A deep learning-based visualization method for dynamic linear radar wall in vehicle monitoring

Abstract

The present application provides a method for visualizing a dynamic linear radar wall for vehicle monitoring based on deep learning, constructs an intelligent vehicle monitoring network, constructs a backbone network of the intelligent vehicle monitoring network based on the Resnet model, and introduces a MAAM module and an attention mechanism into the backbone network; the MAAM module can accurately obtain the features of the input image, reduce the amount of calculation and prevent overfitting, thereby improving the recognition accuracy of the intelligent vehicle monitoring network; by introducing a channel attention mechanism in the shallow convolution of the backbone network Resnet and a spatial attention mechanism in the deep convolution, it is ensured that the intelligent vehicle monitoring network can pay more attention to the contour feature information and position feature information in the input image, making the intelligent vehicle monitoring network more suitable for the recognition of trigger objects in images captured by vehicle monitoring lenses.

Description

Technical Field

[0001] The present invention relates to the field of intelligent transportation technology, and in particular to a vehicle-mounted monitoring dynamic linear radar wall visualization method based on deep learning. Background Art

[0002] With the widespread application of intelligent transportation around the world, driverless technology has brought great changes to the automotive industry, and the Advanced Driver Assistance System (ADAS) is one of the key technologies to achieve driverless driving, which can fully ensure the safety of automobile driving. Vehicle-mounted millimeter-wave radar is an important sensor device for ADAS, and plays an irreplaceable role in automobile traffic warning, blind spot detection, adaptive cruise control, etc. Vehicle-mounted millimeter-wave radar can be divided into long-range radar (LRR), medium-range radar (MRR) and short-range radar (SRR) according to the detection distance, and can be divided into 24GHz narrow-band radar (24.00-24.25GHz), 24GHz ultra-wideband radar (24.25-24.65GHz), 77GHz radar (76-77GHz) and 79GHz radar (77-81GHz) according to the frequency of use. Its specific application direction is: combining radar, vision and data fusion software to realize parking assistance, lane change assistance, forward collision warning and other functions. The realization of these functions requires the use of vehicle-mounted radar ranging technology.

[0003] The patent with publication number CN117289281B is a patent technology that our company is currently using. It discloses an intelligent visualization method for vehicle monitoring based on ADAS-DMRW, which can display the radar wall in 2D according to the detection distance of the radar, and can also display the safety distance and alarm distance of the radar wall. However, in actual use, it is found that when the user changes the viewing angle of the radar wall through the virtual camera, the radar wall displayed in 2D will not be displayed. Figure 1 The front of the vehicle is enlarged to enter the perspective of the front micro-view. The red linear structure in the figure is the 2D radar wall in the top view. The display effect of the radar wall in the front micro-view is not ideal. At the same time, based on actual market demand, in addition to displaying the distance change between the vehicle and the obstacle based on the dynamic radar wall, it is also necessary to prompt the specific obstacle type so that the driver can control his driving behavior more accurately. However, the existing intelligent visualization method for vehicle-mounted monitoring cannot distinguish the obstacle type. Summary of the invention

[0004] In order to solve the problems that the existing vehicle-mounted monitoring intelligent visualization method cannot achieve ideal results of 2D radar wall when changing the observation angle and cannot distinguish the types of obstacles, the present invention provides a vehicle-mounted monitoring dynamic linear radar wall visualization method based on deep learning, which can dynamically display the radar wall in 3D and can display the distance between the obstacle and the vehicle based on the dynamic radar wall, and at the same time realize the recognition of the dynamic linear radar wall trigger.

[0005] The technical solution of the present invention is as follows: a method for visualizing a dynamic linear radar wall for vehicle-mounted monitoring based on deep learning, characterized in that it comprises the following steps:

[0006] S1: A dynamic linear radar wall with 3D effect is constructed based on the linear radar wall display method;

[0007] S2: Building an intelligent vehicle monitoring network based on deep learning network model;

[0008] The intelligent vehicle monitoring network includes a backbone network built based on the Resnet model, and a MAAM module and an attention mechanism are introduced into the backbone network;

[0009] The intelligent vehicle monitoring network includes: an input layer, a MAAM module, a first convolution module, a second convolution module, a third convolution module, a fourth convolution module, a global average pooling layer, a fully connected layer and an output layer connected in sequence;

[0010] Each convolution module consists of 4 consecutive convolution layers; 2 residual blocks are introduced in each convolution module;

[0011] In the first convolution module, a channel attention mechanism is introduced in the shallow convolution, and a spatial attention mechanism is introduced in the deep convolution;

[0012] The MAAM module includes: a maximum pooling operation Maxpool and an average pooling operation avgpool, and its expression formula is:

[0013] ;

[0014] Among them, α and β are two hyperparameter coefficients;

[0015] S3: Collect multiple sets of business scene images based on the cameras in front and behind the vehicle body;

[0016] The business scene images include images of radar walls triggered by different triggers;

[0017] S4: Preprocessing the business scene images to obtain a training data set;

[0018] Training the intelligent vehicle monitoring network based on the training data set to obtain the trained intelligent vehicle monitoring network;

[0019] S5: collecting multiple groups of calibration images in real scenes, and performing visual detection on the trained intelligent vehicle monitoring network;

[0020] Repeat steps S3 to S5 until the trained intelligent vehicle monitoring network is obtained;

[0021] S6: Use the trained intelligent vehicle monitoring network to be set in the intelligent vehicle monitoring system, and realize visual monitoring of obstacles based on the dynamic linear radar wall.

[0022] It is further characterized by:

[0023] In the first convolution module, a channel attention mechanism is introduced in the shallow convolution, and a spatial attention mechanism is introduced in the deep convolution;

[0024] The channel attention mechanism enables the network model to pay more attention to the contour feature information, and its output is M c (F):

[0025] ;

[0026] Where σ is the Sigmoid activation function, AvgPool(F) means concatenating the average pooling and MaxPool(F) maximum pooling results along the channel axis; MLP means multi-layer perception mechanism;

[0027] The spatial attention machine enables the network model to pay more attention to the spatial location information in the image or video data, and its output is M s (F):

[0028] ;

[0029] Where σ is the Sigmoid activation function, f^(7*7) is a 7*7 convolution operation, and [AvgPool(F);MaxPool(F)] means concatenating the average pooling and maximum pooling results along the channel axis;

[0030] The trigger objects include: curbstones, grass, trees, roadblocks, animals, people, walls, motor vehicles, non-motor vehicles and other obstacles;

