structure, and they cannot be correctly descriptor-matched or geometrically verified with points in the map.

[0151] When the survival rate is low, it's usually because over-enhancement (too high a contrast limiting parameter value) amplifies noise, generating a large number of false and unstable feature points. By reducing the contrast limiting parameter value, the system adopts a more conservative enhancement strategy. While this may reduce the total number of extracted features, the goal is to significantly improve the quality and reliability of these features, thereby increasing the number of successful matches. The ultimate goal is to improve the ratio. This is a strategy of trading quality for quantity, which is crucial for the stability and accuracy of mapping.

[0152] The fundamental purpose of gradually reducing the contrast limiting parameter is to resolve the inherent contradiction between "enhancing details" and "suppressing noise" in low-light image enhancement. When a low matching survival rate is detected in a certain region, it indicates that although the current enhancement strategy extracts a large number of features, most of them are false or unstable features generated by noise amplification and cannot be used for accurate visual mapping. By gradually reducing the enhancement intensity through the attenuation factor, the strategy actively shifts from "aggressive" to "conservative". Its core function is to suppress noise amplification and prioritize feature quality rather than quantity. This means that although the total number of extracted features may be reduced, the signal-to-noise ratio and reliability of each retained feature point are significantly improved, thereby fundamentally improving the success rate of feature matching.

[0153] Lowering the contrast limiting parameter means a more stringent "cropping" of the histogram, which directly suppresses the contrast amplification ability of the CLAHE algorithm in local areas. This stricter cropping changes the image enhancement from "aggressive" to "conservative"; it effectively suppresses noise amplification and the generation of artifacts, although this may increase the total number of extracted feature points. While the number of features decreases (some weak true features may also go undetected temporarily), the remaining feature points exhibit significantly improved signal-to-noise ratio and stability. These "high-quality" feature points have a much higher probability of passing subsequent matching verification, thus increasing the number of successful matches. To improve or maintain stability, by using Feedback is used to gradually adjust the contrast limiting parameters. The system no longer relies on preset fixed parameters, but adaptively finds the most suitable processing intensity for the current environment. This enables the solution to cope with a variety of low-light environments, even dynamically changing ones, greatly enhancing its robustness.

[0154] The optimized candidate feature points replace the corresponding feature points in the visual map. The method is as follows: for matches where the survival rate has been improved and exceeds the survival rate threshold... The tiles are used to replace the newly generated candidate feature points that meet the matching conditions with the corresponding old feature points in the visual map, and the descriptors and 3D position information of the relevant map points are updated to complete the iterative optimization of the visual map.

[0155] Iterative optimization is not a one-time process, but a continuous loop: image processing → feature extraction → performance evaluation → parameter adjustment → image reprocessing; this process allows the system to automatically and gradually find the optimal solution even when the initial parameters are not ideal. Especially in scenes where the environment changes gradually, such as when the light gradually dims, this mechanism can ensure that the system maintains good performance.

[0156] Updating the descriptor ensures that map points have the latest and most accurate visual appearance information, while updating the 3D position optimizes the geometric accuracy of map points using new and more reliable observation data. This step ensures that the long-term consistency and accuracy of the visual map are constantly improving itself.

[0157] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0158] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0159] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0160] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A visual mapping method for low-light environments, characterized in that, The specific steps include: Step 1: Separate the luminance channel from the current input image in the HSV color space, divide the luminance channel into several tiles and calculate its local statistics, and generate the contrast limiting parameter for each tile based on the local statistics; Step 2: Apply the CLAHE algorithm based on the contrast limiting parameter to obtain the enhanced luminance channel, and combine it with the original chrominance channel to reconstruct the enhanced image. Select candidate feature points from the enhanced image that correspond one-to-one with the feature points in the visual map. Combine the contrast limiting parameter and local statistics to generate the enhancement parameter vector of the tile where the candidate feature points are located, and calculate the comprehensive confidence score of each candidate feature point. Step 3: If any candidate feature point and its corresponding feature point meet the matching conditions, then it is used as a map point. The matching conditions include that the vector distance between the enhanced parameter vector of the tile where the candidate feature point is located and the tile where the corresponding feature point is located is lower than the dynamic threshold, and the comprehensive confidence score of the candidate feature point exceeds the preset score threshold. Step 4: Generate the matching survival rate of each tile based on map points and candidate feature points, and optimize the brightness channel of the current input image until the matching survival rate exceeds the survival rate threshold. Replace the corresponding feature points in the visual map with the optimized candidate feature points to obtain the final visual map. The logic for generating the matching survival rate of each tile based on map points and candidate feature points is as follows: definition To increase the total number of candidate feature points generated by tile k in the image, The match survival rate is the number of candidate feature points that are successfully matched and adopted as map points. Defined as: when Below the survival rate threshold If this occurs, the contrast limiting parameter of the tile is updated in a decay manner: in, As the attenuation factor, The optimized contrast limiting parameters, The contrast limiting parameters before optimization; Based on optimized contrast limiting parameters Then, re-execute the CLAHE algorithm to enhance the brightness channels of the current and subsequent input images, and re-execute steps 2, 3, and 4 until... This generates a new set of candidate feature points; The optimized candidate feature points replace the corresponding feature points in the visual map. The method is as follows: for matches where the survival rate has been improved and exceeds the survival rate threshold... The newly generated candidate feature points that meet the matching conditions are replaced with the corresponding old feature points in the visual map, and the descriptors and 3D position information of the relevant map points are updated, thereby completing the iterative optimization of the visual map.

