A target texture recognition method, device, recognition equipment and storage medium

Abstract

The embodiment of the present invention relates to a method, apparatus, recognition device and storage medium for target texture recognition, the method comprising: obtaining brightness data and extracting a feature vector from the brightness data; evaluating the feature vector using an offline module to obtain a proximity coefficient; performing spatial frequency filtering on the proximity coefficient; judging whether the feature vector is a target feature based on the proximity coefficient after the spatial frequency filtering, wherein the target feature corresponds to the feature template of the target; if so, calculating the position and direction of the target feature in the brightness data. There is no need to temporarily change the radar's sampling rate, rotation speed and other working states, avoiding data interference caused by switching states, being able to effectively identify target textures, and reducing the processing cost of targets.

Description

Technical Field

[0001] The embodiments of the present invention relate to the technical field of recognition equipment, and in particular to a target texture recognition method, device, recognition equipment and storage medium. Background Art

[0002] With the development of SLAM (Simultaneous Localization And Mapping) technology, lidar has been widely used in the control and navigation systems of mobile robots. In order to have a comprehensive and multi-dimensional perception of the surrounding environment, different types of sensors such as infrared ranging, cameras, IMU and GPS are also used in the control and navigation systems of mobile robots to improve the accuracy of the overall system.

[0003] However, when adding additional sensors to consumer-grade robot products (such as smart sweeping robots and drones), they are often subject to many limitations such as mold design and platform computing power. By using the additional brightness information output by existing laser radars for target recognition, similar functions can be achieved without adding additional equipment, which can effectively control production costs. Summary of the invention

[0004] The purpose of the embodiments of the present invention is to provide a target texture recognition method, device, recognition equipment and storage medium, which do not require temporary changes to the radar's sampling rate, rotation speed and other working states, avoid data interference caused by switching states, can effectively recognize target textures, and reduce the processing cost of targets.

[0005] In a first aspect, an embodiment of the present invention provides a method for identifying target texture, the method comprising:

[0006] Acquire brightness data, and extract feature vectors from the brightness data;

[0007] Using an offline module to evaluate the feature vector to obtain an approximation coefficient;

[0008] Performing spatial frequency filtering on the approximation coefficient;

[0009] Based on the approximation coefficient after the spatial frequency filtering process, determining whether the feature vector is a target feature, wherein the target feature corresponds to a feature template of a target;

[0010] If so, the position and direction of the target feature in the brightness data are calculated.

[0011] In some embodiments, the using an offline module to evaluate the feature vector to obtain a similarity coefficient includes:

[0012] Calculating the Euclidean distance or direction cosine between the feature vector and the feature template to obtain an original evaluation parameter;

[0013] After normalizing the original evaluation parameters, the first approximation coefficient is obtained.

[0014] In some embodiments, the using an offline module to evaluate the feature vector to obtain a similarity coefficient includes:

[0015] Using a pre-trained feature classifier to evaluate the feature vector to obtain an original cost coefficient;

[0016] After normalizing the original cost coefficient, a second approximation coefficient is obtained.

[0017] In some embodiments, the method further comprises:

[0018] Move the recognition device to a preset sampling position and collect radar point cloud data of the preset sampling position, wherein the preset sampling position is distributed within a preset sampling area, and the radar point cloud data includes position and brightness information;

[0019] Determining radar point cloud data corresponding to the target position in the radar point cloud data according to the position information in the point cloud data and the position information of the target;

[0020] Extract features based on brightness information in the radar point cloud data corresponding to the target position to obtain feature positive samples;

[0021] Obtaining negative sample data other than the radar point cloud data corresponding to the target position in the radar point cloud data, and performing feature extraction on the negative sample data to obtain a plurality of feature negative samples;

[0022] The feature template is obtained according to the feature positive sample and the feature negative sample.

[0023] In some embodiments, after obtaining the feature template, the method further includes:

[0024] Based on the feature positive samples and the labels corresponding to the feature positive samples, and the feature negative samples and the labels corresponding to the feature negative samples, a classifier model is trained to obtain the feature classifier.

[0025] In some embodiments, the classifier model is W T xb, training a classifier model based on the feature positive sample and the label corresponding to the feature positive sample, and the feature negative sample and the label corresponding to the feature negative sample to obtain the feature classifier, including:

[0026] Input the feature positive samples and the labels corresponding to the feature positive samples, and the feature negative samples and the labels corresponding to the feature negative samples into the classifier model one by one, and obtain the parameters W and b of the classifier model with the goal of optimizing the cost function;

[0027] Wherein, the optimization cost function is ∑1-y k (W T -b).

[0028] In some embodiments, after moving the recognition device to a preset sampling position and collecting radar point cloud data of the preset sampling position, the method further includes:

[0029] Correcting the position in the radar point cloud data based on a preset map;

[0030] A data segment of interest is extracted from the radar point cloud data.

[0031] In some embodiments, after acquiring the brightness data, the method further includes:

[0032] The real-time brightness data is processed with a sliding window of dynamic length to intercept a section of local brightness data.