[0031] The dynamic linear radar wall uses the arc line of the radar wall as the bottom line to display the 3D radar wall gradient effect in three dimensions. The specific display method is as follows:

[0032] Construct an adaptive configuration module, wherein the adaptive configuration module dynamically displays the 3D radar wall based on the arc line of the radar wall, through the perspective projection matrix P, combined with the virtual camera position of the real radar wall, and in coordination with the change of the pitch angle of the virtual camera;

[0033] The perspective of the virtual camera simulates the perspective of the driver observing the radar wall;

[0034] The perspective projection matrix P is:

[0035] ;

[0036] Wherein, l, b, n are the XYZ coordinates of the lower left corner of the near plane of the object seen by the virtual camera, r, t, f are the XYZ coordinates of the upper right corner of the far plane of the object seen by the virtual camera; γ is the adaptively configured value after the intelligent vehicle monitoring network determines the triggering radar wall object, γ∈[0,1];

[0037] The calculation method of the radar wall arc line comprises the following steps:

[0038] a1: Build a visual radar wall;

[0039] The visualized radar wall includes: two groups of radar walls located at the front and rear directions of the vehicle to be calculated respectively; each group of radar walls includes: a farthest radar wall and a nearest radar wall parallel to each other;

[0040] The farthest radar wall and the nearest radar wall each include N arc segment-shaped radar wall arc segments;

[0041] Connecting the two end points of the radar wall arc segment with a straight line can obtain a straight line segment, which is recorded as: radar wall segment;

[0042] Any point on the farthest radar wall and the nearest radar wall can find a symmetrical point based on the X-axis and the Y-axis in the visualized radar wall;

[0043] When the distance between the vehicle to be calculated and the obstacle is the farthest distance, the farthest radar wall is displayed; when the distance between the vehicle to be calculated and the obstacle is the closest distance, the closest radar wall is displayed; when the vehicle to be calculated gradually approaches the obstacle from the farthest distance, the dynamic radar wall is displayed;

[0044] The radar wall segments on each radar wall are of equal length, and adjacent radar wall segments are connected at the first position; each radar wall segment includes two coordinate points indicating its start and end points, and the radar wall segment coordinate points constitute the coordinates of the farthest radar wall and the nearest radar wall;

[0045] Each vehicle-mounted radar has a corresponding radar wall segment on the farthest radar wall and the nearest radar wall, and is responsible for scanning between the radar and the radar wall segment;

[0046] The nearest radar wall meets the following conditions:

[0047] Assume that the angle between the lines connecting the two endpoints of the nearest radar wall and the origin of the coordinate system is the FOV of the radar wall. Then the angle α between the line connecting one endpoint of the nearest radar wall and its nearest radar and the X-axis is half of the FOV of the radar wall.

[0048] The distances between the two end points of the nearest radar wall and the radar closest to it are the shortest safety distance n1, and the distance between the intersection of the radar wall and the X-axis and the radar closest to the X-axis is the shortest safety distance n1;

[0049] The farthest radar wall meets the following conditions:

[0050] The distances between the two endpoints of the farthest radar wall and the radar closest to it are the farthest safety distance f1, and the distance between the intersection of the radar wall and the X-axis and the radar closest to the X-axis is the farthest safety distance f1;

[0051] a2: Get the number of vehicle-mounted radars installed on the vehicle to be calculated. Assume that the number of vehicle-mounted radars installed on the vehicle to be calculated is 2N;

[0052] Number each vehicle-mounted radar: start from the front of the vehicle and number the radars in a clockwise direction. The radars installed at the front of the vehicle are numbered from 1 to N, and the radars installed at the rear of the vehicle are numbered from N+1 to 2N.

[0053] The coordinates of the head and tail endpoints of the radar wall segment on the nearest radar wall corresponding to the numbered radar are:

[0054] [NRi (NRXi, NRYi), NLi (NLXi, NLYi)], (NRXi, NRYi) is the head coordinate of the line segment, (NLXi, NLYi) is the tail coordinate of the line segment; i is the radar serial number, and its value is 1, 2.....2N;

[0055] The coordinates of the head and tail endpoints of the radar wall segment on the farthest radar wall corresponding to each radar are:

[0056] [FRi (FRXi, FRYi), FLi (FLXi, FLYi)], where i is 1, 2, ... 2N;

[0057] a3: based on the number of vehicle-mounted radars, the radar wall FOV and the radar detection distance, determine the farthest safety distance f1 and the closest safety distance n1;

[0058] a4: According to the known conditions of radar wall FOV, shortest safety distance n1 and longest safety distance f1, a data relationship is constructed. When N has different values, the coordinates of each point on the radar wall segment corresponding to all the radar wall arc segments are calculated according to the mathematical relationship;

[0059] a5: calculating the radar wall arc line according to the coordinates of each point on the radar wall segment and the center angle of each radar wall arc line;

[0060] In step a5, when N=4, the calculation method of the radar wall arc line is as follows:

[0061] b1: The radars installed on the vehicle to be calculated include: radars A, B, C, D installed on the front of the vehicle and radars D', C', B', A' installed on the rear of the vehicle, and the corresponding serial numbers are 1 to 8;

[0062] The following point coordinates are obtained by calculation:

[0063] The endpoint coordinates of the radar wall segment corresponding to the nearest radar wall:

[0064] [(NRXi,NRYi)、(NLXi,NLYi)], i is 1, 2, ... 8;

[0065] The endpoint coordinates of the radar wall segment corresponding to the farthest radar wall:

[0066] [(FRXi, FRYi), (FLXi, FLYi)], where i is 1, 2, ... 8;

[0067] Vehicle radar coordinates: A: (a1, b1), B: (a2, b2);

[0068] b2: confirm the center angle of each arc line of the radar wall;

[0069] b3: According to the positional relationship between the vehicle-mounted radar and the radar wall, the following relationship is established:

[0070] NRY1 = k1*NRX1 + c1;

[0071] FRY1 = k2*FRX1 + c2;

[0072] The values ​​of k1, k2, c1 and c2 are obtained by solving;

[0073] b4: solve the arc length of the radar wall;

[0074] Assume that point J is the midpoint of the radar wall segment [NR1 (NRX1, NRY1), NL1 (NLX1, NLY1)], then:

[0075] The coordinates of point J are: ((NLX1+NRX1) / 2,(NRY1+NLY1) / 2);

[0076] Assume that point O is the center of the radar wall arc [NR1, NL1], then connect point J and point O to get the straight line y3: y3=k3*x+c3;