2. The visual mapping method in a low-light environment according to claim 1, characterized in that: The logic for dividing the luminance channel into several tiles and calculating its local statistics is as follows: The luminance channel image is divided into M×N rectangular regions of the same size as tiles, with 0% overlap between each tile. For each tile obtained from the division, its local statistics are calculated. These local statistics include the mean and standard deviation of the gray values ​​of all pixels in the tile, which are used to characterize the brightness level and contrast fluctuation of the region.

3. The visual mapping method in a low-light environment according to claim 2, characterized in that: The logic for obtaining the enhanced luminance channel by applying the CLAHE algorithm based on the contrast limiting parameter is as follows: The contrast limiting parameter is generated based on the local noise level calculated from local statistics. For the k-th tile, its noise level is defined as follows: The contrast limiting parameter is defined as follows: The specific calculation formula is as follows: in, and Let be the mean and standard deviation of the pixel value of the k-th tile, respectively. It is a positive correction constant. Let k be the noise level of the k-th tile. Let k be the contrast limiting parameter for CLAHE. This is a limiting function, indicating that... Limited to the range Even if inside, The range of values ​​is within Inside, k is the index of the tile.

4. The visual mapping method in a low-light environment according to claim 3, characterized in that: The logic for selecting candidate feature points from the enhanced image that correspond one-to-one with feature points in the visual reconstruction is as follows: Contrast limiting parameters based on each tile The CLAHE algorithm is applied to enhance the luminance channel, and the enhanced luminance channel is combined with the original chrominance channel to reconstruct the enhanced image. The FAST feature point detection algorithm is used to extract feature points from the original image in the visual mapping to obtain a feature point set. The same FAST feature point detection algorithm is used to extract feature points from the enhanced image to obtain a candidate feature point set. For each feature point in the visual mapping, based on its predicted projection position in the image, a subset of spatially adjacent candidate feature points is selected in the enhanced image within a circular local search window with a preset radius. The descriptor distance between the feature point and each feature point in the subset of candidate feature points is calculated. The nearest neighbor to second nearest neighbor ratio method is used for preliminary matching and screening. The RANSAC-based geometric verification of the fundamental matrix is ​​performed on the descriptor matching pairs. A one-to-one correspondence with the original feature points is established only when the candidate feature points pass the position constraints, descriptor matching and geometric verification simultaneously.

5. The visual mapping method in a low-light environment according to claim 4, characterized in that: The logic for calculating the overall confidence score for each candidate feature point is as follows: For a candidate feature point p, calculate its comprehensive confidence score. The score is derived from a weighted combination of the normalized FAST corner response value and the normalized gradient magnitude. The specific calculation formula is as follows: in, The value is the normalized FAST corner response value at candidate feature point p, which is obtained by calculating the maximum grayscale difference between candidate feature point p and surrounding pixels. The gradient magnitude at candidate feature point p is the predefined normalized value, which is synthesized from the x and y gradient components calculated by the Sobel operator. This is a weighting coefficient used to adjust the contribution ratio of the FAST response value and gradient magnitude to the overall score; its value range is... p is the index of the candidate feature point; The contrast limiting parameter and local statistics are combined to generate the enhancement parameter vector of the tile containing the candidate feature point, i.e., the enhancement parameter vector of the k-th tile. The contrast limiting parameter of the k-th tile and its local statistics are defined as follows: .

6. The visual mapping method in a low-light environment according to claim 5, characterized in that: If any candidate feature point and its corresponding feature point satisfy the matching condition, then the logic for treating it as a map point is as follows: If any candidate feature point meets the matching criteria and is adopted as a map point, the augmentation parameter vector of its corresponding tile is... and the overall confidence score of the candidate feature point This is stored as metadata bound to the map point; The matching condition includes that the vector distance between the tile containing the candidate feature point and the enhanced parameter vector of the corresponding feature point is less than a dynamic threshold. The specific logic is as follows: Calculate the weighted Euclidean distance between the augmented parameter vector of the tile containing the current candidate feature point and the historical augmented parameter vector stored in the metadata of the map point to which the corresponding original feature point is bound. And compare this distance with a dynamic threshold. Satisfaction is determined by comparison, as defined below: in, Let be the enhancement parameter vector of the k-th tile where the current candidate feature point is located. This is the enhancement parameter vector for the tile containing the map point corresponding to the candidate feature point. It is a diagonal weighted matrix. As the baseline threshold, This is the scaling factor. The degree of fluctuation of the parameter vector within the spatial neighborhood of the tile: in, This is the set of spatial neighborhood tile indices centered at the current k-th tile. The number of tiles in the spatial neighborhood tile index set. This is the arithmetic mean of all augmentation parameter vectors within the spatial neighborhood tile index set. Let j be the enhancement parameter vector of the j-th tile in the spatial neighborhood tile index set, where j is the index of the spatial neighborhood tile index set; When the vector distance criterion is Furthermore, when the overall confidence score associated with the map point exceeds the preset score threshold, and both of these criteria are met simultaneously, the candidate feature point is used as the map point.

Images