CN112327265A

CN112327265A - Division and treatment detection method based on semantic segmentation network

Info

Publication number: CN112327265A
Application number: CN202011147125.9A
Authority: CN
Inventors: 胡程; 王锐; 周超; 李思伟
Original assignee: Beijing Institute of Technology BIT
Current assignee: Beijing Institute of Technology BIT
Priority date: 2020-10-23
Filing date: 2020-10-23
Publication date: 2021-02-05

Abstract

The invention discloses a division and treatment detection algorithm based on a semantic segmentation network. The invention realizes the clutter region segmentation of the range-Doppler domain by means of a semantic segmentation network, sets a detection strategy according to the characteristics of main bodies and edges of different clutter regions, selects a proper detector and reasonable detection parameters from an alternative detector group to complete divide-and-conquer detection, and effectively inhibits false alarms while detecting a target to the maximum extent. The technology has important significance for radar detection of low-altitude small targets. Compared with a detection plane normalization method, the method does not need detection plane conversion, has smaller calculated amount, avoids the influence of a parameter estimation result on the detection plane conversion, and has stronger robustness.

Description

Division and treatment detection method based on semantic segmentation network

Technical Field

The invention belongs to the technical field of low-altitude small target radar detection, and particularly relates to a division and treatment detection method based on a semantic segmentation network.

Background

The low-altitude small target is a flying target with the flying height below 1000 meters and the radar reflection sectional area less than 2 square meters. The low-altitude small target is closely related to human production life, according to relevant data statistics, biological targets such as birds and the like are main causes of bird attack events, and artificial targets such as unmanned aerial vehicles and the like are main tools of terrorist attacks. The radar is a powerful tool for detecting the low-altitude small target, and has important significance for warning in advance and improving public safety.

In the radar detection of the low-altitude small target, due to the unique motion characteristic of the target, the target is very easily influenced by various clutter such as ground vegetation, weather cloud and rain and the like, and different types of clutter present different clutter areas in a range Doppler domain. The traditional target detection method adopts a single detector to carry out Chinese Constant False Alarm detection (CFAR) in a range-Doppler domain, cannot give consideration to the detection performance of different clutter areas, and is easy to generate False alarms and missing detection. The current processing idea is to identify clutter areas and adopt a differentiation processing strategy, after identifying clutter areas, the idea of distribution indexing is mostly adopted, namely, a detection plane containing different clutter areas is normalized into index distribution of the same parameter, and then target detection is carried out, but the method needs to convert each value of the detection plane, has huge calculation amount and depends heavily on a parameter estimation result, and cannot form a uniform detection plane when the parameter estimation is inaccurate, so that the detection performance is deteriorated rapidly.

Disclosure of Invention

In view of this, the invention provides a divide-and-conquer detection method based on a semantic segmentation network, which can adopt different detectors to detect different clutter distribution areas, thereby improving detection accuracy.

A divide and conquer detection method based on semantic segmentation network processes the echo signal of radar to form image; cutting the image, and identifying the cut image through a semantic segmentation network to obtain an identification result, wherein the identification result comprises a range-Doppler domain clutter-free area, a rain clutter area and a ground clutter area; and then, different detectors and corresponding retrieval strategies are adopted for detecting Doppler domain clutter-free areas and rain clutter.

Preferably, the method for processing the echo signal of the radar to form the image includes:

performing pulse compression and coherent accumulation on an echo signal of the radar to obtain a range-Doppler domain data matrix; then converting the range-Doppler domain data matrix into an energy ratio matrix; and converting the obtained energy ratio matrix into picture pixel values with the same size in a linear mapping mode to finally form an image.

Preferably, the semantic segmentation network performs neural network learning before use, and the specific method is as follows:

and labeling each pixel in the image by combining program pre-labeling and manual adjustment, and performing neural network learning by taking a labeling result and the image as input objects of a semantic segmentation network.

Preferably, the search policy is:

(1) the unit to be detected is located in the ground clutter region

Target detection is not performed because the ground clutter region is a non-detection region;

(2) the unit to be detected is located in a clutter-free area

When the unit to be detected is positioned in the clutter-free area, a CA-CFAR detector is adopted;

(3) the unit to be detected is located in the rain clutter area

When the unit to be detected is located in the rain clutter area, a CA-CFAR detector is adopted;

and after the detector is selected, setting detection parameters according to the detectors selected in different clutter areas, and completing target detection.