[0077] Based on the known conditions: the distances from the points on the perpendicular bisector to the points at both ends are the same, and the central angle of each arc line of the radar wall is known;

[0078] According to the mathematical relationship, we can get: line segment OJ, line segment ONL1, line segment ONR1 and line segment NL1NR1;

[0079] Based on the known coordinates of the endpoints of the radar wall segment, the values ​​of k3 and c3 are obtained by calculation;

[0080] b5: Calculate the coordinates of point O through the straight line y3=k3*x+c3, the coordinates of point J, and the distance of line segment OJ;

[0081] Then, according to the coordinates of the center O and the radius ONL1 of the circle, the coordinates and length of the radar wall arc line [NL1, NR1] are calculated;

[0082] In step b5, assuming that the central angle of each radar wall arc is 60°, the following relationship holds:

[0083] Line segment ONL1 = line segment ONR1 = line segment NL1NR1;

[0084] Then, based on the mathematical relationship, we can calculate:

[0085] k3=-1 / ((NRY1-NLY1) / (NRX1-NLX1))=(NLX1-NRX1) / (NRY1-NLY1);

[0086] c3=(NRY1+NLY1) / 2-((NLX1-NRX1)*(NLX1+NRX1)) / 2(NRY1-NLY1);

[0087] y3=((NLX1-NRX1) / (NRY1-NLY1))*x+(NRY1+NLY1) / 2-((NLX1-NRX1)*(NLX1+NRX1)) / 2(NRY1-NLY1).

[0088] The present application provides a method for visualizing a dynamic linear radar wall for on-board monitoring based on deep learning. The intelligent on-board monitoring network constructed by the method comprises a backbone network of the intelligent on-board monitoring network based on a Resnet model, and a MAAM module and an attention mechanism are introduced into the backbone network. The MAAM module can accurately obtain the features of the input image, reduce the amount of calculation and prevent overfitting, thereby improving the recognition accuracy of the intelligent on-board monitoring network. The channel attention mechanism is introduced into the shallow convolution of the backbone network Resnet, and the spatial attention mechanism is introduced into the deep convolution, thereby ensuring that the intelligent on-board monitoring network can pay more attention to the contour feature information and position feature information in the input image, thereby making the intelligent on-board monitoring network more suitable for the recognition of triggers in images collected by on-board monitoring lenses. The method constructs a dynamic linear radar wall based on a linear radar wall display method to display the distance between obstacles and vehicles. On this basis, the dynamic linear radar wall trigger can be accurately identified based on the intelligent on-board monitoring network, thereby greatly improving the usability of the on-board dynamic radar wall. BRIEF DESCRIPTION OF THE DRAWINGS

[0089] Figure 1 This is a comparison chart of the effect of changing the viewing angle of a 2D radar wall in the prior art;

[0090] Figure 2 This is a diagram of the intelligent vehicle monitoring network model in this application;

[0091] Figure 3 This is the MAAM module structure diagram;

[0092] Figure 4 It is the residual block structure diagram;

[0093] Figure 5 Design diagram for a linear radar wall with four front and four rear radars;

[0094] Figure 6 Design diagram for a linear radar wall with three radars in front and three in the back;

[0095] Figure 7 Design diagram for a linear radar wall with five front and five rear radars;

[0096] Figure 8 This is an example of the 3D view radar wall rendering in this application;

[0097] Fig. 9 A comparison chart of the detection effects of different triggers triggering the intelligent vehicle monitoring network;

[0098] Fig.10 A comparison chart of the display effects of the 3D radar wall in this application for trigger objects at different heights. DETAILED DESCRIPTION

[0099] The present application includes a deep learning-based vehicle monitoring dynamic linear radar wall visualization method, which includes the following steps.

[0100] S1: A dynamic linear radar wall with 3D effect is constructed based on the linear radar wall display method.

[0101] S2: Build an intelligent vehicle monitoring network based on deep learning network model.

[0102] The backbone network used in the deep learning-based intelligent vehicle monitoring network of this application is Resnet-MAAM, in which the activation function is the RELU function, and in order to improve the recognition accuracy, the channel attention module is introduced in the shallow convolution, and the spatial attention module is introduced in the deep convolution, forming a new network structure model.

[0103] like Figure 1 As shown, the intelligent vehicle monitoring network includes: an input layer, a MAAM module, a first convolution module, a second convolution module, a third convolution module, a fourth convolution module, a global average pooling layer, a fully connected layer and an output layer connected in sequence.

[0104] Each convolution module consists of 4 consecutive convolution layers; 2 residual blocks are introduced in each convolution module, and each residual block consists of two consecutive convolution layers and a skip connection.

[0105] In the first convolution module, the channel attention mechanism is introduced in the shallow convolution, and the spatial attention mechanism is introduced in the deep convolution.

[0106] Input the monitoring images received by the vehicle monitoring system into Figure 1 In the intelligent vehicle monitoring network model shown, the type of trigger is identified, which specifically includes the following steps:

[0107] Input layer: receives RGB images of size 224*224;

[0108] Convolutional layer: 4 convolutional layers, each of which uses a 3*3 convolution kernel and Relu activation function to extract local features of the image;

[0109] Residual blocks: 8 residual blocks, each of which consists of two convolutional layers and a skip connection to solve the problems of gradient explosion and gradient disappearance;

[0110] Global average pooling layer: performs global average pooling on the feature map and converts the feature map into a one-dimensional vector;

[0111] Fully connected layer: contains a fully connected layer of size 1000 for classification output;

[0112] Output layer: Use the softmax activation function to generate the probability distribution of the trigger category; in this embodiment, all triggers are divided into 10 categories, including: people, motor vehicles, curbs, grass, trees, walls, roadblocks, non-motor vehicles, animals and other obstacles.

[0113] Figure 1 In the above figure, MAA-pool-Moudle is the MAAM module introduced above; RELU is the activation function, which is expressed as: f(x)=max(0,x), where x is the input value and f(x) is the output of the activation function. It has nonlinear properties, sparsity, high computational efficiency and alleviates the gradient vanishing problem.

[0114] Conc7*7: 7*7 convolution; 64, 128, 256, 512 are the number of channels; 112, 56, 28, 14, 7 are the image sizes, for example: 112*112, 56*56; stride=2 / 1: the step size is 2 / 1; padding=0: padding is 0.

[0115] FC is the fully connected layer, 512 is consistent with the number of channels of the average pooling layer, and 1000 is the size of the fully connected layer.