[0033] In a second aspect, an embodiment of the present invention further provides a target texture recognition device, the device comprising:

[0034] A feature vector acquisition module, used to acquire brightness data and extract feature vectors from the brightness data;

[0035] An evaluation module, used for evaluating the feature vector using an offline module to obtain an approximation coefficient;

[0036] A filtering module, used for performing spatial frequency filtering on the approximation coefficient;

[0037] A judgment module, used for judging whether the feature vector is a target feature based on the approximation coefficient after the spatial frequency filtering process, wherein the target feature corresponds to a feature template of a target;

[0038] A calculation module is used to calculate the position and direction of the target feature in the brightness data.

[0039] In a third aspect, an embodiment of the present invention further provides an identification device, the identification device comprising:

[0040] at least one processor, and

[0041] A memory, wherein the memory is communicatively connected to the at least one processor, the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can perform the above method.

[0042] In a fourth aspect, an embodiment of the present invention further provides a non-volatile computer-readable storage medium, wherein the computer-readable storage medium stores computer-executable instructions. When the computer-executable instructions are executed by a recognition device, the recognition device executes the method described above.

[0043] In a fifth aspect, an embodiment of the present invention further provides a computer program product, which includes a computer program stored on a non-volatile computer-readable storage medium, and the computer program includes program instructions. When the program instructions are executed by an identification device, the identification device executes the above method.

[0044] Compared with the prior art, the present application has at least the following beneficial effects: the target texture recognition method, apparatus, recognition device and storage medium of the embodiments of the present invention obtain brightness data without temporarily changing the sampling rate of the brightness data collected by the radar; and the feature vector extracted from the brightness data can better describe the characteristics of the target and improve the stability of recognition; the feature vector is evaluated by an offline module to obtain an approximation coefficient, and the approximation coefficient can preliminarily judge the recognition result; in order to reduce the influence of noise and abnormal data in the approximation coefficient, the approximation coefficient is subjected to spatial frequency filtering, and based on the approximation coefficient after the spatial frequency filtering, it is judged whether the feature vector is a target feature, wherein the target feature corresponds to the feature template of the target, and when the feature vector is a target feature, the position and direction of the target feature in the brightness data are calculated, which can effectively suppress the disturbance caused by the specific data and improve the accuracy of target texture recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] One or more embodiments are exemplarily described by pictures in the corresponding drawings, and these exemplified descriptions do not constitute limitations on the embodiments. Elements with the same reference numerals in the drawings represent similar elements, and unless otherwise stated, the figures in the drawings do not constitute proportional limitations.

[0046] Figure 1 is a schematic structural diagram of an embodiment of an identification device of the present invention;

[0047] Figure 2 It is a flow chart of an embodiment of the target texture recognition method of the present invention;

[0048] Figure 3is a schematic diagram of preset sampling positions of an embodiment of a target texture recognition method of the present invention;

[0049] Figure 4 It is a schematic diagram of a characteristic template of a target in one embodiment of the target texture recognition method of the present invention;

[0050] Figure 5A is a schematic diagram of a target of an embodiment of a method for identifying target texture of the present invention;

[0051] Figure 5B yes Figure 5A Sample diagram of corresponding brightness;

[0052] Figure 6 is a waveform diagram of a similarity coefficient of an embodiment of a target texture recognition method of the present invention;

[0053] Figure 7 It is a comparison diagram of radar point cloud data before and after alignment at a preset sampling position of an embodiment of a target texture recognition method of the present invention;

[0054] Figure 8 is a comparison diagram of the waveform of the approximation coefficient of an embodiment of the target texture recognition method of the present invention before and after filtering;

[0055] Fig. 9 It is a schematic structural diagram of an embodiment of a target texture recognition device of the present invention. DETAILED DESCRIPTION

[0056] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several variations and improvements may be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0057] The terms "first", "second", etc. in the specification and claims of the present application are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable under appropriate circumstances, so that the embodiments of the present application can be implemented in an order other than those illustrated or described here, and the objects distinguished by "first", "second", etc. are generally of one type, and the number of objects is not limited. For example, the first object can be one or more. In addition, "and / or" in the specification and claims represents at least one of the connected objects, and the character " / " generally indicates that the objects associated with each other are in an "or" relationship.

[0058] The recognition device can be a robot, such as an intelligent sweeping robot, a drone, etc. The recognition device is equipped with a laser radar. There is no need to temporarily change the working state of the radar, such as the sampling rate and rotation speed, to avoid data interference caused by switching states, and the target texture can be effectively recognized, and the processing cost of the target can be reduced.

[0059] The identification device may be a device with a vision-based distance measurement navigation system, but is not limited thereto.

[0060] Figure 1 The hardware structure of the identification device 100 is schematically shown. Figure 1 As shown, the recognition device 100 includes a memory 11 and a processor 12, and the processor 12 is connected to the memory 11. Those skilled in the art can understand that Figure 1 The structure shown does not constitute a limitation on the identification device. The identification device may include more or fewer components than shown in the figure, or combine or separate some components, or arrange the components differently.

