A polarization three-dimensional target recognition method based on the ORB operator

Abstract

The present invention discloses a polarization three-dimensional target recognition method based on the ORB operator, which relates to the technical field of target recognition. First, a matching object recognition library is established, in which the median filtering method and the method of delimiting a specific area are used to process the image; the ORB operator and the polarization gradient normal vector are used for object recognition. The requirements for equipment for polarization three-dimensional recognition are relatively low compared to three-dimensional imaging technologies such as laser scanning and binocular vision. The prior art consumes too much memory of the computer. The local feature matching recognition of the present invention can greatly reduce the calculation cost, thereby accelerating the recognition speed. Therefore, the present invention has high target recognition accuracy, and the amount of operation data of this method is small, which can improve the recognition speed.

Description

Technical Field

[0001] The present invention relates to the technical field of target recognition, and particularly relates to a polarization three-dimensional target recognition method based on ORB operators. Background Art

[0002] At present, the development prospect of three-dimensional recognition technology is relatively good, and there are the following two existing three-dimensional target recognition solutions:

[0003] Three-dimensional target recognition based on point cloud - Classical point cloud target recognition algorithms can identify obstacle targets from the point cloud data obtained by sensors such as lidar, and calculate data information such as their spatial positions, sizes, geometric shapes, and motion orientations, and describe the target contour information by fitting a bounding box or a polygon contour. The classical point cloud target recognition process includes processes such as point cloud preprocessing, region segmentation, obstacle clustering, object type recognition, and bounding box fitting. The overall process is relatively complex and requires multiple steps of processing such as segmenting, clustering, and classifying the point cloud.

[0004] This three-dimensional target recognition based on point cloud has the following defects: The overall process of the point cloud target recognition algorithm is relatively cumbersome and requires processing such as segmenting, clustering, and classification matching of the point cloud. In addition, due to the characteristics of incomplete target object data and low data density of the point cloud, classical point cloud target recognition algorithms can support the implementation of low-level perception tasks, but there are certain deficiencies in the target recognition accuracy and it is difficult to meet the requirements of the autonomous driving system for target recognition accuracy.

[0005] Three-dimensional recognition algorithm based on deep learning - Deep learning has a network structure with a high degree of abstraction ability, and can identify abstract patterns through a multi-layer convolutional neural network, and can solve the problems existing in traditional methods. At present, various neural networks with special neural processing units and complex structures emerge in an endless stream, such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), deep neural networks (DNNs), etc. It is a collection of algorithms that model high-complexity data through multiple non-linear transformations. Its essence is to use machine learning models with multiple hidden layers and a large amount of training data to represent and learn useful feature information as fully as possible. In the field of face recognition, good results have been achieved by performing deep learning on the rigid regions of faces, and higher accuracy has been achieved in the field of robotics.

[0006] The three-dimensional recognition algorithm based on deep learning has the following defects: When applying the three-dimensional recognition technology of deep learning methods to process the same data set, the larger the model scale, the higher the training accuracy, but the slower the training speed; the three-dimensional model library reserve is not rich. For the same image, different recognition results will be generated under the guidance of different models. The establishment of the model library not only requires the use of the object's own models, such as color, texture, shape, etc., but also requires models of the relationships between objects.

[0007] Therefore, there is currently no technical solution that can achieve target recognition using 3D imaging technology, and can solve the problem of low target recognition accuracy in the 3D target recognition solution based on point cloud, and at the same time solve the problems of large data volume and slow recognition speed in the 3D recognition algorithm based on deep learning. Summary of the Invention

[0008] In view of this, the present invention provides a polarization 3D target recognition method based on the orb operator, which can, on the basis of obtaining the normal vector of the object surface, utilize the feature matching of the orb operator to perform the function of target recognition on the object, with high target recognition accuracy, and the amount of operation data of this method is small, which can improve the recognition speed.

[0009] To achieve the above object, the technical solution of the present invention includes the following steps:

[0010] Step 1: Construct a matching database. The matching database contains polarization images of various objects. The standard polarization image is an image of the object surface obtained by a polarization camera. Each pixel point in the polarization image has normal vector gradient information; each polarization image in the matching database is divided into matching areas, and multiple matching areas are delimited for each polarization image.

[0011] Step 2: Obtain the image Y to be recognized, and select a polarization image Y' from the matching database.

[0012] Step 3: Match the two images Y and Y' through the orb algorithm to obtain the matching pairs of the two images. The two images have multiple matching pairs, and each matching pair contains the pixel point A from Y and the pixel point A' from Y'.

[0013] Step 4: Determine the matching area in the polarization image Y' where the pixel point A' in all the matching pairs is located. Select the matching area with more than three matching pairs in the polarization image Y' as the area to be recognized. If there is no matching area with more than three matching pairs, then select the next polarization image from the matching database as Y', and return to Step 3.

[0014] In the area S to be recognized j select the three matching pairs with the highest matching degree.

[0015] Determine the normal vector gradients n1, n2, n3 of the positions of the pixel points A from Y in the three matching pairs with the highest matching degree in the image Y to be recognized, and calculate the vectors respectively to obtain a first vector residual triangle composed of three vectors and obtain the three included angles θ1, θ2, θ3 in the first vector residual triangle.