[0116] avgpool: average pooling layer doublepool: bidirectional pooling layer; reduce parameters through pooling layer.

[0117] like Figure 2 As shown in the figure, the MAAM module includes: the maximum pooling operation Maxpool and the average pooling operation avgpool. The 2*2 four-color on the left side of the figure is Maxpool; the blue part is the feature map of different channels; the 1*4 four-color module on the right is Avgpool; the standard Avgpool is square, that is, n*n. For a specific data set, a deformed Avgpool, 1*n 2 By performing weighted averaging with the normal Maxpool, the features of the image can be accurately obtained, the amount of calculation can be reduced, and overfitting can be prevented, so that the training effect can be better.

[0118] The MAAM module expression formula is:

[0119] ;

[0120] Among them, α and β are two hyper-parameter coefficients. In this embodiment, α=0.6 and β=0.4 are used for training, and the result is optimal.

[0121] In this application, MAAM is introduced into the Resnet model; MAAM can better maintain the statistical characteristics of the features by calculating the average value within the region, making the model more robust to small changes in the features, and is suitable for scenarios where feature stability needs to be maintained. Since MAAM does not involve parameter selection, it can better prevent overfitting during training and improve the generalization ability of the model; at the same time, MAAM smoothes the feature map to reduce the impact of noise and make the features smoother and continuous; MAAM can better extract significant local features by selecting the maximum value within the region, and is suitable for scenarios where significant features need to be emphasized; and because MAAM selects the maximum value calculation, it can significantly reduce the amount of calculation of subsequent layers and improve the computational efficiency of the model; MAAM introduces nonlinear characteristics, which helps the model learn more complex feature representations.

[0122] In the first convolution module, the channel attention mechanism is introduced in the shallow convolution and the spatial attention mechanism is introduced in the deep convolution;

[0123] The channel attention mechanism enables the network model to pay more attention to the contour feature information, and its output is M c (F):

[0124] ;

[0125] Where σ is the Sigmoid activation function, AvgPool(F) means concatenating the average pooling and MaxPool(F) maximum pooling results along the channel axis; MLP means multi-layer perception mechanism;

[0126] The spatial attention machine enables the network model to pay more attention to the spatial location information in the image or video data, and its output is M s (F):

[0127] ;

[0128] Where σ is the Sigmoid activation function, f^(7*7) is a 7*7 convolution operation, and [AvgPool(F);MaxPool(F)] means concatenating the average pooling and maximum pooling results along the channel axis.

[0129] The residual block structure is as follows Figure 3 As shown in the figure, Identity is the identity mapping, F(x) is the residual, weightlayer is the ordinary convolution, and relu is the activation function. In the ResNet model, the output of a certain layer is allowed to directly skip one or more layers and connect to the input of the subsequent layers, ensuring that some layers do not make any meaningful changes, but can still transmit shallow information without causing too much loss to the gradient.

[0130] The residual block is expressed as:

[0131] H(x)=F(x)+x;

[0132] Among them, H(x) is the observed value; x is the estimated value, which is a skip connection, and the feature response of the previous layer output; F(x) is the residual;

[0133] When F(x)=0, an identity mapping H(x)=x is formed, ensuring easier fitting of the residual.

[0134] S3: Collect multiple sets of business scene images based on the cameras in front and behind the vehicle body;

[0135] The business scene images include images of radar walls triggered by different triggers.

[0136] Specifically, multiple sets of business scene images are collected in different cars at different times and locations using the front and rear cameras of the car body. The images include images of radar walls triggered by people walking, images of radar walls triggered by dynamic and static vehicles, and images of other radar walls triggered by roadblocks. The areas of people, cars, roadblocks and other radar walls triggered in the images are marked, and then preprocessed (randomly enhanced) operations are performed, including resizing, random translation in horizontal and vertical directions, random rotation, random hue change, random brightness change, random contrast change, and random addition of masks (simulating occlusion and stains).

[0137] Since the size of the scene input image for training is 1920*1536*3, and the Resnet18 input is 224*224*3, you can resize the scene image to 224*224*3 by simply using the resize operation in the code.

[0138] S4: Preprocess the business scene images to obtain a training data set;

[0139] The intelligent vehicle monitoring network is trained based on the training data set to obtain the trained intelligent vehicle monitoring network.

[0140] In the specific training process, the back propagation algorithm is used to perform fast iterative convergence using the Adam optimizer. The loss function used is the cross entropy loss function:

[0141] ;

[0142] epcho is 100, which means 100 rounds of training. Based on the accuracy after training, set the business logic and distinguish 10 types of triggers.

[0143] S5: Collect multiple sets of calibration images in real scenes and perform visual detection on the trained intelligent vehicle monitoring network;

[0144] Steps S3 to S5 are repeated until a trained intelligent vehicle monitoring network is obtained.

[0145] S6: Use the trained intelligent vehicle monitoring network to set up the intelligent vehicle monitoring system and realize visual monitoring of obstacles based on the dynamic linear radar wall.

[0146] This method is applicable to the identification of radar wall triggers in various scenarios such as complex backgrounds, standard and non-standard, distorted and non-distorted. In actual scenarios, text and voice warning messages can be added based on the trigger recognition results to remind the driver: Warning, there is an obstacle in a certain direction of the vehicle, please pay attention to the surrounding environment. Examples of the application of radar trigger recognition based on intelligent vehicle monitoring networks in real scenarios are as follows: Fig. 9 As shown, when the trigger object is a person, the warning message will prompt: there is a person behind the vehicle; when the trigger object is a motor vehicle, the warning message will prompt that there is a motor vehicle behind the vehicle; when the trigger object is a tree, the prompt message will prompt: there is a tree behind the vehicle.

[0147] The dynamic linear radar wall in the present application is based on a plurality of arc segments connected end to end to form an arc-shaped radar wall. The two end points of each arc segment are connected to form a straight line segment, and each radar wall includes a plurality of arc segments, that is, each radar wall corresponds to a plurality of straight line segments connected end to end, and these straight line segments can be obtained by any existing algorithm for solving the straight line segment radar wall. This method continues to calculate the coordinates of the points on the arc line corresponding to each arc segment on the basis of the straight line segment, and then obtains the arc-shaped radar wall. In this embodiment, the arc-shaped radar wall composed of arc segments is a 3D dynamic radar wall obtained by further improvement on the 2D radar wall implemented by the existing ADAS-DMRW algorithm of our company.

