Texture-oriented uniform FAST feature extraction method, device, medium and equipment

Abstract

The embodiment of the present invention discloses a texture-oriented uniform FAST feature extraction method, which comprises: constructing a texture-oriented uniform feature extractor; obtaining a texture histogram of a feature picture to be extracted from a feature picture through the texture-oriented uniform FAST feature extractor; determining the contrast and information entropy of the texture histogram of the feature picture to be extracted; presetting a basic threshold according to the contrast and information entropy; determining an optimized threshold according to the basic threshold, a contrast weight and an information entropy weight; extracting feature points of the feature picture to be extracted through the optimized threshold and determining a feature point density distribution map; dynamically adjusting the division of a quadtree according to the density and the number of feature points and optimizing the distribution structure of feature points.

Description

Technical Field

[0001] The present invention relates to the technical field of image feature extraction, and in particular to a texture-oriented uniform FAST feature extraction method, device, medium and equipment. Background Art

[0002] Existing VSLAM feature extractors are particularly sensitive to texture and illumination changes in indoor environments, and are prone to missed detections and false detections in changing environments;

[0003] In addition, the aggregation and redundancy of feature points in texture-rich areas make map construction overly dependent on local information, increasing the risk of mapping failure. Especially when hardware resources are limited, these problems will significantly affect the accuracy of positioning and mapping. Summary of the invention

[0004] Based on this, it is necessary to propose a texture-oriented uniform FAST feature extraction method, device, medium and equipment to address the above problems.

[0005] A texture-oriented uniform FAST feature extraction method, the method comprising:

[0006] Constructing texture-oriented uniform feature extractor;

[0007] The texture histogram of the image to be extracted is obtained by using a texture-oriented uniform FAST feature extractor for the image to be extracted.

[0008] Determine the contrast and information entropy of the texture histogram of the feature image to be extracted;

[0009] Presetting a basic threshold value according to the contrast and information entropy;

[0010] Determine an optimization threshold according to the basic threshold, contrast weight and information entropy weight;

[0011] Extracting feature points of the image to be extracted using the optimized threshold and determining a feature point density distribution map;

[0012] The quadtree division is dynamically adjusted according to the density and number of feature points and the distribution structure of feature points is optimized.

[0013] Preferably, before obtaining a texture histogram of the image to be extracted by a texture-oriented uniform FAST feature extractor, the method further comprises performing a dedistortion process on the image to be extracted to obtain a corrected image to be extracted:

[0014] For each coordinate point in the image to be extracted, it is represented as x and y, and the coordinate point after dedistortion is represented as x distorted anddistorted , the distorted coordinates are:

[0015] x distorted =x(1+k 1 r 2 +k 2 r 4 +k 3 r 6 )+2p 1 xy+p 2 (r 2 +2x 2 )

[0016] y distorted =y(1+k 1 r 2 +k 2 r 4 +k 3 r 6 )+2p 2 xy+p 1 (r 2 +2y 2 )

[0017] Among them, k1, k2, k3, p1, and p2 are distortion parameters, x and y are the coordinates of the point on the normalized plane, and r is the distance between the point on the plane and the origin of the normalized plane.

[0018] Preferably, determining the contrast and information entropy of the texture histogram of the feature picture to be extracted specifically includes:

[0019] According to Determine the contrast C and information entropy E, where M and N are the dimensions of the image, representing the number of rows and columns of pixels respectively, and f(i,j) represents the grayscale value of the pixel with coordinates (i,j) in the image.

[0020] Preferably, the presetting of the basic threshold according to the contrast and information entropy specifically includes:

[0021] When it is determined that the contrast C is low, the basic threshold is set to C 1 ', where C 1 'The value is set to (0,3);

[0022] When the contrast C is determined to be high, the basic threshold is set to C 2 ', where C 2 ' is set to (20, +∞);

[0023] When the contrast C is determined to be medium, the basic threshold is set to T according to the information entropy E. b '.

[0024] Preferably, when the contrast C is determined to be medium, the basic threshold T is determined according to the information entropy E. b ', specifically including: when determining that the information entropy E is low, the preset basic threshold T b 'Set to C' 1 Add the first compensation value; when it is determined that the information entropy E is high, the preset basic threshold T b 'Set to C' 1 A second compensation value is added, and the second compensation value is greater than the first compensation value.

