Clustering algorithm for Terahertz radar extended target tracking

Abstract

The invention discloses a clustering algorithm for Terahertz radar extended target tracking. The clustering algorithm comprises three steps of drawing an azimuth distance map by using original data, carrying out two-dimensional unit average constant false alarm rate detection on an azimuth distance dimension, and clustering by a connected domain labeling method. In the first step, a grid interpolation method is used for carrying out zero interpolation on the reflection intensity of discrete points, so that subsequent calculation of background noise is facilitated; in the second step, a threshold value is calculated by setting related parameters, and part of clutter interference is filtered out; in the third step, adjacent data points are communicated through an eight-field method, then morphological closed operation is designed to merge adjacent targets, the problem of excessive segmentation is effectively avoided, then a minimum area threshold value is added to filter clutter interference, effective targets are reserved, and finally the clustering effect is more remarkable through target center point labeling. The clustering algorithm for Terahertz radar extended target tracking is low in cost, high in processing speed and remarkable in effect, and has certain application value in the fields of Terahertz radar extended target tracking and the like.

Description

Technical Field

[0001] This invention relates to the field of terahertz radar technology, and in particular to a clustering algorithm for extended target tracking in terahertz radar. Background Technology

[0002] In recent years, terahertz radar has seen significant development and application value in the field of target tracking due to its unique advantages such as large bandwidth, high resolution, Doppler sensitivity, and anti-jamming capabilities. Extended target tracking is one of the hot research directions in the field of target tracking. Extended targets refer to targets that cannot be simply regarded as point targets in target tracking. These targets occupy multiple resolution cells in the sensor and generate multiple measurement points. This means that the target information collected by the radar is not just a simple representation of a point target.

[0003] However, most traditional multi-target tracking algorithms are based on the point assumption, allocating measurement data to specific targets through data association between measurement points and target points. If the target does not meet the point target characteristics, inaccuracies will occur during the data association process, further leading to a decline in tracking performance, and even target loss. The increase in the amount of point cloud measurement data and the uncertainty of its location also increase the complexity and challenge of data association, making these multi-target tracking algorithms difficult to effectively handle complex situations such as an increased number of targets and dense clutter environments, and unsuitable for tracking multiple extended targets. Therefore, applying clustering algorithms to extended targets to enhance the accuracy of data association is a core issue in the field of terahertz radar target tracking.

[0004] This invention aims to provide a new approach for extended target clustering in terahertz radar. Summary of the Invention

[0005] The purpose of this invention is to disclose a clustering algorithm for extended target tracking in terahertz radar, which includes preliminary processing of raw data to draw an azimuth-distance map, using the azimuth-distance dimension to perform two-dimensional cell average constant false alarm rate detection, and using the connected component labeling method to cluster the detected objects for the next step of target tracking.

[0006] The terahertz radar can acquire raw data after measuring an object, including four types of data: azimuth angle data, elevation angle data, range data, and reflection intensity data.

[0007] This invention is a clustering algorithm for extended target tracking in terahertz radar. First, the original data needs to be processed. This invention first plots an azimuth-range map using azimuth as the horizontal axis and range as the vertical axis, filling the map with linearly normalized reflection intensity. When detecting an object, the terahertz radar in this invention first fixes an azimuth angle, then evenly divides the range from 0 to the maximum range into 1024 range gates at that azimuth angle, sequentially collecting reflection intensity data for each range gate. After collecting 1024 data points at the current azimuth angle, the radar changes the azimuth angle and continues collecting 1024 range gate data points at the new azimuth angle. Therefore, both range and azimuth are discretely sampled. When plotting the azimuth-range map, a grid interpolation method is needed to interpolate the reflection intensity data of the discrete points.

[0008] The linear normalization transforms the data linearly to the [0,1] interval, and its formula is as follows:

[0009] X_norm=(X-X_min) / (X_max-X_min)

[0010] The grid interpolation method uses the griddata function to interpolate the reflection intensity of discrete points.

[0011] This invention is a clustering algorithm for extended target tracking in terahertz radar. After preliminary processing of the original data, the second part of this invention uses the azimuth distance dimension to perform two-dimensional cell average constant false alarm rate (CFAR) detection. The goal of CFAR detection is to set a threshold that is high enough to limit false alarms to a tolerable range, but low enough to allow target detection.