[0148] The dynamic linear radar wall uses the arc line of the radar wall as the bottom line, and displays the 3D radar wall gradient effect in three dimensions. The specific display method is as follows:

[0149] Construct an adaptive configuration module, first calculate the specific coordinates of the arc line of the radar wall, and use the adaptive configuration module based on the arc line of the radar wall, through the perspective projection matrix P, combined with the virtual camera position of the real radar wall, and the change of the pitch angle of the virtual camera to dynamically display the 3D radar wall. Among them, the perspective of the virtual camera simulates the perspective of the driver observing the radar wall.

[0150] The perspective projection matrix P in this application is:

[0151] ;

[0152] In specific implementation, l, b, n, r, t, f, the distance between the virtual camera and the actual object, and the pitch angle of the virtual camera can be configured according to the real vehicle data; among them, l, b, n are the XYZ point coordinates of the lower left corner of the near plane of the object seen by the virtual camera, and r, t, f are the XYZ point coordinates of the upper right corner of the far plane of the object seen by the virtual camera; by obtaining the lower left corner of the near plane and the upper right corner of the far plane of the object, combined with the configuration of the virtual camera position and the pitch angle of the virtual camera, the parameters with the best radar wall effect can be adjusted; γ is the adaptive configuration value after the intelligent vehicle monitoring network Resnet-MAAM determines the triggering radar wall object, γ∈[0,1]. By modifying the coefficients of the perspective projection matrix, the display effect of the radar wall can be made more realistic and accurate.

[0153] In this application, the radar wall is optimized in terms of visual display. The usual matrices include perspective projection matrix and orthogonal projection matrix. According to the calculated linear radar wall point coordinates, after using the orthogonal projection matrix to convert them into the coordinates displayed on the screen, the display effect is poor, and the stereoscopic effect cannot be achieved under the micro-viewing angle. For example Figure 1 As shown; after using the perspective projection matrix to convert the calculated linear radar wall point coordinates to those displayed on the screen, the display effect is more optimized than the orthogonal projection matrix, but the three-dimensional effect is still not obvious under the micro-viewing angle. Therefore, this application optimizes the perspective matrix P according to the adaptively configured value after the Resnet-MAAM network determines the triggering radar wall object, which can make the radar wall effect under the micro-viewing angle better.

[0154] The specific method for setting γ is to set the γ value corresponding to each trigger object according to the height change of the trigger object. In this embodiment, 0~1 is divided into 10 equal parts, and the γ value corresponding to each trigger object is assigned based on the order of: curbstone, grass, tree, roadblock, animal, person, wall, motor vehicle, non-motor vehicle, and other obstacles, for example: γ 人 = 0.1, γ 机动车 =0.2......γ 其他障碍物 =1, the perspective projection matrix is ​​optimized through γ, so that the viewing angle change effect of the 3D radar wall is more adapted to the type of trigger, thereby improving the display effect of the 3D radar wall. Fig.10 Shown is a comparison chart of the display effects of the 3D radar wall in this application for different triggers. The figure compares the two effects of the trigger being a curb and a wall. It can be seen that the effect of the red radar wall has a significant display difference. In the micro-viewing angle of the front of the vehicle, in order to allow the driver to see the object that triggers the radar wall in the blind spot more clearly, this method designs radar wall effects of different heights for different types of triggers; to prompt the driver; when it is unclear, you can switch to other perspectives to further confirm the triggered object to avoid the problem of the vehicle being scratched by objects that are lower than the curb.

[0155] The specific calculation method of the radar wall arc line includes the following steps.

[0156] a1: Build a visual radar wall. Figure 5 , Figure 6 , Figure 7 As shown, the radar wall in this application has the following features:

[0157] The visualized radar wall includes: two groups of radar walls located at the front and rear directions of the vehicle to be calculated; each group of radar walls includes: a farthest radar wall and a nearest radar wall parallel to each other;

[0158] The farthest radar wall and the nearest radar wall each include N arc segments of radar walls in the shape of arc segments;

[0159] Connecting the two end points of the radar wall arc segment with a straight line can obtain a straight line segment, which is recorded as: radar wall segment;

[0160] Any point on the farthest radar wall and the nearest radar wall can find a symmetrical point based on the X-axis and Y-axis in the visualized radar wall;

[0161] When the distance between the vehicle to be calculated and the obstacle is the farthest, the farthest radar wall is displayed; when the distance between the vehicle to be calculated and the obstacle is the shortest, the shortest radar wall is displayed; when the vehicle to be calculated gradually approaches the obstacle from the farthest distance, the dynamic radar wall is displayed;

[0162] The radar wall segments on each radar wall are of equal length, and adjacent radar wall segments are connected at the first position; each radar wall segment includes two coordinate points indicating its start and end points, and the radar wall segment coordinate points constitute the coordinates of the farthest radar wall and the nearest radar wall;

[0163] Each vehicle-mounted radar has a corresponding radar wall segment on the farthest radar wall and the nearest radar wall, and is responsible for scanning between the radar and the radar wall segment;

[0164] The latest radar wall meets the following conditions:

[0165] Assume that the angle between the two connecting lines of the nearest two end points of the radar wall and the origin of the coordinate system is the radar wall FOV. Then the angle α between the connecting line of one end point of the nearest radar wall and its nearest radar and the X-axis is half of the radar wall FOV. Radar wall FOV: is the horizontal detection range angle covered by the preset radar wall.

[0166] The distances between the two endpoints of the nearest radar wall and the radar closest to it are the shortest safety distance n1, and the distance between the intersection of the radar wall and the X-axis and the radar closest to the X-axis is the shortest safety distance n1;

[0167] The farthest radar wall meets the following conditions:

[0168] The distances between the two end points of the farthest radar wall and the radar closest to it are the farthest safety distance f1, and the distance between the intersection of the radar wall and the X-axis and the radar closest to the X-axis is the farthest safety distance f1.

[0169] a2: Get the number of vehicle-mounted radars installed on the vehicle to be calculated. Assume that the number of vehicle-mounted radars installed on the vehicle to be calculated is 2N;

[0170] Number each vehicle-mounted radar: start from the front of the vehicle and number the radars in a clockwise direction. The radars installed at the front of the vehicle are numbered from 1 to N, and the radars installed at the rear of the vehicle are numbered from N+1 to 2N.