[0025] Preferably, the step of determining the optimization threshold according to the basic threshold, contrast weight and information entropy weight specifically includes: according to T=T b +50×ρ c ×ρ E Determine the optimization threshold, where ρ C and ρ E They are positively correlated with image contrast C and information entropy E,

[0026] Preferably, extracting feature points of the image to be extracted using the optimized threshold and determining a feature point density distribution map specifically includes: Determine the feature point density distribution map, where the value k at each position represents the specific area range S R The number of feature points within.

[0027] A texture-oriented uniform FAST feature extraction device, the device comprising:

[0028] A building block for constructing a texture-oriented uniform feature extractor;

[0029] The texture extraction module is used to obtain the texture histogram of the image to be extracted through the texture-oriented uniform FAST feature extractor;

[0030] The module for obtaining information of the picture to be extracted is used to determine the contrast and information entropy of the texture histogram of the feature picture to be extracted;

[0031] An optimization module, configured to preset a basic threshold value according to the contrast and information entropy; and determine an optimization threshold value according to the basic threshold value, contrast weight and information entropy weight;

[0032] The feature extraction module is used to extract the feature points of the picture to be extracted through the optimization threshold and determine the density distribution map of the feature points; it is also used to dynamically adjust the quadtree division according to the density and the number of feature points and optimize the distribution structure of the feature points.

[0033] A computer-readable storage medium stores a computer program, which, when executed by a processor, causes the processor to perform the following steps:

[0034] Constructing texture-oriented uniform feature extractor;

[0035] The texture histogram of the image to be extracted is obtained by using a texture-oriented uniform FAST feature extractor for the image to be extracted.

[0036] Determine the contrast and information entropy of the texture histogram of the feature image to be extracted;

[0037] Presetting a basic threshold value according to the contrast and information entropy;

[0038] Determine an optimization threshold according to the basic threshold, contrast weight and information entropy weight;

[0039] Extracting feature points of the image to be extracted using the optimized threshold and determining a feature point density distribution map;

[0040] The quadtree division is dynamically adjusted according to the density and number of feature points and the distribution structure of feature points is optimized.

[0041] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the following steps:

[0042] Constructing texture-oriented uniform feature extractor;

[0043] The texture histogram of the image to be extracted is obtained by using a texture-oriented uniform FAST feature extractor for the image to be extracted.

[0044] Determine the contrast and information entropy of the texture histogram of the feature image to be extracted;

[0045] Presetting a basic threshold value according to the contrast and information entropy;

[0046] Determine an optimization threshold according to the basic threshold, contrast weight and information entropy weight;

[0047] Extracting feature points of the image to be extracted using the optimized threshold and determining a feature point density distribution map;

[0048] The quadtree division is dynamically adjusted according to the density and number of feature points and the distribution structure of feature points is optimized.

[0049] Compared with the prior art, the present invention obtains a texture-oriented feature point extractor by improving the traditional VSLAM feature point extractor, then inputs the image to be extracted into the texture-oriented feature point extractor to obtain a basic threshold for extraction, finally, inputs the basic threshold into the texture-oriented feature point extractor to obtain an optimized threshold, further determines the feature point extraction result of the image to be extracted according to the optimized threshold, and further homogenizes the extracted feature point map to obtain a result of the feature point map as the final extraction result of the image to be extracted.

[0050] Compared with the traditional way of applying the image feature point extractor to the task of the VSLAM front end, the present invention improves the feature point extractor to obtain a texture-oriented feature point extractor, then inputs the image to be extracted into the feature point extractor for feature point extraction, and then further homogenizes the extracted feature points, and divides and extracts the image areas with different feature point densities. The feature points extracted after the homogenization are the final extraction results, so that the present invention can realize feature extraction in different surrounding pixel situations through the texture-oriented extractor, thereby improving the accuracy of feature point extraction, and further improving the robustness and stability of the feature extraction algorithm in the VSLAM system. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0052] in:

[0053] Figure 1 A flowchart of a texture-oriented uniform FAST feature extraction method is provided for an embodiment of the present invention;

[0054] Figure 2 A schematic diagram of an eight-layer pyramid model in a texture-oriented uniform FAST feature extraction method is provided for an embodiment of the present invention;

[0055] Figure 3 The statistical results of contrast and information entropy of randomly selected samples;

[0056] Figure 4 This is the experimental scene diagram.