[0061] Among them, the memory 11, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer executable programs and modules. The memory 11 may include a program storage area and a data storage area, wherein the program storage area can store an operating system, at least one application required for a function; the data storage area can store data created according to the use of the identification device, etc. In addition, the memory 11 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory 11 may optionally include a memory remotely arranged relative to the processor 12, and these remote memories may be connected to the identification device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0062] The processor 12 is the control center of the recognition device, and uses various interfaces and lines to connect various parts of the entire recognition device. By running or executing software programs and / or modules stored in the memory 11 and calling data stored in the memory 11, it executes various functions of the recognition device and processes data, thereby monitoring the recognition device as a whole, such as realizing the target texture recognition method described in any embodiment of the present invention.

[0063] The processor 12 may be one or more, Figure 1 A processor 12 is taken as an example. The processor 12 and the memory 11 may be connected via a bus or other means. Figure 1In the example, the connection through the bus is taken. The processor 12 may include a central processing unit (CPU), a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a controller, a field programmable gate array (FPGA) device, etc. The processor 12 may also be implemented as a combination of computing devices, for example, a combination of a DSP and a microprocessor, a plurality of microprocessors, one or more microprocessors combined with a DSP core, or any other such configuration.

[0064] The target texture recognition method, apparatus, recognition device and storage medium provided by the embodiments of the present invention can be executed by the recognition device 100. Specifically, in some embodiments, the method can be executed by the processor 12.

[0065] Figure 2 A schematic diagram of a target texture recognition method provided by an embodiment of the present invention is shown in FIG. Figure 2 As shown, the method includes:

[0066] 101: Acquire brightness data, and extract a feature vector from the brightness data.

[0067] When performing target texture recognition, the recognition device can obtain real-time brightness data through the laser radar, and intercept the real-time brightness data to obtain a section of local brightness data.

[0068] In this case, there is no limitation on the surface structure of the target texture. The brightness information obtained when the laser radar samples certain specific texture structures presents significant frequency characteristics. Therefore, the texture design can be appropriately adjusted while ensuring the spatial frequency characteristics, such as using black and white stripes, gradient colors, changing texture size or shape, etc.

[0069] In some embodiments, before acquiring brightness data, the method may further include:

[0070] The real-time brightness data is processed with a sliding window of dynamic length to intercept a section of local brightness data.

[0071] Specifically, when the recognition device recognizes the target texture online, it needs to deal with a continuous data stream. Therefore, it is necessary to set a buffer of a certain length and a data sliding window. The buffer is used to record historical brightness data, and the sliding window is used to intercept the brightness sampling data that may correspond to the potential target. That is, the real-time brightness data is processed by a sliding window of dynamic length to obtain the said section of local brightness data.

[0072] It is understandable that in order to meet the needs of different application scenarios, the length of the buffer should be dynamically adjusted according to the current radar operating status to ensure that the window length can cover the data segment corresponding to the entire potential target.

[0073] After obtaining a section of local brightness data, the local brightness data is preprocessed and then subjected to feature extraction to obtain a set of feature vectors.

[0074] Correspondingly, the local brightness data is first preprocessed, and then a common sequence feature extraction algorithm can be used to calculate the feature vector. For example, in one embodiment, the feature vector can be constructed by using juxtaposed and rearranged image moments.

[0075] 102: Evaluate the feature vector using an offline module to obtain an approximation coefficient.

[0076] When identifying the target texture, the feature vector can be evaluated using an offline module, which includes a feature template obtained in the offline processing stage and a trained feature classifier. A low computing power method can be used for evaluation, that is, the feature vector is compared and calculated with the feature template to obtain a first approximation coefficient, and the feature template is obtained by an offline learning method based on a small sample. A high computing power method can also be used, that is, the feature vector is evaluated using a pre-trained feature classifier to obtain an original cost coefficient; after normalizing the original cost coefficient, a second approximation coefficient is obtained.

[0077] Furthermore, when intercepting the real-time brightness data, it is necessary to set a suitable interception interval according to the sampling position. In some embodiments, the acquisition of the feature template may include:

[0078] Move the recognition device to a preset sampling position and collect radar point cloud data of the preset sampling position, wherein the preset sampling position is distributed within a preset sampling area, and the radar point cloud data includes position and brightness information;

[0079] Determining radar point cloud data corresponding to the target position in the radar point cloud data according to the position information in the point cloud data and the position information of the target;

[0080] Extract features based on brightness information in the radar point cloud data corresponding to the target position to obtain feature positive samples;

[0081] Obtaining negative sample data other than the radar point cloud data corresponding to the target position in the radar point cloud data, and performing feature extraction on the negative sample data to obtain a plurality of feature negative samples;

[0082] The feature template is obtained according to the feature positive sample and the feature negative sample.

[0083] Specifically, the recognition device is moved to a preset sampling position, and radar point cloud data of the preset sampling position is collected. The preset sampling position can be evenly distributed in the preset sampling area, or the preset sampling position is not distributed in the preset sampling area. The setting of the preset sampling position is not limited here. The radar point cloud data includes position and brightness information. The preset sampling area refers to an area where the laser radar is expected to have a high detection rate for the target. It can be set according to the adaptability of the application scenario. It can be a sector, rectangle, or other area with regular shape and symmetry. It can also be an open area or a closed area. It is not limited here.

