An Adaptive Feature Point Extraction Method and Device

Abstract

The present invention relates to an adaptive image feature point extraction method and device. First, the present invention acquires the image data collected by a camera, extracts feature points using the FAST algorithm, and then compares the extracted feature point parameters with a preset parameter threshold. If the feature point parameters are greater than or equal to the parameter threshold, the ORB feature extraction algorithm is automatically used to extract feature points from the image data. If the feature point parameters are less than the parameter threshold, the SURF feature point extraction algorithm is automatically used to extract feature points from the image data. The present invention reduces the feature point calculation amount and lowers the cost of the calculation platform.

Description

Technical Field

[0001] The present invention relates to the technical field of autonomous driving, and particularly to an adaptive feature point extraction method and device. Background Art

[0002] With the rapid development of intelligent networked vehicles in China, autonomous driving vehicles have been widely used in various industries. In some specific scenarios, such as port transport vehicles, logistics carts, park tour vehicles, road sweepers, automatic parking in underground garages, etc., they have been widely applied. A common feature of these scenarios is the low-speed scenario. Therefore, based on the vision method, extracting feature points in these scenarios is of great significance for constructing the map of the scenario, and the map construction based on vision feature points can reduce the dependence on the platform computing power, which is a very meaningful research direction. Summary of the Invention

[0003] The present invention provides an adaptive feature point extraction method and device for solving the technical problems existing in the prior art.

[0004] The technical solution of the present invention to solve the above technical problems is as follows:

[0005] On the one hand, the present invention provides an adaptive image feature point extraction method, which is characterized by including the following steps:

[0006] Obtain the image data collected by the camera;

[0007] Extract feature points using the FAST algorithm,

[0008] Compare the extracted feature point parameters with the preset parameter threshold;

[0009] If the feature point parameters are greater than or equal to the parameter threshold, automatically use the ORB feature extraction algorithm to extract feature points from the image data;

[0010] If the feature point parameters are less than the parameter threshold, automatically use the SURF feature point extraction algorithm to extract feature points from the image data.

[0011] Furthermore, the method further includes collecting feature points for various specific scenarios using the FAST algorithm to obtain the feature point empirical threshold, and generating a feature point empirical threshold table according to the feature point empirical threshold and the scenario information.

[0012] Furthermore, it also includes: when obtaining the image data collected by the camera, judging the scenario information of the image data collected by the camera;

[0013] When extracting feature points from the image data using the FAST algorithm, query the feature point empirical threshold table according to the scenario information to obtain the feature point parameter threshold corresponding to the scenario where the image data is located.

[0014] Further, the scene information includes scene location information. Based on the scene location information, the scene where the image data collected by the camera is located is determined, and then the feature point experience threshold table is queried to obtain the feature point parameter threshold corresponding to the scene where the image data is located.

[0015] Further, after the image data collected by the camera is obtained, Gaussian smoothing processing is performed on the image data.

[0016] On the other hand, the present invention also provides an adaptive image feature point extraction device, including:

[0017] A data acquisition module for acquiring the image data collected by the camera;

[0018] A first feature extraction module that extracts feature points using the FAST algorithm,

[0019] A comparison and judgment module that compares the extracted feature point parameters with the preset parameter thresholds;

[0020] A second feature extraction module for extracting feature points from the image data using the ORB feature extraction algorithm when the feature point parameters are greater than or equal to the parameter thresholds;

[0021] A third feature extraction module for extracting feature points from the image data using the SURF feature point extraction algorithm when the feature point parameters are less than the parameter thresholds.

[0022] Further, the device further includes a threshold table construction module that collects feature points for various specific scenes using the FAST algorithm to obtain feature point experience thresholds, and generates a feature point experience threshold table based on the feature point experience thresholds and the scene information.

[0023] Further, the data acquisition module is further configured to judge the scene information where the image data collected by the camera is located;

[0024] The first feature extraction module queries the feature point experience threshold table according to the scene information to obtain the feature point parameter threshold corresponding to the scene where the image data is located.

[0025] Further, the scene information includes scene location information. Based on the scene location information, the scene where the image data collected by the camera is located is determined, and then the feature point experience threshold table is queried to obtain the feature point parameter threshold corresponding to the scene where the image data is located.

[0026] Further, a preprocessing module for performing Gaussian smoothing processing on the image data.

[0027] The beneficial effects of the present invention are as follows: A trajectory map is constructed using machine vision. The key lies in extracting feature points from different picture frames. Due to different scenarios, the number of feature points existing in different captured scenarios and the difficulty of extracting feature points vary. This patent adopts an adaptive method to adaptively use different feature extraction methods in different scenarios. The present invention reduces the computational amount of feature points, lowers the cost of the computing platform, and improves the accuracy of feature extraction in different scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a flowchart of an adaptive image feature point extraction method provided by an embodiment of the present invention;

[0029] Figure 2 It is a schematic structural diagram of an adaptive image feature point extraction device provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0030] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only used to explain the present invention and are not intended to limit the scope of the present invention.