[0057] Figure 5 This is a qualitative comparison diagram of TOHF;

[0058] Figure 6Experimental results on the Room dataset. DETAILED DESCRIPTION

[0059] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0060] The embodiment of the present invention provides a texture-oriented uniform FAST feature extraction method, such as Figure 1 As shown, the method includes:

[0061] Step 101: construct a texture-oriented uniform feature extractor;

[0062] Specifically, image extraction based on VSLAM is achieved through a texture-oriented uniform feature extractor. The feature point extractor is a feature point extractor composed of transforming feature maps and extracting feature points. The texture-oriented uniform FAST feature extractor can be specifically divided into an image preprocessing module, a FAST corner point detection module, a BRIEF descriptor calculation module, and a feature point screening module.

[0063] The image and processing module includes image acquisition function, dedistortion function, grayscale function, and image pyramid function.

[0064] Step 102: obtaining a texture histogram of the image to be extracted using a texture-oriented uniform FAST feature extractor;

[0065] Specifically, the texture features include gray-level co-occurrence matrix (GLCM), local binary pattern (LBP), and Gabor filter response;

[0066] Before obtaining the texture histogram, the feature image to be extracted is preprocessed, including:

[0067] Performing a dedistortion process on the image to be extracted to obtain a corrected image to be extracted;

[0068] After acquiring the image through the sensor, if the camera lens has distortion (such as radial distortion and tangential distortion), the image needs to be dedistorted. This step is to eliminate the image distortion caused by the lens shape in order to more accurately estimate the camera motion and build a map. For each coordinate point in the image to be extracted, it is represented by x and y, and the coordinate point after dedistortion is represented by x distorted and distorted , the distorted coordinates are:

[0069] x distorted =x(1+k 1 r 2 +k 2 r 4 +k 3 r 6 )+2p 1 xy+p 2 (r 2 +2x 2 )

[0070] y distorted =y(1+k 1 r 2 +k 2 r 4 +k 3 r 6 )+2p 2 xy+p 1 (r 2 +2y 2 )

[0071] Among them, k1, k2, k3, p1, and p2 are distortion parameters, x and y are the coordinates of the point on the normalized plane, and r is the distance between the point on the plane and the origin of the normalized plane.

[0072] Performing grayscale processing on the corrected image to be extracted to obtain a grayscale image to be extracted;

[0073] The image may be a color image. Although color information may be important in some applications, the feature point extractor usually performs feature point extraction on a grayscale image because it is more sensitive to grayscale changes and can better save computing performance when calculating FAST points, making the processing more efficient.

[0074] A pyramid of multi-level grayscale images to be extracted is constructed according to the grayscale images to be extracted.

[0075] Since it is not possible to extract FAST points better for distant images in the actual image acquisition process, the image pyramid level is used, such as Figure 2 As shown in the figure, an image pyramid usually consists of multiple levels, each corresponding to a different image resolution. The higher the level, the lower the resolution and the more blurred the image. For pyramid levels of different scales, the same feature extraction algorithm can be used to detect key points in the image at different scales. In this way, objects at different distances can be detected and the stability of feature point detection under different conditions can be improved.

[0076] Determine candidate corner points in a multi-level pyramid grayscale image to be extracted; specifically include: selecting any pixel in the multi-level pyramid grayscale image to be extracted as a center point; determining a neighborhood of the center point; determining the intensity of each pixel in the neighborhood and the intensity of the center point; if the number of pixels in the neighborhood with an intensity greater than the center point is greater than a first threshold, determine the center point as a candidate corner point;

[0077] Determine the gradient information in the neighborhood around the candidate corner point, and determine the main direction of the candidate corner point based on it;

[0078] The pixel area around the candidate corner point is divided into several sub-areas, and the local gradient information relative to the main direction is calculated for each sub-area, and the BRIEF descriptor with rotation invariance is determined based on it:

[0079] The rotation-invariant BRIEF descriptor obtained through the above process is used to perform feature matching on candidate corner points.

[0080] Step 103: Determine the contrast and information entropy of the texture histogram of the feature image to be extracted;

[0081] Specifically, according to

[0082]

[0083] Determine the contrast C and information entropy E, where M and N are the dimensions of the image, representing the number of rows and columns of pixels respectively, and f(i,j) represents the grayscale value of the pixel with coordinates (i,j) in the image.

