CN115902800A - Intelligent detection and trend prediction method of millimeter-wave radar high-voltage lines - Google Patents

Abstract

The present invention provides a method for intelligent detection and direction prediction of millimeter-wave radar high-voltage lines, comprising the following steps: step 1, using the radar echo pattern within an antenna scanning period as the input of the convolutional neural network; step 2, setting the first threshold , keep the line point detection results greater than the first threshold; step 3, use the line point detection results as input, and use the multi-curve fitting method to obtain the preliminary detection results of multiple power lines; step 4, for each power line Perform pre-order traversal of the preliminary detection results, and connect each node obtained through the traversal in turn to obtain the power line detection results expressed in the form of line segments; step five, perform smoothing operations on each power line detection result sequence expressed in the form of line segments; step six, A second threshold is set, the electric power lines represented in the form of line segments whose confidence degree or length is smaller than the second threshold are eliminated, and the remaining electric power lines represented in the form of line segments are output as detection results.

Description

Intelligent detection and trend prediction method of millimeter-wave radar high-voltage lines

technical field

The invention relates to the technical field of radar collision avoidance, in particular to a method for intelligent detection and trend prediction of millimeter-wave radar high-voltage lines.

Background technique

Helicopters colliding with power lines are a global problem. In the statistics of flight accidents in recent years, low-altitude obstacles such as chimneys and high-voltage lines have become the main factors that threaten flight safety, and high-voltage lines are often the most difficult to find when the field of vision is poor, so they have become the most dangerous obstacles in low-altitude flight. Since the wavelength of the millimeter wave is close to the size of the power line, the echo of the power line in the millimeter wave band has the most obvious Bragg effect, and is less affected by climate and light, so the millimeter wave anti-collision radar system has become the main anti-collision detection system for low altitude aircraft choose.

The existing millimeter-wave high-voltage line detection algorithm uses constant false alarm detection to extract suspected power line points and the location of power towers, then uses Hough transform and other line detection methods to extract power lines, and finally uses feature classifiers for classification. Ma Qirong et al. proposed a high-voltage line detection algorithm that jointly utilizes Hough transform and support vector machine classifier. The power line detection is completed by extracting the power mean value, average peak distance and variance of the power line, but this method uses the Hough transform to detect the straight line, which is powerless to change the direction of the power line. In the Chinese invention patent CN106529416A, the millimeter-wave echo image is first divided into blocks, and the Hough transform is used to extract straight lines for each divided area, and the features of the line segments are extracted using a decision tree to determine whether each line segment is a power line. However, the detection result of this method is a separated straight line segment instead of a complete high-voltage line, and it is also unable to detect the change of the direction of the power line in the block area. Chinese invention patent CN107561509A uses constant false alarm detection to extract suspected power line points in millimeter wave echo images, uses Kalman filter algorithm to connect line points to complete power line position detection, and uses support vector machine algorithm to complete power line identification and judgment. However, this method is only tested on simulation data, and the calculation of millimeter-wave radar echoes according to the theoretical RCS of the power line when simulating the experimental data cannot be well adapted to the situation where there are a large number of clutter areas in the actual scene, and the Bragg effect is not obvious.

Contents of the invention

In view of this, the embodiment of this specification provides an intelligent detection and direction prediction method for millimeter-wave radar high-voltage lines, so as to improve the obstacle avoidance ability of helicopters flying at low altitudes and ensure the safety of pilots and passengers.

The embodiment of this specification provides the following technical solutions: a method for intelligent detection and direction prediction of millimeter-wave radar high-voltage lines, including the following steps: Step 1, using the radar echo pattern within one antenna scanning period as the input of the convolutional neural network; step 2. Set the first threshold, and keep the line point detection results greater than the first threshold; step 3, use the line point detection results as input, and use the multi-curve fitting method to obtain the preliminary detection results of multiple power lines; step 4 1. Perform preorder traversal on the preliminary detection results of each power line, and connect each node obtained through the traversal in turn to obtain the power line detection results expressed in the form of line segments; Smoothing operation; step 6, setting a second threshold, removing the electric power lines represented by line segments whose confidence degree or length is smaller than the second threshold, and outputting the remaining electric power lines represented by line segments as detection results.

Further, step 1 includes: setting the classification loss function as l _c =-2|c _p -c _g | ^λ c _g logc _p , where c _p and c _g are the category prediction confidence and the real category result respectively, λ Parameters for balancing hard samples.

其中a_p，a_g分别为角度预测结果和真实线点切线角度，d为该点离标签中最近的线点的距离，D为调节角度损失的参数。Further, step 1 also includes: setting the angle prediction using the absolute value loss function

where a _p and a _g are the angle prediction result and the tangent angle of the real line point respectively, d is the distance from the point to the nearest line point in the label, and D is the parameter to adjust the angle loss.

