Millimeter wave radar road target detection method based on lightweight deployment model

Abstract

The invention discloses a millimeter wave radar road target detection method based on a lightweight deployment model, which can obviously improve the accuracy and generalization ability of road target identification in a complex traffic scene. Compared with an original YOLOv5 model, a k-means + + method is used for optimizing a preset anchor frame; the Neck part is embedded into a mixed ECA and CBAM attention module, efficient channel attention (ECA) and a convolution block attention module (CBAM) are fused in a feature pyramid network (FPN + PAN), and dynamic weighting of channels and spatial information is achieved; a Backbone part is replaced by a proposed MG hybrid network structure, and a self-set Ghost-IRB is added, so that redundant feature calculation is reduced; and the loss function part adopts the combination of Alpha-IoU positioning loss and Focal Loss classification loss, so that the regression robustness of the bounding box is improved, and the problem of class imbalance is relieved. And the information reflected by the radar is better used to realize road target identification in a complex road scene and realize stable detection deployment with lower cost.

Description

Technical Field

[0001] The present invention relates to the field of computer vision and target detection technology, and in particular to a millimeter-wave radar road target detection method based on a lightweight deployment model. Background Art

[0002] Millimeter-wave radar has strong anti-interference capabilities and can still work stably under complex lighting and bad weather conditions. It has the advantages of all-weather and all-day operation, and is low-cost and easy to deploy. Therefore, it plays an important role in road target detection.

[0003] In the existing technology, the constant false alarm rate (CFAR) detection algorithm is widely used in radar target detection. This algorithm mainly relies on physical models such as radar signals and environmental clutter. In complex and changeable actual scenarios, it often faces the problem of model mismatch, which affects the detection effect. The detection method based on deep learning usually uses optical images as data input, but there are large differences in features between the processed radar images and optical images, which makes the deep learning method limit the improvement of detection performance to a certain extent. In addition, the deployment cost of the model itself is high, and the computing resource consumption is large, and the cost-effectiveness for edge road detection applications is not high. Therefore, there is an urgent need for a new road target detection method that can cooperate with radar reflection information to perform stable complex road target recognition while further reducing deployment costs and promoting cost-effective full road detection coverage. Summary of the Invention

[0004] Purpose of the invention: To solve the problems mentioned in the background technology, the present invention discloses a millimeter-wave radar road target detection method based on a lightweight deployment model. By improving the yolov5 model, using the MG (MobileNetv3-Ghost) hybrid structure, adding a self-designed Ghost-IRB, and embedding a hybrid ECA-CBAM module to obtain an improved model, while stably using radar reflection information to perform road target recognition, it further reduces the amount of calculation and improves channel redundancy, thereby reducing the cost of detection model deployment.

[0005] Technical solution:

[0006] The present invention discloses a method for detecting road targets using a millimeter-wave radar based on a lightweight deployment model, the method comprising the following steps:

[0007] S1 receives the data stream from the radar RF front end, uses Doppler FFT to obtain the target velocity and non-coherent accumulation method to obtain the RD map, and annotates the radar RD map as a data set for network training and verification;

[0008] S2 builds an improved YOLOv5 model. The input is the radar RD image scaled to 640×640 resolution. The K-means++ algorithm is used to re-cluster the data.

[0009] The Backbone part uses a MobileNetv3-Ghost hybrid structure instead of the original CSPDarknet as the backbone feature extraction network. The MobileNetv3-Ghost hybrid structure replaces the convolution kernel of the Stem layer of the original CSPDarknet with a smaller convolution kernel, and replaces the CSPBlock of the stage layer with Ghost-IRB.

[0010] The Neck part embeds a hybrid ECA-CBAM module in each feature fusion path of FPN+PAN in the original YOLOv5;

[0011] S3 uses the dataset to train the improved YOLOv5 model, converts the radar data into RD graphs, and inputs them into the trained improved YOLOv5 model to obtain target detection results.

[0012] Furthermore, the detailed steps of S1 are as follows:

[0013] Millimeter-wave radar uses frequency-modulated continuous waves to obtain reflected echo signals. The transmitted signal and the echo signal are mixed and passed through a low-pass filter to obtain an intermediate frequency signal. The distance to the reflection point is then obtained through distance FFT, and the velocity of the reflection point is obtained through Doppler FFT. Finally, the two-dimensional information is mapped to an image to obtain a radar RD map. The LabelImg tool is used to annotate the radar RD map, which is used as a dataset for network training and validation.