[0016] Determine the normal vector gradient of the position of the pixel point A' from Y' among the three matching sub-areas with the highest matching degree in the polarization image Y'. Calculate the vectors respectively Obtain a second vector residual triangle composed of three vectors And obtain the three included angles within the second vector triangle as

[0017] For the area S to be recognized j For the corresponding two-vector residual triangle, calculate the matching loss

[0018] Step Five: For all areas to be recognized, the total loss after matching is as follows: N is the number of areas to be recognized.

[0019] If the total loss Loss is less than the critical value δ, it is determined that the images Y and Y' match successfully; otherwise, select another polarization image Y' from the matching database and return to Step Three.

[0020] For the successfully matched images Y and Y', the output target recognition result is that the target in the image Y to be recognized is the object in the polarization image Y'.

[0021] Preferably, in Step One, each polarization image in the matching database is divided into matching areas, and multiple matching areas are delimited for each polarization image. The specific division method includes the following steps: for an image of a certain object, use the orb operator to find all the feature points and calculate the corresponding descriptors; for all the descriptors, use the DBSCAN algorithm to find multiple clusters as the finally delimited multiple matching areas.

[0022] Beneficial effects:

[0023] 1. The present invention provides a polarization three-dimensional target recognition method based on the orb operator. First, a matching object recognition library is established, and the median filtering method and the method of delimiting specific areas are used to process the images; the orb operator and the polarization gradient normal vector are used for object recognition. The polarization three-dimensional recognition has lower requirements for equipment compared to three-dimensional imaging technologies such as laser scanning and binocular vision. The prior art consumes too much memory of the computer. The local feature matching recognition of the present invention can greatly reduce the calculation cost, thereby accelerating the recognition speed. Therefore, the present invention has high target recognition accuracy, and the amount of operation data of this method is small, which can improve the recognition speed. Specific embodiments

[0024] The present invention will be described in detail below in conjunction with the embodiments.

[0025] The present invention provides a polarized three-dimensional target recognition method based on the ORB operator, which is a method for target recognition using polarized three-dimensional images.

[0026] As an emerging and important technology, polarized three-dimensional reconstruction has the following advantages compared with traditional three-dimensional reconstruction methods because it can restore the convex information on the object surface: (1) higher reconstruction accuracy, (2) simple equipment, and (3) affordable price. Since the polarized three-dimensional reconstruction technology introduces the polarization degree information of the object, it can obtain richer object surface information, thereby obtaining higher reconstruction accuracy. The method of three-dimensional reconstruction using polarization information mainly obtains the azimuth angle and zenith angle of the object surface normal to determine the object surface normal by using the polarization characteristics of the object's reflected light, and finally performs discrete integration on the object surface normal to obtain the final three-dimensional image. In particular, the recognition based on polarized three-dimensional has lower requirements for equipment compared with three-dimensional imaging technologies such as laser scanning and binocular vision.

[0027] Therefore, using polarized three-dimensional images for target recognition can greatly reduce the computational cost, thereby accelerating the recognition speed.

[0028] A polarized three-dimensional target recognition method based on the ORB operator provided by the present invention includes the following steps:

[0029] Step 1: Construct a matching database. The matching database contains polarized images of various objects. The standard polarized image is an image of the object surface obtained by a polarized camera. Each pixel point in the polarized image has normal vector gradient information. Each polarized image in the matching database is divided into matching regions, and multiple matching regions are defined for each polarized image.

[0030] The method for dividing the matching regions is as follows:

[0031] For an image of a certain object, use the ORB operator to find all feature points and calculate the corresponding descriptors. For all the descriptors, use the DBSCAN algorithm to find multiple clusters, which are the finally defined multiple matching regions.

[0032] DBSCAN (Density-Based Spatial Clustering of Applications with Noise, density-based spatial clustering algorithm with noise) is a density-based spatial clustering algorithm. This algorithm divides regions with sufficient density into clusters and discovers clusters of arbitrary shapes in a spatial database with noise. It defines a cluster as the largest set of density-connected points.

[0033] The descriptors are carefully selected based on mathematical methods. The SIFT descriptor is among the best. First, it uses the distribution characteristics of the gradient directions of the pixels in the key point neighborhood to assign direction parameters to each key point, thus ensuring the rotation invariance and scale invariance of the feature points. Then, it statistically calculates the gradients in the local area centered on the feature point to generate a 128-dimensional gradient feature vector, and normalizes the feature vector to remove the influence of illumination. The feature points generated through the above steps have properties such as rotation invariance, scale invariance, and illumination invariance.

[0034] Step 2: Obtain the image Y to be recognized, and select a polarized image Y' from the matching database.

[0035] Step 3: Match the two images Y and Y' through the orb algorithm to obtain the matchers of the two images. There are multiple matchers for the two images, and each matcher contains a pixel point A from Y and a pixel point A' from Y'.