[0084] There are no special requirements for the setting of preset sampling positions. You can adopt a layout that covers different distances and angles. Depending on the distance of the target and the radar sampling rate, the number of sampling positions obtained on the same target is different, such as Figure 3 As shown, it includes sampling position A, sampling position B and sampling position C. The number of lines represents the number of sampling points, that is, the number of scanning lines. The radar obtains data one by one through scanning lines. Under the condition of a constant rotation speed, the closer to the scanning line, the denser the target area that can be scanned. Conversely, the farther from the scanning line, the sparser the scanned data. Therefore, sampling position A is farthest from the target, and the scanned data is relatively sparse. Sampling position C is closest to the target, and the scanned data is relatively dense.

[0085] Therefore, sampling can be performed at sampling position A, sampling position B and sampling position C respectively to obtain the corresponding radar point cloud data, and these different numbers of points will be interpolated and resampled to obtain a fixed length of brightness information. For example: the radar point cloud data of the distant sampling position A has few points, only 24; the radar point cloud data of the nearby sampling position C has more points, 52. After interpolation and resampling, they will be normalized into 32 points, so that the information at different distances will be comparable.

[0086] Then, according to the position information in the point cloud data and the position information of the target, the radar point cloud data corresponding to the position of the target in the radar point cloud data is determined. That is, the radar point cloud data at the preset sampling position includes radar point cloud data other than the target, and the radar point cloud data of the target in the radar point cloud data can be preliminarily determined by the position of the radar point cloud data at the preset sampling position and the position of the target.

[0087] Finally, feature extraction is performed based on the brightness information in the radar point cloud data corresponding to the target position to obtain a feature positive sample; negative sample data other than the radar point cloud data corresponding to the target position is obtained from the radar point cloud data, and feature extraction is performed on the negative sample data to obtain multiple feature negative samples; the feature template is obtained based on the feature positive sample and the feature negative sample. Figure 4 As shown, Figure 4 Schematic diagram of the target feature template.

[0088] The target may be a structure corresponding to the target texture, such as a black and white equidistant stripe structure, such as Figure 5A As shown, Figure 5A is a schematic diagram of the target. Figure 5B For Figure 5A Sample diagram of corresponding brightness. The feature template is used for subsequent comparison calculation to obtain the approximation coefficient.

[0089] In a low computing power example, the calculation of the approximation coefficient includes two steps: first, the Euclidean distance between the feature vector and the feature template is calculated, and then the Euclidean distance is mapped to a value between 0 and 1 using an exponential function, and the value is regarded as the first approximation coefficient. The exponential function is not a unique model, and a one-to-one monotonic mapping in a general sense can achieve a similar effect.

[0090] The characteristic vector is compared with the characteristic template, and the waveform corresponding to the obtained approximation coefficient is as follows: Figure 6 As shown. Figure 6 It is obvious that the waveforms of the two stages of 100-200 and 600-800 are relatively regular, and can basically be determined as the target features corresponding to the target to be measured.

[0091] In addition to using a low-computing method to compare and calculate a set of feature vectors with the pre-acquired target feature template to obtain the original approximation coefficient, a high-computing method can also be used to obtain the approximation coefficient. In the high-computing model, the calculation method of the approximation coefficient is mainly determined by the classifier model. In a typical example based on SVM (Support Vector Machine), the output of the classifier can also be normalized to the Euclidean distance of the high-dimensional data point corresponding to the feature vector in the hyperplane relative to the segmentation boundary. Similarly, the Euclidean distance can also be mapped to a numerical range of 0-1, thereby serving as an expression of the approximation coefficient.

[0092] In some embodiments, the method may further include:

[0093] Using a feature classifier to evaluate the feature vector to obtain an original approximation coefficient;

[0094] After normalizing the original cost coefficient, a second approximation coefficient is obtained.

[0095] Specifically, when a high-computing-power classifier is used for calculation, real-time brightness data is acquired, and the real-time brightness data is dynamically intercepted to obtain a section of local brightness data; the local brightness data is preprocessed and feature extracted to obtain a set of feature vectors; the feature vectors are evaluated using a feature classifier to obtain an original approximation coefficient; the original approximation coefficient is subjected to spatial frequency filtering; based on the approximation coefficient after the spatial frequency filtering, it is determined whether the feature vector is a target feature, wherein the target feature corresponds to a feature template of the target; if so, the position and direction of the target feature in the brightness data are calculated.

[0096] Compared with the low-computing power method, the high-computing power method uses the SVM classifier model for recognition. Similarly, combined with spatial filtering processing, it can effectively suppress the disturbance caused by specific data and improve the accuracy of target texture recognition.

[0097] Of course, in this case, the evaluation of the feature vector is not limited to the high computing power or low computing power method, and the feature vector can be evaluated by both methods at the same time, which is not limited here.

[0098] In some embodiments, in order to obtain the feature classifier, after obtaining the feature template, the method may further include:

[0099] Based on the feature positive samples and the labels corresponding to the feature positive samples, and the feature negative samples and the labels corresponding to the feature negative samples, a classifier model is trained to obtain the feature classifier.