Further, step three includes:

Step 3.1, create an empty tree, and represent all line point detection results greater than the first threshold in the form of a linked list;

Step 3.2, the first line point in the linked list starts from the root node of each red-black tree in turn, and compares the size of the nodes in the red-black tree layer by layer until the node closest to the red-black tree is found;

Step 3.3. Calculate the distance, angle difference, position difference and position intervals in different directions between the line point detection results greater than the first threshold and the closest node.

Step 3.4. According to the calculation results in step 3.3, when the distance, angle difference degree, position difference degree and direction interval all meet the set conditions, merge the current line point with the node in the tree, and jump to step 3.7;

Step 3.5. When only the distance does not meet the set conditions, insert the line point into the tree as a leaf node, perform left or right rotation on the red-black tree according to the balance of the current tree, and jump to step 3.8.

Step 3.6. Compare the line point with the next tree, and then go to step 3.2. If all the trees have been compared, put the line point at the end of the linked list first, and then go to step 3.2.

Step 3.7, check whether the nodes in the merged tree and its adjacent nodes still meet the conditions in step 3.4, if so, go to step 3.8, otherwise go to step 3.6;

Step 3.8, remove the line points from the linked list, and then jump to step 3.2;

Step 3.9. If the linked list has not completed any merging or inserting operations into the tree within a loop, create a new tree with the first line point in the linked list as the root node;

Step 3.10, repeat the above steps until the linked list is empty.

Compared with the prior art, the beneficial effects that can be achieved by at least one of the above-mentioned technical solutions adopted in the embodiments of this specification at least include: using the convolutional neural network to automatically extract the peak features in the radar echo image, which overcomes the traditional CFAR peak detection. Setting the detection threshold does not reduce the problem of missed detection while reducing false alarms. Furthermore, this method relies on the prediction of the direction of the line point, overcomes the problem that the existing power line detection method can only detect straight lines, and can still achieve accurate detection under the condition that the direction of the power line changes.

Description of drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the accompanying drawings that need to be used in the embodiments. Obviously, the accompanying drawings in the following description are only some embodiments of the present application. Those of ordinary skill in the art can also obtain other drawings based on these drawings without any creative effort.

Fig. 1 is a schematic flow chart of an embodiment of the present invention;

Fig. 2 is a schematic flow chart of the power line detection algorithm in the embodiment of the present invention.

Detailed ways

Embodiments of the present application will be described in detail below in conjunction with the accompanying drawings.

It should be noted that, in the case of no conflict, the embodiments in the present application and the features in the embodiments can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and examples.

As shown in Figures 1 and 2, an embodiment of the present invention provides a method for intelligent detection and trend prediction of millimeter-wave radar high-voltage lines, which specifically includes the following steps:

Step 1: Take the radar echo map within one antenna scanning period as the input of the CNN network. Among them, the CNN network adopts the ENet network model. The feature extraction part of the network is composed of seven stage modules, and each stage module is composed of several bottleneck structures. The output of the feature extraction network is respectively connected to two branches, and each branch contains two fully connected layers, respectively predicting the line point confidence at each pixel point and the tangent line angle at the line point. The tangent angle is in radians, and the value range is [0, π). The line point confidence prediction adopts the classification prediction method, and cross entropy is used as the loss function. In order to improve the prediction accuracy of the network for difficult samples, a calculation method similar to the focal loss function (focal loss) is used, but the sampling probability of positive and negative samples can be set during the training process, and the problem of sample imbalance is not easy to occur, so focal loss The calculation items used to alleviate the category imbalance in loss are deleted. The calculation formula of the loss function l _c in category prediction is:

l _c ＝-2|c _p -c _g | ^λ c _g logc _p (1)

Among them, c _p and c _g are the category prediction confidence and the real category result respectively, and λ is the parameter for balancing difficult samples, which is set to 2 in the method.

Angle prediction uses an absolute value loss function. Since the tangent angle prediction results at non-power line points are meaningless, the angle loss function l _a is set to be negatively correlated with the distance between the point and the nearest line point in the label during training, and its calculation formula is:

Among them, a _{p and} a _g are the angle prediction result and the tangent angle of the real line point respectively, d is the distance from the point to the nearest line point in the label, and D is the parameter for adjusting the angle loss, which is set to 30 in the method.