Preferably, the search strategy further considers that when non-homogeneous units are mixed:

(1) the unit to be detected is located in a clutter-free area

When the reference unit is mixed into a non-homogeneous unit, namely, other clutter region interference exists at the edge of the reference unit, an SO-CFAR detector is adopted;

(2) the unit to be detected is located in the rain clutter area

When the reference cell half-window is mixed with a non-homogeneous cell of the clutter-free zone, a GO-CFAR detector is adopted;

when non-homogeneous units are mixed in other forms, GO-CFAR detectors are used.

Has the advantages that:

1. the clutter region segmentation of a range-Doppler domain is realized by means of a semantic segmentation network, a detection strategy is formulated according to the characteristics of main bodies and edges of different clutter regions, a proper detector and reasonable detection parameters are selected from an alternative detector group to complete divide-and-conquer detection, and false alarms are effectively suppressed while a target is detected to the maximum extent. The technology has important significance for radar detection of low-altitude small targets. Compared with a detection plane normalization method, the method does not need detection plane conversion, has smaller calculated amount, avoids the influence of a parameter estimation result on the detection plane conversion, and has stronger robustness.

Drawings

Fig. 1 is an overall algorithm flow.

FIG. 2 is a mapping of energy ratio to image pixel.

FIG. 3(a) shows the flow of the pre-labeling procedure.

FIG. 3(b) is a diagram illustrating the recognition result of the pre-annotation process.

FIG. 4 is a diagram of a divide and conquer detector.

Fig. 5(a) is a view of the actual measurement data collection scenario 1.

Fig. 5(b) is a view of the actual measurement data collection scenario 2.

Fig. 6(a) is a schematic diagram of the range-doppler domain of the rain clutter.

Fig. 6(b) is a schematic diagram of the ground clutter range-doppler domain.

Fig. 7(a) is a schematic diagram of data picture input.

Fig. 7(b) is a data picture label schematic.

Fig. 8(a) shows a clutter recognition result — a picture input diagram.

Fig. 8(b) is a schematic diagram of the clutter recognition result-network output.

Fig. 8(c) is a clutter recognition result-image label.

FIG. 9(a) is the result of the partition test method- -CA-CFAR example 1

FIG. 9(b) shows the result of the partition and measure method- -GO-CFAR example 1

FIG. 9(c) shows the result of the divide and conquer test method- -SO-CFAR example 1

FIG. 9(d) shows the result of the divide and conquer method- -example of clutter recognition result 1

FIG. 9(e) shows the results of the divide-and-conquer test method- -divide-and-conquer test example 1

FIG. 10(a) is the result of the partition test method- -CA-CFAR example 2

FIG. 10(b) shows the result of the partition and measure method- -GO-CFAR example 2

FIG. 10(c) is the result of the divide and conquer test method- -SO-CFAR example 2

FIG. 10(d) shows the result of the divide-and-conquer method- -example of clutter recognition result 2

FIG. 10(e) shows the results of the divide-and-conquer test method- -divide-and-conquer test example 2

Detailed Description

The invention is described in detail below with reference to the accompanying drawings and two embodiments.

The invention provides a division and treatment detection method based on a semantic segmentation network. The method is based on the basic idea that clutter region division of a range-Doppler domain is achieved through a semantic segmentation network, and then appropriate detectors and detection parameters are selected from an alternative detector group according to attributes of a detection unit and a reference unit to conduct target detection, so that detection performance under a complex clutter environment is improved. The whole algorithm flow is shown in fig. 1.

Step one, performing pulse compression and coherent accumulation on an echo signal of the radar to obtain a range-Doppler domain data matrix. On the basis, the range-doppler domain data matrix is converted into an energy ratio matrix according to equation (1).

Wherein p is_i,jIs the (i, j) th element, h, of the energy ratio matrix P_i,jIs the (i, j) th element of the range-doppler domain data matrix H, and N represents the number of elements of the matrix.

And step two, mapping the energy ratio matrix in the step one into an image pixel value according to the formula (2).

Wherein the content of the first and second substances,pixel_i,jrepresenting the (i, j) -th element pixel value, floor (·) representing the rounding operation, S_minAnd representing a mapping truncation threshold, k represents a mapping slope, and then converting the obtained energy ratio matrix into picture pixel values with the same size in a linear mapping mode, wherein the upper limit of the obtained image pixel values is 255, and the lower limit of the obtained image pixel values is 0. The specific mapping relationship is shown in fig. 2. The specific expression is shown as formula (3).

Wherein S is_maxRepresenting the maximum in the energy ratio matrix.

And step three, cutting the image obtained in the step two according to the used semantic segmentation network requirement and hardware performance to obtain an image slice. For example, the deep v3+ network requires that the width and height of the input image be greater than 320 pixels.