[0100] Specifically, similar to obtaining the characteristic template of the target, first move the recognition device to the preset sampling position and collect the radar point cloud data of the preset sampling position, the preset sampling position is distributed in the preset sampling area (not limited to the distribution of the preset sampling position here), and the radar point cloud data includes position and brightness information; according to the position and the position of the target, determine the radar point cloud data of the target in the radar point cloud data; then, based on the brightness information in the radar point cloud data of the target, perform feature extraction to obtain a feature positive sample, obtain negative sample data in the radar point cloud data except the radar point cloud data of the target, perform feature extraction on the negative sample data, and obtain multiple feature negative samples. Feature extraction can be performed using a common sequence feature extraction algorithm. The radar point cloud data of the target corresponds to the target position, which is the feature positive sample x P ={x Pi}, i = 1…N; the corresponding non-target position in the radar point cloud data except the radar point cloud data of the target is the feature negative sample x N ={x Ni}, j = 1…M; Different from the feature template for obtaining the target, the classifier model is trained based on the feature positive sample and the label corresponding to the feature positive sample, and the feature negative sample and the label corresponding to the feature negative sample to obtain the feature classifier. The label corresponding to the feature positive sample is set to 1, and the label corresponding to the feature negative sample is set to -1. The feature vector and the label are combined into a binary tuple {x Ck ,y Ck}, C∈{P,N}, k∈{i,j}, and the classifier model can be W T xb, specifically, the feature positive sample and the label corresponding to the feature positive sample, and the feature negative sample and the label corresponding to the feature negative sample are input into the classifier model one by one, with the optimization cost function as the goal, to obtain the parameters W and b of the classifier model; wherein the optimization cost function is ∑[1-y(W T xb)].

[0101] In some instances of classifier implementation, some additional processing steps can be introduced, such as performing various transformations on sample data to augment the data set, performing PCA projection on features, introducing kernel functions for nonlinear segmentation, etc. Of course, these steps should be regarded as minor improvements on the original classifier rather than fundamental distinctions.

[0102] In some embodiments, after moving the recognition device to a preset sampling position and collecting radar point cloud data of the preset sampling position, the method further includes:

[0103] Correcting the position in the radar point cloud data based on a preset map;

[0104] A data segment of interest is extracted from the radar point cloud data.

[0105] Specifically, in the offline processing stage, preprocessing is also performed. The preprocessing is similar to the preprocessing process of the local brightness data, specifically, the alignment and interception of the original data. After the recognition device is moved to the preset sampling position and the radar point cloud data of the preset sampling position is collected, the position in the radar point cloud data is corrected based on the preset map. For example, if the radar point cloud data of the preset sampling position is not aligned with the corresponding position in the preset map, the radar point cloud data of the preset sampling position can be corrected and aligned, such as Figure 7 As shown, Figure 7The comparison diagram of the radar point cloud data at the preset sampling position before and after alignment is shown, wherein the outer frame with scale is the coordinate system of the preset map, and the radar point cloud data at the preset sampling position is on the coordinate system. By correcting the position in the radar point cloud data based on the preset map, the image obtained by the laser radar is aligned with the actual scene. Then, the data segment of interest is intercepted from the radar point cloud data, and the data segment of interest can be the data segment concave in the upper part of the figure.

[0106] When intercepting, you can use the method of manually setting the range, or you can adaptively complete it according to the location information of the map. Since feature extraction and other operations will be performed later, there is no need to set too high requirements for the accuracy of data interception here. Figure 7 As shown, the radar point cloud data of interest is the upper concave part, which is intercepted as the radar point cloud data of interest. In addition, the radar point cloud data of interest includes brightness information data corresponding to the target texture.

[0107] 103: Perform spatial frequency filtering on the approximation coefficient.

[0108] 104: Based on the approximation coefficient after the spatial frequency filtering process, determine whether the feature vector is a target feature, wherein the target feature corresponds to a feature template of a target.

[0109] 105: If yes, calculate the position and direction of the target feature in the brightness data.

[0110] When the waveform of the approximation coefficient is obtained, it can be preliminarily determined whether there is a target feature corresponding to the target to be measured. However, due to the influence of noise and abnormal data, if the original approximation coefficient is directly used to determine whether the target exists, the result lacks stability. Therefore, it is necessary to further perform spatial frequency filtering on the approximation coefficient.

[0111] Specifically, by agreeing on the target texture, the waveform of the approximation coefficient can form a specific frequency pattern near the target signal, and this part of the signal can generate a more stable response after being processed by the bandpass filter. Therefore, judging whether the target exists or not according to the response after the bandpass filter can effectively suppress the disturbance of the signal itself and improve the accuracy of the judgment result. The parameters of the bandpass filter should be calculated according to the physical size, and can also be directly estimated according to the sampling data.

[0112] In one embodiment, the target texture is set to an equal-length staggered pattern of three white segments and two black segments, so the corresponding original brightness signal will form a "mountain"-shaped waveform. When a sliding window passes through this segment of the signal, the original proximity coefficient between the feature vector of the intercepted signal and the feature template will form a high-low staggered waveform. The reason is that when the characteristic waveforms of the two are partially aligned, a larger proximity response will be generated, and conversely, a very small response will be generated when they are partially staggered. And because the texture itself is of equal length intervals, the intervals between the large and small responses should also be of equal length. From the perspective of a continuous signal, it can be understood that the proximity waveform will produce high and low oscillations of a specific frequency on this segment. For example Figure 8 As shown in FIG. 1 , it is a comparison diagram of the waveform of the approximation coefficient before and after filtering processing. It is obvious that after filtering processing, the waveform at the target feature corresponding to the target to be measured is strengthened, which can effectively suppress the disturbance caused by specific data and improve the recognition accuracy.