Step 2: Set the threshold to 0.5, remove the points whose confidence is less than or equal to the threshold, and only keep the line point detection results greater than the threshold. In order to further reduce the calculation amount, 25% of the line points are randomly reserved from these remaining line points for subsequent calculation steps;

Step 3: Use the filtered line point confidence and angle prediction results as input, and use the designed multi-curve fitting method to obtain the preliminary detection results of the power line. The multi-curve fitting method is realized based on the red-black tree, and the specific method is as follows:

(3.1) First create an empty tree, and represent all line points in the form of a linked list;

(3.2) The first line point in the linked list starts from the root node of each red-black tree in turn, and compares the size of the nodes in the red-black tree layer by layer until the node closest to the tree is found. The calculation method of the line point and the size of the node in the tree is: first calculate the vertical line of the direction of the node in the tree, and express it with the standard formula of the straight line. Then substitute the line point into the calculation formula of the vertical line. If the result is positive, the line point is considered to be larger than the node in the tree, otherwise the line point is considered to be smaller than the node in the tree.

(3.3) Calculate the distance between the line point and the nearest node, the degree of angle difference, the degree of position difference and the position interval in different directions; where, the distance is divided into the projected distance d _x relative to the tangent direction of the node and d x relative to the tangent of the node Projection distance d _y in the vertical direction. The angle difference d _ang measures the difference between the prediction result of the line point tangent line and the angle prediction result between two adjacent nodes in the tree, and its calculation formula is:

d _ang ＝|aa _left |+|aa _right |-|a _left -a _right | (3)

Where a, a _left , and a _right are the angles of the line point, and the angles of the tree nodes on the left and right sides of the line point respectively. The position difference degree d _pos measures the smoothness between the line point to be inserted and the two adjacent nodes, and is defined as the difference between the angle formed by the line point and the left and right nodes and π. Both d _pos and d _ang are expressed in radians.

The position interval in different directions calculates the distance between the line point and the farthest merged point of the node in the tree in different directions. By eliminating the points that are too far away from the merged point set of the current tree node, a certain range around the correct line point can be suppressed. The purpose of false alarm interference that appears in the In the present invention, a circle of 360° is divided into 36 directions on average, and each direction covers an area of 10°, and only the points falling in this direction and the farthest point from the center point of the node are retained among the merged points. Since the position of the node center will change every time it is merged, in order to reduce the amount of calculation, the node center point always uses the position of the first point in the node as the center point when calculating the position interval index in different directions.

(3.4) According to the calculation results in 3.3, when the projection distance in the tangential direction is less than 20 and the projection distance in the tangential and perpendicular direction is less than 30, if the angle difference is less than 0.6 and the position difference is less than 0.8, the line point and the node in the tree Merge, and jump to 3.7. Merging is carried out in a weighted manner, and the weight is the line point confidence predicted by the CNN model in step 1. The node position and angle update formulas are:

Among them, conf _poi is the confidence degree of the line point, x _poi , y _poi and θ _poi are the horizontal and vertical coordinates and the tangent direction of the line point respectively. x _old , x _new , y _old , y _new are respectively the abscissa of the node in the tree before and after the update and the ordinate before and after the update. θ _new is the tangent direction of the node in the updated tree.

(3.5) If the projection distance in the tangent direction is less than 40, and the projection distance in the tangent and perpendicular direction is less than 50, but the distance does not meet the requirements for merging, and the other indicators meet the requirements, insert the line point into the tree as a leaf node of the current tree node , perform left or right rotation on the red-black tree according to the balance of the current tree, and jump to 3.8;

(3.6) Compare the line point with the next tree, and jump to 3.2. If all the trees have been compared, first place the line point at the end of the linked list, and then jump to 3.2;

(3.7) check whether the node in the merged tree and its adjacent nodes still meet the angle difference degree and position difference degree condition in 3.4, if satisfied, then jump to 3.8, otherwise jump to 3.6;

(3.8) Remove the line points from the linked list and jump to 3.2;

(3.9) If the linked list does not complete any merging or inserting operations into the tree within a loop, a new tree is created with the first line point in the linked list as the root node;

(3.10) loop the above process until the linked list is empty;

Step 4: Perform preorder traversal on each tree obtained in step 3, and connect each node obtained in sequence to obtain the preliminary detection result of the power line in the form of a line segment;

Step 5: Perform a smoothing operation on each power line detection result sequence. When smoothing, keep the end points of the first and last ends of the line unchanged. For the middle node, smoothing depends on the position of a node before and after it. Remember that the coordinates of the point to be smoothed and its left and right nodes are (x ₁ , y ₁ ), (x ₀ , y ₀ ), (x ₂ , y ₂ ), and the calculation formula is:

为平滑后的线点坐标。In the formula v is an intermediate variable,

is the smoothed line point coordinates.

Step 6: Calculate the sum of the confidences of all line points contained in each power line, and arrange the power line sequence in descending order. Accumulate the confidence of the power line sequence, and when the proportion of the accumulated confidence to the total confidence of all power lines exceeds the threshold (set to 0.8 in the method), delete the remaining power lines in the sequence, and delete the power lines with a length less than 80 at the same time. Suppresses the interference of false alarms in the detection results of the peak detection model.