[0031] Figure 1 It is a flowchart of an adaptive image feature point extraction method provided by an embodiment of the present invention. As Figure 1 shown, the method of the present invention includes the following steps:

[0032] S1. Use the FAST algorithm to collect feature points for various specific scenarios, obtain the empirical threshold of feature points, and generate an empirical threshold table of feature points according to the empirical threshold of feature points and scene information; The specific scenarios can be underground garages, above-ground parking lots, ports, specific parks, etc. The scene information includes scene location information, specifically, it can be GPS information. Each scene in the empirical threshold table of feature points corresponds to a parameter threshold of feature points.

[0033] S2. Obtain the image data collected by the camera; Determine the scene information of the image data collected by the camera; When obtaining the image data collected by the camera, the GPS information of the camera can be obtained simultaneously, and the scene in which the camera is located can be determined through the GPS information.

[0034] S3. First perform Gaussian smoothing processing on the image data and then use the FAST algorithm to extract feature points. Query the empirical threshold table of feature points according to the scene information to obtain the parameter threshold of feature points corresponding to the scene where the image data is located; For example, when the GPS is located in an underground garage, the parameter threshold of feature points corresponding to the underground garage needs to be queried in the empirical threshold table of feature points.

[0035] Using the FAST algorithm to extract feature points includes:

[0036] 1. Select a pixel P from the image. Next, we will determine whether it is a feature point. First, set its brightness value to Ip;

[0037] 2. Set an appropriate threshold t;

[0038] 3. Consider a discretized Bresenham circle with a radius equal to 3 pixels centered on this pixel. There are 16 pixels on the boundary of this circle;

[0039] 4. If there are n consecutive pixel points on this circle of 16 pixels, and their pixel values are either all greater than Ip + t or all less than Ip - t, then it is a feature point;

[0040] 5. Save the feature points of the first frame.

[0041] S4. Compare the extracted feature point parameters with the preset parameter thresholds, that is, compare the extracted feature point parameters with the feature point parameter thresholds corresponding to the scene where the image data is located.

[0042] If the feature point parameters are greater than or equal to the parameter thresholds, automatically use the ORB feature extraction algorithm to extract feature points from the image data;

[0043] The ORB feature extraction algorithm first uses the FAST algorithm for feature point detection, and then performs BRIEF feature description. BRIEF is a binary descriptor, and its description vector consists of multiple 0s and 1s. Here, the 0s and 1s encode the size relationship between two pixels (such as p and q) near the feature point. BRIEF is exactly such an algorithm that generates feature descriptors based on binary coding and uses the Hamming distance for feature matching. Since BRIEF is just a feature descriptor, it is necessary to detect and locate feature points in advance. The FAST algorithm can be used to detect feature points. On this basis, the BRIEF algorithm is used to establish feature descriptors. Randomly select several point pairs (p, q) in the neighborhood of the feature point, and compare the gray values of these point pairs. If I(p) > I(q), then encode it as 1, otherwise encode it as 0. In this way, a binary coding string of a specific length can be obtained, that is, the BRIEF feature descriptor.

[0044] If the feature point parameters are less than the parameter thresholds, automatically use the SURF feature point extraction algorithm to extract feature points from the image data.

[0045] SURF feature point extraction algorithm:

[0046] First, use the Hessian matrix to detect feature points. This matrix is the second-order derivative matrix in the x and y directions, which can measure the local curvature of a function. The determinant value represents the variation around a pixel point, and the feature points are taken at the extreme points of the determinant value. Replace the Gaussian filter in SIFT with a square filter and use the integral image to calculate the values at the four corners of the filter square. Secondly, locate the feature points through interpolation of the neighboring information of the feature points. Then, calculate the Haar wavelet transform in the x and y directions of the pixel points around the feature points, and add the transformed values in the x and y directions within a certain angular range in the x-y plane to form a vector. The longest vector among all vectors (i.e., the one with the largest x and y components) is the direction of this feature point. Finally, after determining the direction of the feature point, the surrounding pixel points need to establish descriptors based on this direction. At this time, take 55 pixel points as a sub-region, and take the range of 2020 pixel points around the feature point, a total of 16 sub-regions. Calculate the total sum of the Haar wavelet transforms Σdx and Σdy in the x and y directions (at this time, the direction parallel to the feature point is x and the direction perpendicular to the feature point is y) within the sub-region, and the total sum of their vector lengths Σ|dx| and Σ|dy|, a total of four quantities, to generate a 64-dimensional descriptor.