[0113] After determining the target feature, the position and direction of the target feature in the brightness data are calculated. That is, according to the position of a section of local brightness data intercepted in the real-time brightness data, the position and direction of the target texture on the map corresponding to the target feature corresponding to the local brightness data can be further calculated. Specifically for plane targets, the position and direction of the target in the brightness data are the coordinates of the center point of the target and the tangent / normal vector.

[0114] The target texture recognition method of the embodiment of the present invention obtains brightness data without temporarily changing the sampling rate of the brightness data collected by the radar; and extracting a feature vector from the brightness data can better describe the characteristics of the target and improve the stability of recognition; the feature vector is evaluated by an offline module to obtain an approximation coefficient, and the approximation coefficient can preliminarily judge the recognition result; in order to reduce the influence of noise and abnormal data in the approximation coefficient, the approximation coefficient is subjected to spatial frequency filtering processing, and based on the approximation coefficient after the spatial frequency filtering processing, it is judged whether the feature vector is a target feature, wherein the target feature corresponds to a feature template of the target, and when the feature vector is a target feature, the position and direction of the target feature in the brightness data are calculated, which can effectively suppress the disturbance caused by the specific data and improve the accuracy of target texture recognition.

[0115] Compared with the prior art, this application has the following advantages:

[0116] 1. By setting multiple sampling positions, the extracted features can better describe the feature space of the target and improve the stability of recognition.

[0117] 2. There are no particularly strict requirements for setting the sampling position, which facilitates the design of the production test process.

[0118] 3. The robot's rotation speed, direction or sampling rate can be appropriately adjusted according to the position of the target, which simplifies the control process and is more suitable for configuration on small embedded systems.

[0119] 4. For systems with strong computing power, a classifier interface based on small sample learning is provided, which can further enhance the recognition ability of target features and thus improve the overall recognition accuracy.

[0120] 5. By controlling the texture of the target, the matching coefficient can form a specific frequency response, and after filtering, the disturbance caused by abnormal data can be effectively suppressed.

[0121] 6. There is no particularly strong dependence between the texture spatial frequency and the target structure. While ensuring the texture frequency characteristics, the structure itself still retains a large degree of processing freedom, which is convenient for controlling manufacturing costs.

[0122] Correspondingly, such as Fig. 9 As shown, the embodiment of the present invention also provides a target texture recognition device, which can be applied to recognition equipment, such as Figure 1 The identification device 100 shown, the target texture identification device 800 includes:

[0123] A feature vector acquisition module 801 is used to acquire brightness data and extract feature vectors from the brightness data;

[0124] An evaluation module 802, used to evaluate the feature vector using an offline module to obtain an approximation coefficient;

[0125] A filtering module 803, configured to perform spatial frequency filtering on the first approximation coefficient;

[0126] A judgment module 804 is used to judge whether the feature vector is a target feature based on the approximation coefficient after the spatial frequency filtering process, wherein the target feature corresponds to a feature template of a target;

[0127] The calculation module 805 is used to calculate the position and direction of the target feature in the brightness data if yes, then if yes, calculate the position and direction of the target feature in the brightness data.

[0128] The target texture recognition method, device, recognition equipment and storage medium of the embodiment of the present invention obtain brightness data without temporarily changing the sampling rate of the brightness data collected by the radar; and extracting feature vectors from the brightness data can better describe the characteristics of the target and improve the stability of recognition; the feature vectors are evaluated by an offline module to obtain an approximation coefficient, and the approximation coefficient can preliminarily determine the recognition result; in order to reduce the influence of noise and abnormal data in the approximation coefficient, the approximation coefficient is subjected to spatial frequency filtering processing, and based on the approximation coefficient after the spatial frequency filtering processing, it is determined whether the feature vector is a target feature, wherein the target feature corresponds to a feature template of the target, and when the feature vector is a target feature, the position and direction of the target feature in the brightness data are calculated, which can effectively suppress the disturbance caused by the specific data and improve the accuracy of target texture recognition.

[0129] In some embodiments with low computing power support, the evaluation module 802 is further used to:

[0130] The feature vector is compared with the feature template to obtain a first approximation coefficient.

[0131] In some embodiments with low computing power support, the evaluation module 802 is further used to:

[0132] Calculating the Euclidean distance or direction cosine between the feature vector and the feature template to obtain an original evaluation parameter;

[0133] After normalizing the original evaluation parameters, the first approximation coefficient is obtained.

[0134] In some embodiments with high computing power support, the evaluation module 802 is further used to:

[0135] Using a pre-trained feature classifier to evaluate the feature vector to obtain an original cost coefficient;

[0136] After normalizing the original cost coefficient, a second approximation coefficient is obtained.