Among them, it should be pointed out that the red-black tree can effectively maintain the balance of the tree and the efficiency of insertion and search when there are many nodes. The ordered data structure can replace the red-black tree to realize the curve fitting function in step 3. At the same time, the CNN-based peak detection function can be realized by using the image segmentation network model with dense point prediction.

The advantage of the embodiment of the present invention is:

(1) Convolutional Neural Network (CNN) is used to pre-extract power line points. In the traditional power line detection method, the constant false alarm detection method is used to extract power line points and tower points, and these methods need to set the detection threshold in advance. If the detection threshold is set too high, it is easy to miss detection, while if the detection threshold is too low, it is easy to generate a large number of false alarms. In this method, the method of deep learning is used to automatically extract the power line points without setting the detection threshold in advance, and the accuracy of power line point detection can be significantly improved.

(2) By predicting the tangent direction of the power line point, the position of the power line can be detected without using the Hough transform, which avoids the assumption that the power line is always a straight line in most existing millimeter-wave radar power line detection methods. At the same time, this method can still achieve a high recognition rate without the power line discrimination step in the traditional method, avoiding excessive dependence on the prior knowledge of high-voltage lines, and has strong scene applicability.

The above is only a specific embodiment of the present invention, and cannot limit the scope of the invention, so the replacement of its equivalent components, or the equivalent changes and modifications made according to the patent protection scope of the present invention, should still fall within the scope of this patent. category. In addition, the technical features and technical features, technical features and technical solutions, and technical solutions and technical solutions in the present invention can be used in free combination.

Claims (4)

1. A millimeter-wave radar high voltage line intelligent detection and trend prediction method, is characterized in that, comprises the following steps: Step 1, using the radar echo image in one antenna scanning period as the input of the convolutional neural network; Step 2, setting a first threshold, and retaining the line point detection results greater than the first threshold; Step 3, using the line point detection results as input, using a multi-curve fitting method to obtain preliminary detection results of multiple power lines; Step 4, perform preorder traversal on the preliminary detection results of each power line, connect each node obtained through the traversal in turn, and obtain the power line detection results expressed in the form of line segments; Step 5, performing a smoothing operation on each power line detection result sequence expressed in the form of a line segment; Step 6: Setting a second threshold, removing electric power lines represented by line segments whose confidence or length is smaller than the second threshold, and outputting remaining electric power lines represented by line segments as detection results. 2. The intelligent detection and trend prediction method for millimeter-wave radar high-voltage lines according to claim 1, wherein said step 1 includes: setting the classification loss function as _lc =-2| _cp - _cg | ^λc _g logc _p , where, c _p , c _g are the category prediction confidence and the real category result respectively, and λ is a parameter for balancing difficult samples. 3. The intelligent detection and direction prediction method for millimeter-wave radar high-voltage lines according to claim 2, characterized in that, said step 1 further includes: setting the angle prediction using an absolute value loss function

where a _p and a _g are the angle prediction result and the tangent angle of the real line point respectively, d is the distance from the point to the nearest line point in the label, and D is the parameter to adjust the angle loss. 4. The intelligent detection and direction prediction method of the millimeter-wave radar high-voltage line according to claim 3, wherein the step 3 includes: Step 3.1, create an empty tree, and represent all line point detection results greater than the first threshold in the form of a linked list; Step 3.2, the first line point in the linked list starts from the root node of each red-black tree in turn, and compares the size of the nodes in the red-black tree layer by layer until the node closest to the red-black tree is found; Step 3.3. Calculate the distance, angle difference, position difference and position intervals in different directions between the line point detection results greater than the first threshold and the closest node. Step 3.4, according to the calculation result in step 3.3, when the distance, angle difference degree, position difference degree and direction interval all meet the set conditions, merge the current line point with the node in the tree, and jump to step 3.7; Step 3.5. When only the distance does not meet the set conditions, insert the line point into the tree as a leaf node, perform left-handed or right-handed operation on the red-black tree according to the balance of the current tree, and jump to step 3.8. Step 3.6. Compare the line point with the next tree, and then go to step 3.2. If all the trees have been compared, put the line point at the end of the linked list first, and then go to step 3.2. Step 3.7, check whether the nodes in the merged tree and its adjacent nodes still meet the conditions in step 3.4, if so, go to step 3.8, otherwise go to step 3.6; Step 3.8, remove the line points from the linked list, and then jump to step 3.2; Step 3.9. If the linked list has not completed any merging or inserting operations into the tree within a loop, create a new tree with the first line point in the linked list as the root node; Step 3.10, repeat the above steps until the linked list is empty.

Images