[0047] Figure 2 FIG. is a schematic structural diagram of an adaptive image feature point extraction device provided by an embodiment of the present invention. As Figure 2 shown, the device includes:

[0048] A threshold table construction module, which uses the FAST algorithm to collect feature points in various specific scenarios, obtains the empirical threshold of feature points, and generates an empirical threshold table of feature points according to the empirical threshold of feature points and scene information; the scene information includes scene position information,

[0049] A data acquisition module, which is used to acquire the image data collected by the camera; and judge the scene information of the image data collected by the camera;

[0050] A preprocessing module, which is used to perform Gaussian smoothing processing on the image data;

[0051] A first feature extraction module, which uses the FAST algorithm to extract feature points, queries the empirical threshold table of feature points according to the scene information, and obtains the parameter threshold of feature points corresponding to the scene where the image data is located;

[0052] A comparison and judgment module, which compares the extracted feature point parameters with the preset parameter threshold; that is, compares the extracted feature point parameters with the parameter threshold of feature points corresponding to the scene where the image data is located;

[0053] A second feature extraction module, which is used to extract feature points from the image data by using the ORB feature extraction algorithm when the feature point parameters are greater than or equal to the parameter threshold;

[0054] A third feature extraction module, configured to perform feature point extraction on image data by using a SURF feature point extraction algorithm when a feature point parameter is less than a parameter threshold.

[0055] A trajectory map is constructed using machine vision. The key lies in extracting feature points from different picture frames. Depending on the scene, the number of feature points present in different captured scenes and the difficulty of feature point extraction vary. This patent adopts an adaptive method, adaptively using different feature extraction methods in different scenes. In terms of calculation speed, FAST >> ORB >> SURF (each differs by an order of magnitude), while in terms of robustness, SURF >> ORB >> FAST. Therefore, the method of the present invention reduces the feature point calculation amount and lowers the calculation platform cost; it improves the accuracy of feature extraction in different scenes.

[0056] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. An adaptive method for extracting image feature points, characterized in that, it includes the following steps: Using the FAST algorithm to collect feature points for various specific scenarios, obtaining the empirical threshold of feature points, and generating an empirical threshold table of feature points according to the empirical threshold of feature points and scene information; Obtaining the image data collected by the camera; when obtaining the image data collected by the camera, judging the scene information of the image data collected by the camera; Using the FAST algorithm to extract feature points; when using the FAST algorithm to extract feature points in the image data, querying the empirical threshold table of feature points according to the scene information to obtain the threshold of feature point parameters corresponding to the scene where the image data is located; Comparing the extracted feature point parameters with the threshold of feature point parameters corresponding to the scene where the image data is located; If the feature point parameters are greater than or equal to the parameter threshold, automatically using the ORB feature extraction algorithm to extract feature points from the image data; If the feature point parameters are less than the parameter threshold, automatically using the SURF feature point extraction algorithm to extract feature points from the image data.

2. The method for extracting image feature points according to claim 1, characterized in that, the scene information includes scene position information, judging the scene where the image data collected by the camera is located according to the scene position information, and then querying the empirical threshold table of feature points to obtain the threshold of feature point parameters corresponding to the scene where the image data is located.

3. The method for extracting image feature points according to claim 1, characterized in that, after obtaining the image data collected by the camera, performing Gaussian smoothing processing on the image data.

4. An adaptive device for extracting image feature points, characterized in that, it includes: A threshold table construction module, which uses the FAST algorithm to collect feature points for various specific scenarios, obtains the empirical threshold of feature points, and generates an empirical threshold table of feature points according to the empirical threshold of feature points and scene information; A data acquisition module, which is used to acquire the image data collected by the camera and is also used to judge the scene information of the image data collected by the camera; A first feature extraction module, which uses the FAST algorithm to extract feature points, queries the empirical threshold table of feature points according to the scene information, and obtains the threshold of feature point parameters corresponding to the scene where the image data is located; A comparison and judgment module, which compares the extracted feature point parameters with the threshold of feature point parameters corresponding to the scene where the image data is located; A second feature extraction module, which is used to extract feature points from the image data using the ORB feature extraction algorithm when the feature point parameters are greater than or equal to the parameter threshold; A third feature extraction module, which is used to extract feature points from the image data using the SURF feature point extraction algorithm when the feature point parameters are less than the parameter threshold.

5. The device for extracting image feature points according to claim 4, characterized in that, the scene information includes scene position information, judging the scene where the image data collected by the camera is located according to the scene position information, and then querying the empirical threshold table of feature points to obtain the threshold of feature point parameters corresponding to the scene where the image data is located.

6. The device for extracting image feature points according to claim 4, characterized in that, A preprocessing module, which is used to perform Gaussian smoothing processing on the image data.

Images