[0171] The coordinates of the head and tail endpoints of the radar wall segment on the nearest radar wall corresponding to the numbered radar are:

[0172] [NRi (NRXi, NRYi), NLi (NLXi, NLYi)], (NRXi, NRYi) is the head coordinate of the line segment, (NLXi, NLYi) is the tail coordinate of the line segment; i is the radar serial number, and its value is 1, 2.....2N;

[0173] The coordinates of the head and tail endpoints of the radar wall segment on the farthest radar wall corresponding to each radar are:

[0174] [FRi (FRXi, FRYi), FLi (FLXi, FLYi)], where i is 1, 2.....2N.

[0175] a3: Based on the number of vehicle-mounted radars, radar wall FOV and radar detection distance, determine the farthest safe distance f1 and the closest safe distance n1.

[0176] a4: According to the known conditions of radar wall FOV, shortest safety distance n1 and longest safety distance f1, a data relationship is constructed. When N has different values, the coordinates of each point on the radar wall segment corresponding to all radar wall arc segments are calculated according to the mathematical relationship;

[0177] The constructed data relationship is as follows:

[0178] The line connecting point NR1 (NRX1, NRY1) and the origin of the coordinate system is denoted as: Fo1;

[0179] The line connecting point NL1 (NLXN, NLYN) and the origin of the coordinate system is denoted as: FoN;

[0180] The angle between Fo1 and FoN is the radar wall FOV, denoted as: angle Fov;

[0181] The angle between the line connecting point NR1 (NRX1, NRY1) and radar A (a1, b1) and the X-axis is α; then there is a relationship: α=Fov / 2;

[0182] According to the radar wall FOV, construct data relationship:

[0183] The line connecting point NR1 (NRX1, NRY1) and the origin of the coordinate system is denoted as: Fo1;

[0184] The line connecting point NL1 (NLXN, NLYN) and the origin of the coordinate system is denoted as: FoN;

[0185] The angle between Fo1 and FoN is the radar wall FOV, denoted as: angle Fov;

[0186] The angle between the line connecting point NR1 (NRX1, NRY1) and radar A (a1, b1) and the X-axis is α; then there is a relationship: α=Fov / 2;

[0187] Construct the data relationship based on the shortest safety distance n1 and the longest safety distance f1:

[0188] The two endpoints of the nearest radar wall in the positive direction of the x-axis are: point (NRX1, NRY1) and point (NLXN, NLYN);

[0189] The distance between point NR1 (NRX1, NRY1) and radar 1 is recorded as: NearD1;

[0190] The distance between point NLN (NLXN, NLYN) and radar N is recorded as: NearDN;

[0191] Then: NearD1= NearDN=n1;

[0192] The two endpoints of the farthest radar wall in the positive x-axis direction are: point (FRX1, FRY1) and point (FLXN, FLYN);

[0193] The distance between point FR1 (FRX1, FRY1) and radar 1 is recorded as: FarD1;

[0194] The distance between point FLN (FLXN, FLYN) and radar N is recorded as: FarDN;

[0195] Then: FarD1= FarDN =f1.

[0196] a5: Calculate the radar wall arc line according to the coordinates of each point on the radar wall segment and the center angle of each radar wall arc line.

[0197] When N=4, the calculation method of the radar wall arc line is as follows:

[0198] b1: The radars installed on the vehicle to be calculated include: radars A, B, C, D installed on the front of the vehicle and radars D', C', B', A' installed on the rear of the vehicle, and the corresponding serial numbers are 1 to 8;

[0199] The following point coordinates are obtained by calculation:

[0200] The endpoint coordinates of the radar wall segment corresponding to the nearest radar wall:

[0201] [(NRXi,NRYi)、(NLXi,NLYi)], i is 1, 2, ... 8;

[0202] The endpoint coordinates of the radar wall segment corresponding to the farthest radar wall:

[0203] [(FRXi, FRYi), (FLXi, FLYi)], where i is 1, 2, ... 8;

[0204] Vehicle-mounted radar coordinates: A: (a1, b1), B: (a2, b2).

[0205] b2: Confirm the center angle of each radar wall arc.

[0206] The center angle of each radar wall arc line in the method is preset according to the vehicle type and the number of vehicle-mounted radars. In the radar wall display method of the method, the center angle of the radar wall arc line is used as a preset condition, and the specific position of the radar wall is adjusted according to the radius corresponding to the radar wall arc line.

[0207] b3: According to the positional relationship between the vehicle-mounted radar and the radar wall, the following relationship is established:

[0208] NRY1 = k1*NRX1 + c1;

[0209] FRY1 = k2*FRX1 + c2;

[0210] Combined with the principle that the distances between the two segments of the line segment are equal based on the perpendicular bisector of the line segment, the values ​​of k1, k2, c1 and c2 are obtained by solving; the specific calculation process is implemented based on the existing method.

[0211] b4: solve the arc length of the radar wall;

[0212] Assume that point J is the midpoint of the radar wall segment [NR1 (NRX1, NRY1), NL1 (NLX1, NLY1)], then:

[0213] The coordinates of point J are: ((NLX1+NRX1) / 2,(NRY1+NLY1) / 2);

[0214] Assume that point O is the center of the radar wall arc [NR1, NL1], then connect point J and point O to get the straight line y3: y3=k3*x+c3;

[0215] Based on the known conditions: the distances from the points on the perpendicular bisector to the points at both ends are the same, and the central angle of each radar wall arc is known;

[0216] According to the mathematical relationship, we can get: line segment OJ, line segment ONL1, line segment ONR1 and line segment NL1NR1;

[0217] Based on the known condition that the endpoint coordinates of the radar wall segment are known, the values ​​of k3 and c3 are obtained by calculation.

[0218] b5: Calculate the coordinates of point O through the straight line y3=k3*x+c3, the coordinates of point J, and the distance of line segment OJ;

[0219] Then, according to the coordinates of the center O and the radius ONL1 of the circle, the coordinates and length of the radar wall arc line [NL1, NR1] are calculated.

[0220] In this embodiment, assuming that the center angle of each radar wall arc is 60°, the following relationship holds:

[0221] Line segment ONL1 = line segment ONR1 = line segment NL1NR1;

[0222] For the straight line y3: y3=k3*x+c3, the product of the slope k3 of Y3 and the slope of the line segment NL1NR1 is -1;

[0223] Then k3=-1 / ((NRY1-NLY1) / (NRX1-NLX1))=(NLX1-NRX1) / (NRY1-NLY1);

[0224] That is: y3=k3*x+c3;

[0225] Substituting the xy coordinates of point J into y3, we get:

[0226] c3=(NRY1+NLY1) / 2-((NLX1-NRX1)*(NLX1+NRX1)) / 2(NRY1-NLY1);

[0227] Then we have:

[0228] y3=((NLX1-NRX1) / (NRY1-NLY1))*x+(NRY1+NLY1) / 2-((NLX1-NRX1)*(NLX1+NRX1)) / 2(NRY1-NLY1).