[0137] In some embodiments, the target texture recognition device 800 further includes:

[0138] The feature template acquisition module 806 is used to:

[0139] Move the recognition device to a preset sampling position and collect radar point cloud data of the preset sampling position, wherein the preset sampling position is distributed within a preset sampling area, and the radar point cloud data includes position and brightness information;

[0140] Determining radar point cloud data corresponding to the target position in the radar point cloud data according to the position information in the point cloud data and the position information of the target;

[0141] Extract features based on brightness information in the radar point cloud data corresponding to the target position to obtain feature positive samples;

[0142] Obtaining negative sample data other than the radar point cloud data corresponding to the target position in the radar point cloud data, and performing feature extraction on the negative sample data to obtain a plurality of feature negative samples;

[0143] The feature template is obtained according to the feature positive sample and the feature negative sample.

[0144] In some embodiments, after the feature template acquisition module 806 executes the step of obtaining the feature template, the target texture recognition device 800 further includes:

[0145] The feature classifier acquisition module 807 is also used for:

[0146] Based on the feature positive samples and the labels corresponding to the feature positive samples, and the feature negative samples and the labels corresponding to the feature negative samples, a classifier model is trained to obtain the feature classifier.

[0147] In some embodiments, the classifier model is W T xb, feature classifier acquisition module 812, is also used for:

[0148] Input the feature positive samples and the labels corresponding to the feature positive samples, and the feature negative samples and the labels corresponding to the feature negative samples into the classifier model one by one, and obtain the parameters W and b of the classifier model with the goal of optimizing the cost function;

[0149] Wherein, the optimization cost function is ∑[1-y(W T xb)].

[0150] In some embodiments, the target texture recognition device 800 further includes a pre-processing module 808 for:

[0151] Correcting the position in the radar point cloud data based on a preset map;

[0152] A data segment of interest is extracted from the radar point cloud data.

[0153] In some embodiments, after the feature vector acquisition module 801 acquires the brightness data, the feature vector acquisition module 801 is further configured to:

[0154] The real-time brightness data is subjected to sliding window processing with a dynamic length to intercept the section of local brightness data.

[0155] It should be noted that the above device has the functional modules and beneficial effects corresponding to the method. For technical details not fully described in the device embodiment, please refer to the method provided in the embodiment of this application.

[0156] The present application also provides a non-volatile computer-readable storage medium, wherein the computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are executed by one or more processors, for example Figure 1 A processor 12 in the embodiment may enable the one or more processors to execute the target texture recognition method in any of the above method embodiments, for example, executing the above described Figure 2 Steps 101 to 105 of the method; implementing Fig. 9 The functions of modules 801-808 in.

[0157] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0158] Through the description of the above embodiments, a person of ordinary skill in the art can clearly understand that each embodiment can be implemented by means of software plus a general hardware platform, or of course by hardware. A person of ordinary skill in the art can understand that all or part of the processes in the above embodiment methods can be implemented by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes of the embodiments of the above methods. Among them, the storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0159] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them. Under the concept of the present invention, the technical features in the above embodiments or different embodiments may also be combined, the steps may be implemented in any order, and there are many other changes in different aspects of the present invention as described above, which are not provided in detail for the sake of simplicity. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they may still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A target texture recognition method, characterized in that: The method comprises: Acquiring brightness data, and extracting a feature vector from the brightness data; Using an offline module to evaluate the feature vector to obtain an approximation coefficient; The method of evaluating the feature vector using an offline module to obtain an approximation coefficient includes: Calculating the Euclidean distance or direction cosine between the feature vector and the feature template to obtain the original evaluation parameter; After normalizing the original evaluation parameters, an approximation coefficient is obtained; Performing spatial frequency filtering on the approximation coefficient; Based on the approximation coefficient after the spatial frequency filtering process, determining whether the feature vector is a target feature, wherein the target feature corresponds to a feature template of a target; If so, the position and direction of the target feature in the brightness data are calculated.

2. The method according to claim 1, characterized in that The method further comprises: Move the recognition device to a preset sampling position and collect radar point cloud data of the preset sampling position, wherein the preset sampling position is distributed within a preset sampling area, and the radar point cloud data includes position and brightness information; Determining radar point cloud data corresponding to the target position in the radar point cloud data according to the position information in the point cloud data and the position information of the target; Extract features based on brightness information in the radar point cloud data corresponding to the target position to obtain feature positive samples; Obtaining negative sample data other than the radar point cloud data corresponding to the target position in the radar point cloud data, and performing feature extraction on the negative sample data to obtain a plurality of feature negative samples; The feature template is obtained according to the feature positive sample and the feature negative sample.

3. The method according to claim 2, characterized in that After obtaining the feature template, the method further includes: Based on the feature positive samples and the labels corresponding to the feature positive samples, and the feature negative samples and the labels corresponding to the feature negative samples, a classifier model is trained to obtain a feature classifier.

4. The method according to claim 3, characterized in that The classifier model is W r xb, training a classifier model based on the feature positive sample and the label corresponding to the feature positive sample, and the feature negative sample and the label corresponding to the feature negative sample to obtain a feature classifier, including: Input the feature positive samples and the labels corresponding to the feature positive samples, and the feature negative samples and the labels corresponding to the feature negative samples into the classifier model one by one, and obtain the parameters W and b of the classifier model with the goal of optimizing the cost function; Wherein, the optimization cost function is ∑1-y k (W r -b).