[0084] Step 104: presetting a basic threshold according to the contrast and information entropy;

[0085] Specifically, high, medium, and low are determined by comparing the contrast C with a critical value of low contrast to medium contrast and a critical value of medium contrast to relatively high contrast.

[0086] By sampling and training multiple standard data sets, the contrast and information entropy obtained are divided into numerical regions, where (0,5) is defined as low contrast, (5,25) is defined as medium contrast, and (25,+∞) is defined as high contrast.

[0087] When it is determined that the contrast C is low, the basic threshold is set to C 1 ', where C 1 'The value is set to (0,3) to retain as many valid feature points as possible;

[0088] When it is determined that the contrast C is medium, the basic threshold is set to be, according to the information entropy E, where T bThe value of ' is set to (3,20) to improve the stability of feature points and reduce the missed detection of feature points due to low texture complexity.

[0089] When the contrast C is determined to be high, the basic threshold is set to C 2 ', where C 2 ' is set to (20, +∞) to ensure the stability of feature points;

[0090]

[0091] Among them, C 1 and C 2 are the low contrast and high contrast thresholds, T b is the basic threshold.

[0092] When the contrast C is determined to be medium, even if the contrast of two images is the same, their information entropy may be greatly different, for example, a wall image (with low information entropy) and a desk image (with high information entropy) taken in an indoor environment. If a unified basic threshold is used, feature points in images with low information entropy are easily missed, and feature points in images with high information entropy are easily misdetected. In order to reduce the missed detection rate and the misdetection rate, this paper adds a threshold value in C. 1 'Based on the statistical information entropy parameter E 1 To set the basic threshold T in general scenarios b ', which is calculated as follows:

[0093]

[0094] When the image information entropy is low, T b 'In C 1 ' is set as the first compensation value to improve the stability of feature points and reduce missed feature points due to low texture complexity; when the image information entropy is high, T b 'In C 1 ' is set as the second compensation value to avoid false detection of feature points due to high texture complexity.

[0095] The first compensation value described in the above process has a value range of (0, 10), and the second compensation value has a value range of (10, 20), and the second compensation value is greater than the first compensation value.

[0096] Step 105: determining an optimized threshold according to the basic threshold, contrast weight and information entropy weight;

[0097] Specifically, because different images in the same contrast and information entropy range still have certain differences in intensity difference and information amount, in order to capture this difference, the contrast weight factor and information entropy weight factor are set respectively to fine-tune the basic threshold to obtain the optimized threshold.

[0098] According to T = T b +50×ρ c ×ρ E Determine the optimization threshold, where ρ C and ρ E They are positively correlated with image contrast C and information entropy E,

[0099] Information entropy and contrast jointly control the fine-tuning value to ensure that the adjustment of the FAST threshold has a reasonable impact. This feature extraction method can intelligently select the FAST threshold in different environments, thereby automatically determining to extract more or fewer feature points, enhancing the algorithm's adaptability to image texture complexity and illumination diversity, and improving the quality of feature points.

[0100] This feature extraction method can intelligently select the FAST threshold in different environments, thereby automatically determining to extract more or fewer feature points, enhancing the algorithm's adaptability to image texture complexity and illumination diversity, and improving the quality of feature points.

[0101] Step 106: extracting feature points of the image to be extracted using the optimized threshold and determining a feature point density distribution map;

[0102] Step 107: Dynamically adjust the quadtree division according to the density and the number of feature points and optimize the distribution structure of the feature points.

[0103] The feature point map is evenly distributed by using a density-aware quadtree uniformization strategy. First, the feature point density distribution map of the image is calculated, where the value k at each position represents the specific area range S R The number of feature points within, as shown in the formula As shown, the quadtree is then divided according to the density of each node. If the density of the current node is greater than the average density of the area where the node is located before the division, the division will continue. Finally, the representative feature points are screened by the maximum response value. In areas with higher density, the present invention will divide more deeply, and in areas with lower density, the division will be stopped. This not only avoids the increase in computational complexity and storage overhead caused by excessive division, and improves system efficiency; it also ensures that the original distribution of feature points is maintained in areas with lower density of feature points, thereby reducing the loss of valid feature points. In short, the present invention enhances the division capability through the perception of the density of feature points, effectively improves the quality of feature points, and thereby ensures the quality of subsequent feature matching and map construction.