[0229] In this method, the coordinates of point O can be calculated through the straight line y3=k3*x+c3, the coordinates of point J and the distance of line segment OJ; according to the coordinates of the center O and the radius ONL1 of the circle, the arc range in the X direction is between NLX1 and NRX1, and the arc range in the Y direction is between NLY1 and NRY1; among them, we set the rotation radius of ONL1 to be controllable, which can affect the strength of the arc.

[0230] Similarly, the coordinates of points c4 and O', the coordinates of the center O', and the radius O'FL1 of the circle can be calculated. The arc range in the X direction is between FLX1 and FRX1, and the arc range in the Y direction is between FLY1 and FRY1. Among them, setting the rotation radius of O'FL1 to be controllable can affect the strength of the arc.

[0231] Similarly, the same calculation method can be used to change the point coordinates to calculate the far and near linear radar walls of the four radar walls in the front and the linear radar wall in the rear.

[0232] like Figure 6 As shown, when there are three front and three rear radars, compared with the calculation formula of the front and four rear radar walls, the two curves y3 and y4 are missing, but the calculation method of the formula is the same, and only 6 calculations are needed for the forward and backward directions respectively.

[0233] like Figure 7 As shown, the calculation formula for the linear radar wall of the front five and rear five radars is exactly the same as that of the front four and rear four radar walls. It should be noted that when there are four radars, the forward direction needs to be calculated 8 times, and when there are five radars, the forward direction needs to be calculated 10 times. The formula is the same, and four points NL5, NR5, FL5, and FR5 are added; the rear direction is the same as the forward direction.

[0234] The display effect of the 3D radar wall constructed by this method is as follows: Figure 8 As shown, it can be seen that the 3D display effects of the 3D dynamic radar wall of the present application are clearly different under different viewing angles.

[0235] After using the technical solution of the present invention, the safety distance and alarm distance of the radar wall can be accurately displayed according to the detection distance of the radar, and it is flexibly applicable to different types of vehicles and different vehicle-mounted radar installation scenarios. On this basis, combined with the deep learning algorithm, the specific situation of obstacles such as people, vehicles, roadblocks, etc. near the vehicle-mounted radar can be accurately displayed, and the driver can be reminded, thereby improving the usability of the radar wall.

Claims

1. A method for visualizing a dynamic linear radar wall for vehicle monitoring based on deep learning, characterized in that: It includes the following steps: S1: A dynamic linear radar wall with 3D effect is constructed based on the linear radar wall display method; S2: Building an intelligent vehicle monitoring network based on deep learning network model; The intelligent vehicle monitoring network includes a backbone network built based on the Resnet model, and a MAAM module and an attention mechanism are introduced into the backbone network; The intelligent vehicle monitoring network includes: an input layer, a MAAM module, a first convolution module, a second convolution module, a third convolution module, a fourth convolution module, a global average pooling layer, a fully connected layer and an output layer connected in sequence; Each convolution module consists of 4 consecutive convolution layers; 2 residual blocks are introduced in each convolution module; In the first convolution module, a channel attention mechanism is introduced in the shallow convolution, and a spatial attention mechanism is introduced in the deep convolution; The MAAM module includes: a maximum pooling operation Maxpool and an average pooling operation avgpool, and its expression formula is: ； Among them, α and β are two hyperparameter coefficients; S3: Collect multiple sets of business scene images based on the cameras in front and behind the vehicle body; The business scene images include images of radar walls triggered by different triggers; S4: Preprocessing the business scene images to obtain a training data set; Training the intelligent vehicle monitoring network based on the training data set to obtain the trained intelligent vehicle monitoring network; S5: collecting multiple groups of calibration images in real scenes, and performing visual detection on the trained intelligent vehicle monitoring network; Repeat steps S3 to S5 until the trained intelligent vehicle monitoring network is obtained; S6: using the trained intelligent vehicle monitoring network to be set in an intelligent vehicle monitoring system, and realizing visual monitoring of obstacles based on the dynamic linear radar wall; The dynamic linear radar wall uses the arc line of the radar wall as the bottom line to display the 3D radar wall gradient effect in three dimensions. The specific display method is as follows: Construct an adaptive configuration module, wherein the adaptive configuration module dynamically displays the 3D radar wall based on the arc line of the radar wall, through the perspective projection matrix P, combined with the virtual camera position of the real radar wall, and in coordination with the change of the pitch angle of the virtual camera; The perspective of the virtual camera simulates the perspective of the driver observing the radar wall; The perspective projection matrix P is: ； Among them, l, b, n are the XYZ point coordinates of the lower left corner of the near plane of the object seen by the virtual camera, r, t, f are the XYZ point coordinates of the upper right corner of the far plane of the object seen by the virtual camera; γ is the value of the adaptive configuration after the intelligent vehicle monitoring network determines the triggering radar wall object, γ∈[0,1].

2. According to claim 1, a method for visualizing a dynamic linear radar wall for vehicle monitoring based on deep learning, characterized in that: In the first convolution module, a channel attention mechanism is introduced in the shallow convolution, and a spatial attention mechanism is introduced in the deep convolution; The channel attention mechanism enables the network model to pay more attention to the contour feature information, and its output is M c (F): ； Where σ is the Sigmoid activation function, AvgPool(F) means concatenating the average pooling and MaxPool(F) maximum pooling results along the channel axis; MLP means multi-layer perception mechanism; The spatial attention machine enables the network model to pay more attention to the spatial location information in the image or video data, and its output is M s (F): ； Where σ is the Sigmoid activation function, f^(7*7) is a 7*7 convolution operation, and [AvgPool(F);MaxPool(F)] means concatenating the average pooling and maximum pooling results along the channel axis.

3. According to claim 1, a method for visualizing a vehicle-mounted monitoring dynamic linear radar wall based on deep learning, characterized in that: The trigger objects include: curbstones, grass, trees, roadblocks, animals, people, walls, motor vehicles, non-motor vehicles and other obstacles.