5. The method according to claim 2, characterized in that: After moving the recognition device to the preset sampling position and collecting the radar point cloud data of the preset sampling position, the method further includes: Correcting the position in the radar point cloud data based on a preset map; Extract interesting data segments from radar point cloud data.

6. The method according to claim 1, characterized in that After acquiring the brightness data, the method further includes: The real-time brightness data is processed with a sliding window of dynamic length to intercept a section of local brightness data.

7. A target texture recognition method, characterized in that: The method comprises: Acquire brightness data, and extract feature vectors from the brightness data; Using an offline module to evaluate the feature vector to obtain an approximation coefficient; The method of evaluating the feature vector using an offline module to obtain an approximation coefficient includes: Using a pre-trained feature classifier to evaluate the feature vector to obtain an original cost coefficient; After normalizing the original cost coefficient, an approximation coefficient is obtained; Performing spatial frequency filtering on the approximation coefficient; Based on the approximation coefficient after the spatial frequency filtering process, determining whether the feature vector is a target feature, wherein the target feature corresponds to a feature template of a target; If so, the position and direction of the target feature in the brightness data are calculated.

8. The method according to claim 7, characterized in that The method further comprises: Move the recognition device to a preset sampling position and collect radar point cloud data of the preset sampling position, wherein the preset sampling position is distributed within a preset sampling area, and the radar point cloud data includes position and brightness information; Determining radar point cloud data corresponding to the target position in the radar point cloud data according to the position information in the point cloud data and the position information of the target; Extract features based on brightness information in the radar point cloud data corresponding to the target position to obtain feature positive samples; Obtaining negative sample data other than the radar point cloud data corresponding to the target position in the radar point cloud data, and performing feature extraction on the negative sample data to obtain a plurality of feature negative samples; The feature template is obtained according to the feature positive sample and the feature negative sample.

9. The method according to claim 8, characterized in that After obtaining the feature template, the method further includes: Based on the feature positive samples and the labels corresponding to the feature positive samples, and the feature negative samples and the labels corresponding to the feature negative samples, a classifier model is trained to obtain a feature classifier.

10. The method according to claim 9, characterized in that The classifier model is W r xb, training a classifier model based on the feature positive sample and the label corresponding to the feature positive sample, and the feature negative sample and the label corresponding to the feature negative sample to obtain a feature classifier, including: Input the feature positive samples and the labels corresponding to the feature positive samples, and the feature negative samples and the labels corresponding to the feature negative samples into the classifier model one by one, and obtain the parameters W and b of the classifier model with the goal of optimizing the cost function; Wherein, the optimization cost function is ∑1-y k (W r -b).

11. The method according to claim 8, characterized in that After moving the recognition device to the preset sampling position and collecting the radar point cloud data of the preset sampling position, the method further includes: Correcting the position in the radar point cloud data based on a preset map; Extract interesting data segments from radar point cloud data.

12. The method according to claim 7, characterized in that After acquiring the brightness data, the method further includes: The real-time brightness data is processed with a sliding window of dynamic length to intercept a section of local brightness data.

13. A target texture recognition device, characterized in that: The device comprises: A feature vector acquisition module, used to acquire brightness data and extract feature vectors from the brightness data; An evaluation module, used for evaluating the feature vector using an offline module to obtain an approximation coefficient; The evaluation module is also used to: calculate the Euclidean distance or direction cosine between the feature vector and the feature template to obtain an original evaluation parameter; and obtain an approximation coefficient after normalizing the original evaluation parameter; A filtering module, used for performing spatial frequency filtering on the approximation coefficient; A judgment module, used for judging whether the feature vector is a target feature based on the approximation coefficient after the spatial frequency filtering process, wherein the target feature corresponds to a feature template of a target; A calculation module is used to calculate the position and direction of the target feature in the brightness data.

14. A target texture recognition device, characterized in that: The device comprises: A feature vector acquisition module, used to acquire brightness data and extract feature vectors from the brightness data; An evaluation module, used for evaluating the feature vector using an offline module to obtain an approximation coefficient; The evaluation module is also used to: evaluate the feature vector using a pre-trained feature classifier to obtain an original cost coefficient; normalize the original cost coefficient to obtain an approximation coefficient; A filtering module, used for performing spatial frequency filtering on the approximation coefficient; A judgment module, used for judging whether the feature vector is a target feature based on the approximation coefficient after the spatial frequency filtering process, wherein the target feature corresponds to a feature template of a target; A calculation module is used to calculate the position and direction of the target feature in the brightness data.

15. An identification device, characterized in that: The identification device comprises: at least one processor, and A memory, the memory is communicatively connected to the processor, the memory stores instructions executable by the at least one processor, the instructions are executed by the at least one processor so that the at least one processor can execute the method according to any one of claims 1 to 12.

16. A non-volatile computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are executed by a recognition device, the recognition device is caused to perform the method according to any one of claims 1 to 12.

Images