[0104] For example, based on multiple experiments on the EuRoc dataset on ORB-SLAM3, the FAST thresholds set to 20 and 7 have good effects. In order to ensure the robustness of TOHF indoors, 60 images were randomly selected for statistical analysis using the indoor environment dataset UMA-VI and the TUM dataset, as shown in Figure 2. Figure 3 As shown, the following formula is defined to measure the effectiveness of the matched feature points in two images.

[0105]

[0106] Among them, N m is the number of feature points after brute force matching, N k1 is the number of feature points extracted from the first image, N k2 is the number of feature points extracted from the second image, and P is the matching rate.

[0107] like Figure 4 As shown in (i), experiments were conducted on multiple sub-datasets of the EuRoc dataset and data recorded by a sweeping robot in a resource-constrained environment (Room dataset), which are specifically presented in Figure 4 (i). The monocular camera used has a resolution of 300,000 pixels and a frame rate of 2 frames per second. Data was recorded as the robot vacuum traversed the entire area of ​​each room along the predetermined path. Although the limitations of the camera configuration impose certain constraints on the image acquisition speed and quality, this also reflects the challenges that actual household vacuum robots may face under resource-constrained conditions, increasing practical considerations.

[0108] The Room dataset is unique in that it has a top-view camera angle, which is in contrast to conventional front-view or side-view cameras, and is designed to better protect user privacy. Combined with the complexity of textures and the limitations of image clarity in home environments, this dataset provides unique challenges in visual localization and mapping tasks. In order to more fully evaluate the robustness of the proposed method, the original images are adjusted for contrast and brightness to simulate the effects of changing lighting conditions in the real world, as shown in Figure 4 (ii) and (iii).

[0109] According to the data in Tables 1, 2, and 3, it can be clearly seen that TOHF shows significant superiority over other algorithms in various scenarios. In simple scenarios, the matching rate of TOHF on the original image is as high as 83.33%, far exceeding FAST, BRISK, and AKAZE, which failed to effectively extract feature points in these scenarios. For general scenes, the false detection problem of the FAST algorithm is relatively significant. Even under the two recommended threshold settings, its matching rate is lower than that of TOHF at 9.99% and 4.20%, respectively. More worryingly, when the scene becomes more complex, especially when the contrast and brightness change, the threshold selection of FAST becomes a critical factor. In the original image, the matching rate of the threshold of 20 in general and complex scenes is higher than that of the threshold of 7, which is 5.79% and 5.14%, respectively. However, in scenes with changes in contrast and brightness, the matching rate of FAST with a threshold of 20 drops significantly, by 3.28% and 16.33%, respectively (see Figure 2). Figure 4 (ii)). In comparison, BRISK, AKAZE, and TOHF all show high stability in various scenarios, among which TOHF performs particularly well, outperforming other algorithms in terms of both stability and matching rate.

[0110] Table 1 Comparison of feature matching results of original images

[0111]

[0112] Table 2 Feature matching results when contrast is reduced by 60% and brightness is reduced by 60

[0113]

[0114] Table 3 Feature matching results when contrast is reduced by 60% and brightness is increased by 80

[0115]

[0116] The performance of the present invention shown in the above experiments has a higher improvement in the ability to extract feature points compared to other algorithms. In order to comprehensively evaluate the performance of the present invention, the extracted feature points are homogenized, and the density-aware homogenization method of the present invention and the quadtree homogenization method are compared. Figure 4 Experiments were conducted on the three representative scenarios shown in (i), and the corresponding qualitative and quantitative results are shown in Figure 5 And as shown in Table 4.

[0117] Table 4 Quantitative comparison of the present invention

[0118]

[0119] In simple scenarios, the experimental results show that the quadtree method improves the matching rate by eliminating a large number of feature points compared to the baseline, but the improvement is only 4.76%, which is still lower than the present invention, as shown in Table 4. This relatively low improvement can be attributed to the fact that the quadtree homogenization is not fine enough under the termination condition, so some valuable feature points have to be discarded. However, it should be emphasized that this termination condition does not always lead to insufficiently fine division. Figure 5 For example, the quadtree method retains more feature points than the present invention. However, by further analyzing the data in Table 6, although the quadtree method retains about 3.39 times more feature points than the present invention, its matching rate is only 0.2% higher than that of the present invention. Therefore, while reducing the computational complexity and storage overhead, the present invention can still maintain a high matching rate and effectively retain sufficient key points.