[0012] The constant false alarm rate (CFAR) detection is a two-dimensional unit average CFAR detection, in which the main parameters include the detected unit N, row reference unit N_TR, column reference unit N_TC, row protection unit N_GR, column protection unit N_GC, false alarm probability P_fa, threshold factor α, background noise_mean, and threshold.

[0013] The parameters, namely row reference unit N_TR, column reference unit N_TC, row protection unit N_GR, column protection unit N_GC, and false alarm probability P_fa, can be adjusted according to requirements.

[0014] The parameter N is calculated using the formula: (2*N_TC+2*N_GC+1)*(2*N_TR+2*N_GR+1)-(2*N_GC+1)*(2*N_GR+1).

[0015] The threshold factor α is calculated using the formula: N*(P_fa^(-1 / N)-1).

[0016] The background noise is calculated by obtaining the average value of the reference cell using the nanmean function; the threshold is calculated using the formula: α*noise_mean.

[0017] The basic principle of constant false alarm rate (CFAR) detection is to compare the intensity of each data point with a threshold, delete data points below the threshold, and leave the remaining data points as the detected objects.

[0018] This invention is a clustering algorithm for extended target tracking in terahertz radar. After preliminary processing of the original data and filtering out data points below a threshold using two-dimensional unit average constant false alarm rate detection based on azimuth distance dimension, since extended targets have many data points in a single measurement, the third part of this invention uses connected component labeling to cluster data points that exceed the threshold and belong to the same object, thus identifying them as the same object.

[0019] A connected component generally refers to an image region consisting of pixels with the same pixel value and adjacent positions. Connected component analysis involves identifying and labeling each connected component in an image. The object of connected component analysis is generally a binarized image. Pixel neighborhood relationships are generally classified into four-neighborhood and eight-neighborhood.

[0020] The neighborhood: The set of points corresponding to pixel (x,y) {(x+p,y+q); (p,q) is a pair of meaningful integers} is called the neighborhood of pixel (x,y).

[0021] The connectivity is defined as follows: pixel f(x,y) = f(x+i,y+j), where i and j exist within the image R. f(x,y) and its neighboring points have 4-point and 8-point neighborhoods. The 4-neighbor neighborhood consists of four points (top, bottom, left, and right). The 8-neighbor neighborhood consists of eight points, including points on the diagonal.

[0022] The labeling process involves assigning labels to each target image region after image segmentation. The labeling method depends on whether a 4-neighborhood or 8-neighborhood method is used; therefore, commonly used labeling methods are divided into 4-connected labeling and 8-connected labeling.

[0023] The connected component labeling method employs an 8-neighbor clustering algorithm to detect targets and merges adjacent targets through morphological closing operations to avoid over-segmentation. Subsequently, a minimum area threshold is set to filter out clutter interference and retain valid targets. Finally, the center point of each target is labeled to make the clustering results clearer and more intuitive. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. The extended target used in this invention is a quadcopter drone.

[0025] Figure 1 The diagram shows a flowchart of the clustering algorithm for extended target tracking in terahertz radar according to the present invention.

[0026] Figure 2 The image shows an example of raw data from a terahertz radar measurement.

[0027] Figure 3 The image shows a quadcopter drone hovering against a background of natural clutter.

[0028] Figure 4 The image shown is a location distance map after interpolation.

[0029] Figure 5 The image shows the constant false alarm rate (CFAR) detection results and statistics before clustering.

[0030] Figure 6 The image shows the constant false alarm rate (CFAR) detection results and statistics after clustering.

[0031] The figures are labeled as follows: 1. The quadcopter drone in the orientation distance map; 2. The quadcopter drone in constant false alarm rate (CFAR) detection before clustering; 3. The quadcopter drone in CFAR detection after clustering. Detailed Implementation

[0032] The following detailed description of a clustering algorithm for extended target tracking in terahertz radar, in conjunction with the accompanying drawings and specific embodiments, illustrates the present invention. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the invention, and should not be construed as limiting the invention.

[0033] The flowchart of the clustering algorithm used for extended target tracking in terahertz radar is shown below. Figure 1 As shown; a schematic diagram of the original measurement data of the terahertz radar on the extended target in the clustering algorithm used for extended target tracking by terahertz radar is shown below. Figure 2 As shown, the data includes azimuth data, pitch data, distance data, and reflection intensity data. Pitch data is not considered in this invention.

[0034] like Figure 3 The image shows a quadcopter drone hovering in a natural clutter background. This invention uses a quadcopter drone as a typical extended target, and it can be seen that the shape of this quadcopter drone is approximately rectangular in the hovering state.