4. According to claim 1, a method for visualizing a dynamic linear radar wall for vehicle monitoring based on deep learning, characterized in that: The calculation method of the radar wall arc line comprises the following steps: a1: Build a visual radar wall; The visualized radar wall includes: two groups of radar walls located at the front and rear directions of the vehicle to be calculated respectively; each group of radar walls includes: a farthest radar wall and a nearest radar wall parallel to each other; The farthest radar wall and the nearest radar wall each include N arc segment-shaped radar wall arc segments; Connect the two end points of the radar wall arc segment with a straight line to obtain a straight line segment, which is recorded as: radar wall segment; Any point on the farthest radar wall and the nearest radar wall can find a symmetrical point based on the X-axis and the Y-axis in the visualized radar wall; When the distance between the vehicle to be calculated and the obstacle is the farthest, the farthest radar wall is displayed; when the distance between the vehicle to be calculated and the obstacle is the shortest, the shortest radar wall is displayed; when the vehicle to be calculated gradually approaches the obstacle from the farthest distance, the dynamic radar wall is displayed; The radar wall segments on each radar wall are of equal length, and adjacent radar wall segments are connected at the first position; each radar wall segment includes two coordinate points indicating its start and end points, and the radar wall segment coordinate points constitute the coordinates of the farthest radar wall and the nearest radar wall; Each vehicle-mounted radar has a corresponding radar wall segment on the farthest radar wall and the nearest radar wall, and is responsible for scanning between the radar and the radar wall segment; The nearest radar wall meets the following conditions: Assume that the angle between the lines connecting the two endpoints of the nearest radar wall and the origin of the coordinate system is the radar wall FOV, then the angle α between the line connecting one endpoint of the nearest radar wall and its nearest radar and the X-axis is half of the radar wall FOV; The distances between the two end points of the nearest radar wall and the radar closest to it are the shortest safety distance n1, and the distance between the intersection of the radar wall and the X-axis and the radar closest to the X-axis is the shortest safety distance n1; The farthest radar wall meets the following conditions: The distances between the two endpoints of the farthest radar wall and the radar closest to it are the farthest safety distance f1, and the distance between the intersection of the radar wall and the X-axis and the radar closest to the X-axis is the farthest safety distance f1; a2: Get the number of vehicle-mounted radars installed on the vehicle to be calculated. Assume that the number of vehicle-mounted radars installed on the vehicle to be calculated is 2N; Number each vehicle-mounted radar: start from the front of the vehicle and number the radars in a clockwise direction. The radars installed at the front of the vehicle are numbered from 1 to N, and the radars installed at the rear of the vehicle are numbered from N+1 to 2N. The coordinates of the head and tail endpoints of the radar wall segment on the nearest radar wall corresponding to the numbered radar are: [NRi (NRXi, NRYi), NLi (NLXi, NLYi)], (NRXi, NRYi) is the head coordinate of the line segment, (NLXi, NLYi) is the tail coordinate of the line segment; i is the radar serial number, and its value is 1, 2.....2N; The coordinates of the head and tail endpoints of the radar wall segment on the farthest radar wall corresponding to each radar are: [FRi (FRXi, FRYi), FLi (FLXi, FLYi)], where i is 1, 2, ... 2N; a3: based on the number of vehicle-mounted radars, the radar wall FOV and the radar detection distance, determine the farthest safety distance f1 and the closest safety distance n1; a4: According to the known conditions of radar wall FOV, shortest safety distance n1 and longest safety distance f1, a data relationship is constructed. When N has different values, the coordinates of each point on the radar wall segment corresponding to all the radar wall arc segments are calculated according to the mathematical relationship; a5: Calculate the radar wall arc line according to the coordinates of each point on the radar wall segment and the center angle of each radar wall arc line.

5. According to claim 4, a method for visualizing a dynamic linear radar wall for vehicle monitoring based on deep learning, characterized in that: In step a5, when N=4, the calculation method of the radar wall arc line is as follows: b1: The radars installed on the vehicle to be calculated include: radars A, B, C, D installed on the front of the vehicle and radars D', C', B', A' installed on the rear of the vehicle, and the corresponding serial numbers are 1 to 8; The following point coordinates are obtained by calculation: The endpoint coordinates of the radar wall segment corresponding to the nearest radar wall: [(NRXi,NRYi)、(NLXi,NLYi)], i is 1, 2, ... 8; The endpoint coordinates of the radar wall segment corresponding to the farthest radar wall: [(FRXi, FRYi), (FLXi, FLYi)], where i is 1, 2, ... 8; Vehicle radar coordinates: A: (a1, b1), B: (a2, b2); b2: confirm the center angle of each arc line of the radar wall; b3: According to the positional relationship between the vehicle-mounted radar and the radar wall, the following relationship is established: NRY1 = k1*NRX1 + c1; FRY1 = k2*FRX1 + c2; The values ​​of k1, k2, c1 and c2 are obtained by solving; b4: Solve for the arc length of the radar wall; Assume that point J is the midpoint of the radar wall segment [NR1 (NRX1, NRY1), NL1 (NLX1, NLY1)], then: The coordinates of point J are: ((NLX1+NRX1) / 2,(NRY1+NLY1) / 2); Assume that point O is the center of the radar wall arc [NR1, NL1], then connect point J and point O to get the straight line y3: y3=k3*x+c3; Based on the known conditions: the distances from the points on the perpendicular bisector to the points at both ends are the same, and the central angle of each arc line of the radar wall is known; According to the mathematical relationship, we can get: line segment OJ, line segment ONL1, line segment ONR1 and line segment NL1NR1; Based on the known coordinates of the endpoints of the radar wall segment, the values ​​of k3 and c3 are obtained by calculation; b5: Calculate the coordinates of point O through the straight line y3=k3*x+c3, the coordinates of point J, and the distance of line segment OJ; Then, according to the coordinates of the center O and the radius ONL1 of the circle, the coordinates and length of the radar wall arc line [NL1, NR1] are calculated.

6. The method for visualizing a dynamic linear radar wall for vehicle monitoring based on deep learning according to claim 5, characterized in that: In step b5, assuming that the central angle of each radar wall arc is 60°, the following relationship holds: Line segment ONL1 = line segment ONR1 = line segment NL1NR1; Then, based on the mathematical relationship, we can calculate: k3=-1 / ((NRY1-NLY1) / (NRX1-NLX1))=(NLX1-NRX1) / (NRY1-NLY1); c3=(NRY1+NLY1) / 2-((NLX1-NRX1)*(NLX1+NRX1)) / 2(NRY1-NLY1); y3=((NLX1-NRX1) / (NRY1-NLY1))*x+(NRY1+NLY1) / 2-((NLX1-NRX1)*(NLX1+NRX1)) / 2(NRY1-NLY1).

Images