And step four, dividing the clutter region into a Doppler region clutter-free region, a rain clutter region and a ground clutter region according to the clutter attributes. Inputting the image slices obtained in the third step into a semantic segmentation network to obtain the recognition results of the network on the range-Doppler domain clutter-free area, the rain clutter area and the ground clutter area;

at the initial use, neural network training (BP learning) is required for the semantic segmentation network. BP learning requires providing labels corresponding to the entries. The tag contains clutter class information corresponding to the input. The characteristics of different clutters need to be considered when marking the clutter classes. For the ground clutter, the distribution is mainly in a central Doppler channel, and the strength relation along the distance direction and the extension degree in the Doppler direction are related to a scene; for rain clutter, as raindrops are filled in the whole radar wave beam, the distance direction and the Doppler direction of a rain clutter area are expanded in different degrees, a block-shaped clutter area is formed, and the area is large. The clutter-free area also presents block distribution, but the intensity is obviously lower than that of the rain clutter area because of no clutter energy interference. And when the picture is marked, the marking precision can be improved by referring to the prior information.

In order to improve the labeling precision and the labeling efficiency, the labels are obtained by combining program pre-labeling and manual adjustment. The program is pre-labeled as a series of pixel fill, pixel culling, and dilation erosion processes, and the specific method flow is shown in fig. 3(a) and 3 (b). The program is redundant, and the parameter adaptability and the method adaptability are poor, so that the labeling and classifying process cannot be independently completed. Therefore, manual adjustment is performed after program pre-labeling, and mainly relates to whether the clutter class is correct, calibrating the edge profile of the clutter and the like.

Fifthly, dividing and detecting by combining the original range-Doppler domain data matrix and the clutter identification result in the fourth step; the structure of the divide and conquer detector is shown in fig. 4, and the specific detection strategy is as follows:

the characteristics of the three distinguished areas are different, and the distribution range of the rain clutter area and the clutter-free area is large and is an area where a target can exist in a large probability; the ground clutter area is caused by a strong ground reflection object, the energy is high, targets falling into the area are often submerged by clutter, and the targets are not easy to detect. Therefore, the ground clutter region is set as a non-detection region, and the non-clutter region and the rain clutter region are set as regions to be detected. The research and actual measurement of predecessors show that the amplitudes of a clutter-free area (background noise) and a rain clutter area can be approximately described by Rayleigh distribution, so that an average CFAR detector is selected for target detection.

Because the areas with different attributes are mainly distributed in a block shape, clutter edges are mixed into non-homogeneous sampling from one side more during target detection, meanwhile, the target distribution is sparse, and the multi-target phenomenon is not obvious. Therefore, in order to combine the advantages of different detectors, different detection strategies are set according to the clutter recognition result and the attributes of the detection unit and the reference unit, and the detection strategies are as follows:

(1) the unit to be detected is located in the ground clutter region

Since the ground clutter region is a non-detection region, no target detection is performed regardless of the attribute of the reference cell.

(2) The unit to be detected is located in a clutter-free area

When the reference unit is not mixed with non-homogeneous units, a CA-CFAR detector is employed.

When the reference unit is mixed with the heterogeneous unit, the edge of the reference unit has the characteristics different from the image of the local clutter area, and other clutter area interferences exist.

(3) The unit to be detected is located in the rain clutter area

When the reference cell half-window is mixed with non-homogeneous cells in the clutter free zone, a GO-CFAR detector is used to counter false alarms generated by clutter edges due to weak mixed cell energy.

When non-homogeneous units are mixed in other forms, the GO-CFAR detector is used uniformly, as it is rare that no more detailed differentiation is made here.

After the detector is selected, differential or non-differential detection parameters are set for different clutter areas (as the case may be), and target detection is completed. Namely: and setting detection parameters corresponding to the detectors according to the detectors selected by different clutter areas and experience to finish target detection.

Example (b):

in this example, the effectiveness of the algorithm is verified using the measured clutter data. The actual measurement clutter data is ground clutter and rain clutter data, and a specific experimental scene is shown in fig. 5(a) and 5 (b). The radar parameters associated with the algorithm are shown in table 1.

TABLE 1 Radar parameters

The range-doppler domain obtained after pulse compression and coherent accumulation of the typical rain clutter and ground clutter echoes is shown in fig. 6(a) and 6 (b).