[0120] In addition, in view of the importance of model complexity under resource-constrained conditions, the time complexity and space complexity of the algorithm are analyzed in detail. The FAST algorithm exhibits the lowest time complexity with its direct feature point detection method. The space complexity of this method is also relatively low, and only the detected feature points need to be stored. In contrast, the AKAZE algorithm needs to store the entire scale space and feature descriptors when providing more complex scale and rotation invariance characteristics, resulting in high time complexity and extremely high space complexity. Through the analysis of the algorithm flow, it can be seen that the time complexity of the present invention in the texture-oriented feature extraction stage and the density-aware homogenization stage is O(MN+k) and O(k 2 ). Its spatial complexity is mainly determined by the feature points and the quadtree structure, which is about O(k). Therefore, the model complexity of the present invention is between FAST and AKAZE.

[0121] Taking into account the number of extracted feature points and the matching rate, although the present invention requires slightly more computing resources and storage space than FAST, it can extract more high-quality feature points. This shows that under resource-constrained conditions, the present invention has achieved a balance between performance and complexity.

[0122] As another example, in order to comprehensively evaluate the actual performance of the present invention in the field of VSLAM, experiments were carried out based on the ORB-SLAM3 framework

[31] , and the ORB-SLAM3 integrated with the present invention was called TOHF-SLAM. To ensure the reliability and statistical stability of the evaluation results, each data set was tested 5 times, and the average value of these results was used as the final evaluation result. Through multiple tests, we can more comprehensively understand the consistency of TOHF-SLAM in different data sets and situations, so as to more accurately evaluate its practical application value.

[0123] Table 7 Experimental results on different sub-datasets of EuRoc dataset

[0124]

[0125]

[0126] According to the data in Table 7, it can be seen that compared with the ORB-SLAM3 method, TOHF-SLAM has improved ATE on all sub-datasets. It is particularly noteworthy that in some specific data sets, such as MH01, the error reduction brought by TOHF-SLAM is more significant, reaching 68.42%. Overall, the average ATE of TOHF-SLAM has been significantly reduced compared to ORB-SLAM3, with a decrease of 37.50%.

[0127] On the MH05 dataset, TOHF-SLAM's performance advantage is particularly significant, and on each dataset, TOHF-SLAM shows the best performance, especially on V103 and V202, where the reprojection error is the smallest, the distribution is more compact, and the center value is lower. In terms of computational efficiency, the average processing time of TOHF-SLAM is 142.0 seconds, slightly higher than ORB-SLAM3's 140.5 seconds. However, this difference is acceptable in practical applications, and TOHF-SLAM can still meet the needs of real-time processing; this shows that TOHF-SLAM has not only improved in accuracy, but also maintained good performance in computational efficiency.

[0128] When conducting experiments on the Room dataset on resource-constrained devices, the complexity of the dataset poses a huge challenge to the VSLAM system. Figure 6 As shown in (a), ORB-SLAM3 fails to track because it cannot extract enough feature points. Figure 3 (a) and the feature point statistics in Table 1. The failure of feature extraction further affects the tracking and mapping capabilities of the system, resulting in a decrease in the performance of ORB-SLAM3 on this particular dataset. In contrast, TOHF-SLAM exhibits stronger robustness and effectiveness. Under the same resource-constrained conditions, TOHF-SLAM is not only able to track continuously, but also the trajectory it generates is basically consistent with the true trajectory measured manually. This shows that TOHF-SLAM has a clear advantage in tracking accuracy. In an environment with constrained resources, TOHF-SLAM exhibits better performance.

[0129] Exemplarily, the present invention provides a texture-oriented uniform feature extractor, which extracts feature points through a two-stage threshold strategy. In the initial stage, the contrast-sensitive channel of the human visual system (HVS) is simulated, and a basic threshold based on image contrast is formulated to adapt to feature point detection in various scenes. In particular, in the context of medium contrast, the texture-sensitive channel of the HVS is further simulated, and the threshold is fine-tuned in combination with information entropy to more effectively adapt to the texture complexity of the image.