[0035] like Figure 4The azimuth-distance map of this invention is plotted with azimuth data on the horizontal axis and distance data on the vertical axis. Since the reflection intensity of different objects varies significantly, the reflection intensity is normalized before filling. Furthermore, because the changes in azimuth and distance data are discrete, a grid interpolation method is used to insert zero values ​​into the discrete points. As shown in the figure, the quadcopter UAV appears as a cluster of bright rectangular areas (marked as 1) in the detection results, while the remaining bright areas are natural background interference. Figure 4 It is easy to see that there are 9 relatively obvious highlighted areas in the orientation distance map, which indicates that there are 9 objects in the background in the orientation distance dimension, including the quadcopter drone.

[0036] The two-dimensional unit average constant false alarm rate detection result in this invention is as follows: Figure 5 As shown, after setting parameters such as row reference unit N_TR, column reference unit N_TC, row protection unit N_GR, column protection unit N_GC, and false alarm probability P_fa according to needs, the detected unit N, threshold factor α, background noise_mean, and threshold can be calculated according to relevant formulas. Then, the intensity data of each data point is compared with the threshold, and data points below the threshold are deleted, leaving the data points that indicate the detected object; for example... Figure 5 As shown, after two-dimensional unit average constant false alarm rate (CFAR) detection, the quadcopter UAV appears as clusters of bright rectangular areas (marked as 2) in the detection results. The remaining bright areas are natural background interference. The detection results show a total of 513 valid targets. Figure 5 The test results show that, although Figure 5 The nine highlighted areas shown in the image can be compared with... Figure 4 The azimuth distance map corresponds to this, but the 513 valid targets counted in the detection results are obviously significantly different from it, indicating that all data points above the threshold were misclassified as independent targets.

[0037] The average constant false alarm rate (CFAR) detection result of the two-dimensional cells after clustering in this invention is as follows: Figure 6 As shown, this invention, based on the 8-neighborhood connectivity labeling method, also designs a morphological closing operation to merge neighboring targets, effectively avoiding over-segmentation, and adds a minimum area threshold to filter out clutter interference and retain valid targets. Figure 6 As shown, after clustering, the average constant false alarm rate (CFAR) detection of the two-dimensional units reveals that the quadcopter UAV appears as clustered bright rectangular areas (marked as 3) in the detection results, while the remaining bright areas are natural background interference. The detection results show a total of 9 valid targets. (Comparison...) Figure 4 / Figure 5 / Figure 6It can be seen that the present invention successfully clustered and separated nine extended targets, including quadcopter drones, from a complex background, and the labeling results of the target center points further verified the effectiveness of the method.

Claims

1. A clustering algorithm for extended target tracking in terahertz radar, characterized in that, Includes the following steps: The raw data is initially processed and plotted as a azimuth-distance map. Two-dimensional cell average constant false alarm rate detection is performed using the azimuth distance dimension; Clustering of detected objects is performed using connected component labeling; Finally, by comparing the different presentations of the quadcopter UAV in the azimuth-distance map, the average 2D constant false alarm rate (CFAR) detection results of the unit before clustering, and the average 2D CFAR detection results of the unit after clustering, the clustering effect of the clustering algorithm used for extended target tracking of terahertz radar is characterized.

2. The original data described in claim 1 is preliminarily processed and plotted as a azimuth-distance map, characterized in that: The original data consists of azimuth data, range data, and reflection intensity data derived from terahertz radar measurements. The azimuth data is used as the horizontal axis, the range data as the vertical axis, and the linearly normalized reflection intensity is used as filler. Zero-padding is performed on the discrete points.

3. The method for two-dimensional unit average constant false alarm rate (CFAR) detection based on azimuth distance dimension according to claim 1, characterized in that: The two-dimensional unit average constant false alarm rate detection calculates a threshold by setting relevant parameters, filters valid data points that meet the threshold, and counts the number of valid data points.

4. The method for clustering detected objects using connected component labeling according to claim 1, characterized in that: The connected component labeling method first adopts the 8-neighborhood method, then designs morphological closing operations to avoid over-segmentation, then sets a minimum area threshold to filter valid targets, and finally performs center point labeling and counts the number of valid targets.

5. The threshold screening of effective targets according to claims 3 and 4, characterized in that: The threshold is a dual threshold, namely, the average constant false alarm rate detection of two-dimensional cells to delete invalid data points and the deletion of invalid targets during clustering. Valid targets are selected through the dual threshold.

Images