The division and treatment detection method based on the semantic segmentation network is adopted to complete the method verification under the measured data, and the specific flow is as follows:

step one, pulse compression and 512-frame coherent accumulation are carried out on the echo of the actually measured clutter data. And (4) converting the obtained range-Doppler domain data matrix into pixel values of the image according to the formulas (4) to (6), labeling the image by using an image morphology method, and manually and repeatedly correcting to obtain a more accurate labeling result. The image is then cropped by cropping to 512 pixel wide and 512 pixel high image slices as a network trained dataset. Fig. 7(a) and 7(b) give examples of data sets.

Step two, the ratio of 7: 3, dividing the data set in the step one into a training set and a testing set, selecting DeepLab v3+ developed by Google by a semantic segmentation network, carrying out training of the network by using the training set, adopting a fine-tuning idea for training, and loading pre-training weight of the network under the VOC2012 data set except that a logits layer for carrying out category judgment needs to be retrained. Specific training parameters are shown in the following table.

TABLE 2 network training parameters

The learning rate is calculated as shown in equation (7).

Wherein current _ lr represents the current learning rate, base _ lr represents the initial learning rate, global _ step represents the current training number, training _ step represents the total training number, and learning _ power represents the attenuation coefficient.

And step three, inputting the test set sample in the data set into a neural network to obtain a clutter recognition result. Fig. 8(a), 8(b) and 8(c) show a set of recognition results.

Step four, according to the characteristics of different clutters, the division detection is carried out on the main body and the edge of different clutter areas, and the detection strategy is as follows:

(1) the unit to be detected is located in the ground clutter region

(2) The unit to be detected is located in a clutter-free area

When the reference cell is mixed with a non-homogeneous cell, an SO-CFAR detector is used to counter the target shadowing effect caused by clutter edges due to the strong energy of the heterogeneous cell.

(3) The unit to be detected is located in the rain clutter area

The specific detection parameters are as follows: two-stage threshold detection is adopted, wherein a fixed threshold detection with a first-stage threshold of 10dB is adopted, two-dimensional CFAR detection is adopted for points exceeding the fixed threshold, and the number of reference units in any direction of the two-dimensional CFAR detection is 48. The preset false alarm rate of the clutter-free area is 10^-6In order to suppress the false alarm caused by tailing in the rain clutter area, the false alarm rate of the rain clutter area is properly increased to 10^-8。

Fig. 9(a) - (e) and fig. 10(a) - (e) show examples of detection results, and it can be seen that the detection result obtained by the method can effectively suppress false alarms caused by ground clutter and rain clutter, and has better detection performance for targets located in a clutter edge environment.

And (4) counting the performances of the method and the CFAR detection methods of all the mean values on the test set. The evaluation results are shown in table 3.

TABLE 3 comparison of the Performance of the test methods

On the premise of basically not losing the target, the divide-and-conquer detection method can effectively inhibit false alarms and has higher detection performance. The validity of the division and treatment detection method based on the semantic segmentation network can be seen through the verification of the actually measured clutter data. The method can improve the detection performance of the low-altitude small target. .

In summary, the above description is only a preferred embodiment of the present invention, and is not intended to limit the scope of the present invention. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A divide and conquer detection method based on semantic segmentation network is characterized in that an echo signal of a radar is processed to form an image; cutting the image, and identifying the cut image through a semantic segmentation network to obtain an identification result, wherein the identification result comprises a range-Doppler domain clutter-free area, a rain clutter area and a ground clutter area; and then, different detectors and corresponding retrieval strategies are adopted for detecting Doppler domain clutter-free areas and rain clutter.

2. The divide-and-conquer detecting method according to claim 1, wherein the echo signal of the radar is processed to form an image by:

3. The divide and conquer detection method of claim 1, wherein the semantic segmentation network performs neural network learning before use, and the specific method is as follows:

4. The divide and conquer detection method of claim 2 or 3, wherein the search strategy is:

(1) the unit to be detected is located in the ground clutter area, and target detection is not carried out;

(2) the unit to be detected is located in a clutter-free area, and a CA-CFAR detector is adopted for target detection;

(3) and the unit to be detected is positioned in the rain clutter area, and a CA-CFAR detector is adopted to detect the target.

5. The divide and conquer detection method of claim 4, wherein the search strategy further considers, when non-homogeneous units are mixed:

(1) the unit to be detected is located in a clutter-free area, and when the reference unit is mixed with the non-homogeneous unit, the SO-CFAR detector is adopted for detection;

(2) the unit to be detected is located in a rain clutter area, and when a reference unit half window is mixed with a non-homogeneous unit without a clutter area, a GO-CFAR detector is adopted for detection; when non-homogeneous units are mixed in other forms, detection is performed using a GO-CFAR detector.