[0130] Calculate the basic threshold T for the input image to be extracted:

[0131] C=Contrast(I) (1)

[0132] E=Entropy(I) (2)

[0133] T b '=f(C,E) ⑶

[0134] AdjustmentFactory(I) = F(ρ C ,ρ E ) (⑷)

[0135] T=T b '+AdjustmentFactory(I) ⑸

[0136] After calculating the basic threshold of the image to be extracted through the above formula, select candidate pixels:

[0137] center_pixel=I(x,y) ⑹

[0138] Select the pixels to compare:

[0139] Select 16 pixels on the circle around the center point (divided into 4 groups, each with 4 points); select 16 pixels on the circle with radius r, divided into 4 groups G 1 ,G 2 ,G 3 ,G 4 ,, 4 points in each group.

[0140] When entering the second stage, we consider that even at the same contrast and information entropy, different images may have slight differences. In order to capture these subtle differences, we borrowed the way HVS handles details and introduced the concept of optimized thresholds. This helps to capture the differences in details in the image more accurately.

[0141] Determine whether it is a key point based on the optimization threshold T:

[0142]

[0143] In Formula 7, 1 is an indicator function, which takes a value of 1 when the condition is met, otherwise it takes a value of 0. Only when count ≥ 12 is the center_pixel considered to be a feature point.

[0144] Non-maximum suppression:

[0145] Non-maximum suppression is performed on all detected feature points P, and only those feature points satisfying p>Neighbors(p) are retained, where p>Neighbors(p) refers to points within the neighborhood of feature point P.

[0146] Return the feature points to keypoints.

[0147] The extracted feature point map and feature point information are homogenized using a density-aware quadtree homogenization strategy.

[0148] First, calculate the feature point density distribution map D:

[0149] D←zeros(size(I))

[0150]

[0151] Here, |·| represents the size of the set, and dist is the distance function.

[0152] Partitioning is done using a density-aware quadtree strategy:

[0153] QuadtreeNode={children[4],range} ⑽

[0154] Root←QuadtreeNode(image_range(I)) (11)

[0155] If N node <threshold:

[0156] Save Node.range and feature point set, otherwise:

[0157] Divide Node.range into 4 sub-ranges and assign 4 child nodes to Node.

[0158] Filter feature points on the feature point map:

[0159] For each leaf node:

[0160] F leaf =select_by_response(Leaf.features) ⑿

[0161] FinalFeatures←FinalFeatures∪F leaf ⒀

[0162] The feature points are filtered using the density-aware quadtree strategy described above and then a feature point set is output.

[0163] In one embodiment, a computer-readable storage medium is provided, storing a computer program, wherein when the computer program is executed by a processor, the processor performs the following steps:

[0164] Step 101: construct a texture-oriented uniform feature extractor;

[0165] Step 102: obtaining a texture histogram of the image to be extracted using a texture-oriented uniform FAST feature extractor;

[0166] Step 103: Determine the contrast and information entropy of the texture histogram of the feature image to be extracted;

[0167] Step 104: presetting a basic threshold according to the contrast and information entropy;

[0168] Step 105: determining an optimized threshold according to the basic threshold, contrast weight and information entropy weight;

[0169] Step 106: extracting feature points of the image to be extracted using the optimized threshold and determining a feature point density distribution map;

[0170] Step 107: Dynamically adjust the quadtree division according to the density and the number of feature points and optimize the distribution structure of the feature points.

[0171] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0172] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0173] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A texture-oriented uniform FAST feature extraction method, characterized in that: The method comprises: Construct a texture-oriented uniform FAST feature extractor; Determine the contrast C and information entropy E of the feature image to be extracted; Presetting a basic threshold according to the contrast C and the information entropy E; Determine an optimization threshold according to the basic threshold, contrast weight and information entropy weight; Extracting feature points of the image to be extracted using the optimized threshold and determining a feature point density distribution map; Dynamically adjust the quadtree division according to the density and number of feature points and optimize the distribution structure of feature points; The step of presetting the basic threshold according to the contrast C and the information entropy E specifically includes: When it is determined that the contrast C is low, the basic threshold is set to C1', wherein the value of C1' is set to (0,3); When it is determined that the contrast C is high, the basic threshold is set to C2', wherein the value of C2' is set to (20, +∞); When the contrast C is determined to be medium, the basic threshold is set to T according to the information entropy E. b '; The step of determining the optimization threshold according to the basic threshold, contrast weight and information entropy weight specifically includes: according to T=T b ′+50×ρ c ×ρ E Determine the optimization threshold, where ρ C and ρ E They are positively correlated with image contrast C and information entropy E, When the contrast C is determined to be medium, the basic threshold T is determined according to the information entropy E. b ', specifically including: when determining that the information entropy E is low, the preset basic threshold T b ' is set to C'1 plus the first compensation value; when it is determined that the information entropy E is high, the preset basic threshold T b ' is set to C'1 plus a second compensation value, and the second compensation value is greater than the first compensation value.