[0036] The specific method is as follows: For a pixel point p in the image to be recognized; set the threshold as t. When the difference in pixel gray values between p and another pixel point around it is greater than the threshold, the two points are considered different points; draw a circle O with P as the center and d as the radius. The pixel points passed by the circle O are selected as the surrounding pixel points, and generally 16 are selected; select the pixel point and the central pixel point; for these 16 pixel points, when the number of pixel points different from the central pixel point is greater than n, then judge the central point as a feature point; calculate the descriptor (each feature point corresponds to calculating one descriptor); match the two-dimensional descriptors of different images pairwise. When the matching value is greater than a certain threshold, the matching feature points are obtained. A pair of matching feature points belonging to two images is recorded as a matcher, and multiple matchers may be obtained for the two images.

[0037] Step 4: Determine the matching area where the pixel point A' in all matchers is located in the polarized image Y'. Select the matching area with more than three matchers in the polarized image Y' as the area to be recognized. If there is no matching area with more than three matchers, then select the next polarized image from the matching database as Y' and return to Step 3.

[0038] In the area S to be recognized j select the three matchers with the highest matching degree.

[0039] Determine the normal vector gradients n1, n2, and n3 of the positions of the pixel points A from Y in the three matchers with the highest matching degree in the image Y to be recognized, and calculate the vectors respectively to obtain a first vector residual triangle composed of three vectors and obtain the three included angles θ1, θ2, and θ3 in the first vector residual triangle.

[0040] Determine the normal vector gradient of the position of the pixel point A' from Y' among the three matching sub-areas with the highest matching degree in the polarization image Y'. Calculate the vectors respectively Obtain a second vector residual triangle composed of three vectors And obtain the three included angles within the second vector triangle as respectively

[0041] For the area S to be recognized j For the corresponding two-vector residual triangle, calculate the matching loss

[0042] Step Five: For all areas to be recognized, the total loss after matching is as follows: N is the number of areas to be recognized.

[0043] If the total loss Loss is less than the critical value δ, it is determined that the images Y and Y' match successfully; otherwise, another polarization image Y' is selected from the matching database, and return to Step Three.

[0044] For the successfully matched images Y and Y', the output target recognition result is that the target in the image Y to be recognized is the object in the polarization image Y'.

[0045] The present invention aims to achieve target recognition by using three-dimensional imaging technology, and has the advantages of low memory consumption of the algorithm, high real-time response, and easy implementation of the principle.

[0046] The present invention can also be used in the aspect of part defect detection, and determine whether a defect occurs at the corresponding location by detecting the surface information of the key point defects of the part.

[0047] In summary, the above are only the preferred embodiments of the present invention, and are not used to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A polarization three-dimensional target recognition method based on the ORB operator, characterized in that It includes the following steps: Step 1: Construct a matching database, which contains polarization images of various objects. The standard polarization image is an image of the object surface obtained by a polarization camera, and each pixel point in the polarization image has normal vector gradient information. Each polarization image in the matching database is divided into matching regions, and multiple matching regions are defined for each polarization image. Step 2: Obtain the image Y to be recognized, and select a polarization image Y' from the matching database. Step 3: Match the two images Y and Y' through the orb algorithm to obtain the matching pairs of the two images. There are multiple matching pairs for the two images, and each matching pair contains a pixel point A from Y and a pixel point A' from Y'. Step 4: Determine the matching region in the polarization image Y' where the pixel point A' is located among all the matching pairs. Select the matching region with more than three matching pairs in the polarization image Y' as the region to be recognized. If there is no matching region with more than three matching pairs, then select the next polarization image from the matching database as Y' and return to Step 3. In the area S to be recognized j select the three matching sub-elements with the highest matching degree; Determine the normal vector gradients n1, n2, and n3 of the position of the pixel point A from Y among the three matching sub-areas with the highest matching degree in the image Y to be recognized, and calculate the vectors respectively to obtain a first vector residual triangle composed of three vectors and obtain three included angles θ1, θ2, and θ3 within the first vector residual triangle; The normal vector gradient of the position of the pixel point A' from Y' among the three matching sub-areas with the highest matching degree in the polarization image Y' Calculate the vectors respectively Obtain a second vector residual triangle composed of three vectors And obtain the three included angles within the second vector triangle, which are respectively For the region S to be recognized j Compute the matching loss for the corresponding two-vector residual triangle Step Five: For all the areas to be recognized, the total loss after matching is as follows: N is the number of areas to be recognized; If the total loss Loss is less than the critical value δ, it is determined that the images Y and Y' are successfully matched; otherwise, select another polarization image Y' from the matching database and return to Step 3. For the successfully matched images Y and Y', the output target recognition result is that the target in the image Y to be recognized is the object in the polarization image Y'.

2. The polarization three-dimensional target recognition method based on the ORB operator according to claim 1, characterized in that In Step 1, each polarization image in the matching database is divided into matching regions, and multiple matching regions are defined for each polarization image. The specific division method includes the following steps: For the image of a certain object, use the orb operator to find all the feature points and calculate the corresponding descriptors. For all the descriptors, use the DBSCAN algorithm to find multiple clusters, which are the finally defined multiple matching regions.

Images