2. The texture-oriented uniform FAST feature extraction method according to claim 1, characterized in that: The method further includes performing a dedistortion process on the image of the feature to be extracted to obtain a corrected image to be extracted: For each coordinate point in the image to be extracted, it is represented as x and y, and the coordinate point after dedistortion is represented as x distorted and distorted , the distorted coordinates are: x distorted =x(1+k1r 2 +k2r 4 +k3r 6 )+2p1xy+p2(r 2 +2x 2 ) <h2 style=";text-align:left;direction:ltr">y<h2 style=";text-align:left;direction:ltr"> distorted <h2 style=";text-align:left;direction:ltr"> =y(1+k1r<h2 style=";text-align:left;direction:ltr"> 2 <h2 style=";text-align:left;direction:ltr"> +k2r<h2 style=";text-align:left;direction:ltr"> 4 <h2 style=";text-align:left;direction:ltr"> +k3r<h2 style=";text-align:left;direction:ltr"> 6 <h2 style=";text-align:left;direction:ltr"> )+2p2xy+p1(r<h2 style=";text-align:left;direction:ltr"> 2 <h2 style=";text-align:left;direction:ltr"> +2y<h2 style=";text-align:left;direction:ltr"> 2 <h2 style=";text-align:left;direction:ltr"> ) Among them, k1, k2, k3, p1, and p2 are distortion parameters, x and y are the coordinates of the point on the normalized plane, and r is the distance between the point on the plane and the origin of the normalized plane.

3. The texture-oriented uniform FAST feature extraction method according to claim 1, characterized in that: The determining of the contrast C and information entropy E of the feature image to be extracted specifically includes: According to Determine the contrast C and information entropy E, where M and N are the sizes of the image to be extracted, representing the number of rows and columns of pixels respectively, and f(i, j) represents the grayscale value of the pixel with coordinates (i, j) in the image to be extracted.

4. The texture-oriented uniform FAST feature extraction method according to claim 1, characterized in that: The extracting feature points of the image to be extracted by the optimized threshold and determining the feature point density distribution map specifically includes: Determine the feature point density distribution map, where the value k at each position represents the specific area range S R The number of feature points within.

5. A texture-oriented uniform FAST feature extraction device, characterized in that: The device comprises: Building module for constructing texture-oriented uniform FAST feature extractor; The module for acquiring the image information to be extracted is used to determine the contrast C and information entropy E of the feature image to be extracted; specifically, when the contrast C is determined to be low, the basic threshold is set to C1', where the value of C1' is set to (0, 3); when the contrast C is determined to be high, the basic threshold is set to C2', where the value of C2' is set to (20, +∞); when the contrast C is determined to be medium, the basic threshold is set to T according to the information entropy E. b '; The optimization module is used to preset a basic threshold according to the contrast C and the information entropy E; and to determine an optimization threshold according to the basic threshold, the contrast weight and the information entropy weight; and is specifically used to determine an optimization threshold according to T=T b ′+50×ρ c ×ρ E Determine the optimization threshold, where ρ C and ρ E They are positively correlated with image contrast C and information entropy E, It is also specifically used to determine that when the information entropy E is low, the preset basic threshold T b ' is set to C'1 plus the first compensation value; when it is determined that the information entropy E is high, the preset basic threshold T b 'Set to C'1 plus the second compensation value, the second compensation value is greater than the first compensation value; The feature extraction module is used to extract the feature points of the picture to be extracted through the optimization threshold and determine the density distribution map of the feature points; it is also used to dynamically adjust the quadtree division according to the density and the number of feature points and optimize the distribution structure of the feature points.

6. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor is caused to perform the steps of the method according to any one of claims 1 to 4.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the method according to any one of claims 1 to 4.

Images