[0014] Furthermore, the steps of re-clustering the data using the K-means++ algorithm described in S2 are as follows:

[0015] Extract the width and height information of all annotation boxes from the training set Normalized to the same height, the complement of the intersection ratio is used as the clustering distance, from {(w i ,h i )} randomly select the first cluster center c1 and then for each sample x i Calculate the closest distance D(x i ), then by probability Select the next center and repeat until k initial centers are selected; in the assignment phase, each target box is assigned to the nearest cluster center: And in the update phase, the cluster center is recalculated as the median of the width and height of all boxes in the cluster, and finally k cluster centers {(w c1 ,h c1 )....(wck ,h ck )}, which is the optimal anchor box set that adapts to the target size, where w i , h i are the original width and height of the i-th target box, C i is the cluster center index to which the i-th sample belongs, IoU is the intersection over union ratio, c j Indicates the width and height of the j-th cluster center, box i is the width and height of the i-th annotation box.

[0016] Furthermore, the MobileNetv3-Ghost hybrid structure described in S2 replaces the original CSPDarknet as the backbone feature extraction network. By integrating lightweight convolution and feature reuse technology, it reduces the number of parameters while maintaining multi-scale feature extraction capabilities. The MG hybrid structure replaces the convolution kernel of the Stem layer of the original CSPDarknet with a smaller convolution kernel, and replaces the CSPBlock of the stage layer with Ghost-IRB. During the replacement process, all convolution layers with stride = 2 must be strictly aligned with the downsampling position of the original CSPDarknet to ensure that the feature map resolution matches the detection head input. When the number of input and output channels is the same, a jump connection is added, otherwise it is replaced with 1×1 GhostConv for channel alignment.

[0017] Furthermore, the specific structure of Ghost-IRB is as follows:

[0018] In the expansion stage, 1×1 GhostConv is used to expand the input channel from Cin to t×Cin, and 3×3 depth-separable convolution is applied. The activation function is h-swish for depth convolution. In the compression stage, 1×1 GhostConv is used to reduce the dimension to the target number of channels C. out ;Given an input F∈R c×w×h , where c is the number of channels, h is the height, and w is the width. After the n×k×k convolution kernel, the feature map F'∈R is obtained. c×h'×w' , the parameters of ghostconv are:

[0019]

[0020] The amount of calculation is:

[0021]

[0022] Among them, d×d is the convolution kernel size of the linear operation, s is the number of linear transformations, s< <c, is the number of output channels at the first transformation, and s-1 is the identity mapping.

[0023] Furthermore, the Neck part of S2 is used to fuse high- and low-level features. The feature maps of different sizes are scaled by upsampling and downsampling convolution operations and then spliced in the channel dimension. The FPN path upsamples the high-level semantic features and splices them with the bottom-level features, while the PAN path downsamples the bottom-level detail features and splices them with the high-level features. The fusion paths include bottom-up and top-down paths. A hybrid ECA-CBAM module is embedded in each feature fusion path of FPN+PAN in the original YOLOv5 model, and the input feature map passes through ECA and CBAM in turn. After the channel dimension is spliced, the dimension is reduced by 1×1 convolution to output the weighted feature map.

[0024] Furthermore, the hybrid ECA-CBAM module operates as follows:

[0025] The steps of the ECA module are as follows: global averaging is used to extract global information of each channel; cross-channel interactions are captured through one-dimensional convolution, the convolution kernel size k is dynamically adjusted, the channel weights are output, and after channel reweighting, the weights are applied to the input feature map;

[0026] The CBAM module is divided into two parts: channel attention and spatial attention:

[0027] The channel attention step is: using average pooling and maximum pooling to aggregate, generating channel weight sharing MLP through the fully connected layer and fusing it with the output weight to perform channel reweighting;

[0028] The spatial attention steps are as follows: concatenating the average and maximum pooling results along the channel dimension, capturing spatial relationships through 7×7 convolution for spatial reweighting, and ECA capturing local cross-channel interactions through lightweight one-dimensional convolution to avoid the high computational complexity of the fully connected layer in CBAM.

[0029] Beneficial effects:

[0030] The present invention improves the yolov5 model by using the MG (MobileNetv3-Ghost) hybrid structure, uses the k-means++ method to optimize the preset anchor frame, and adopts the intersection-over-completion of union rather than the Euclidean distance as the clustering distance, thereby avoiding the error caused by the excessively large aspect ratio of the target object and improving the model recognition accuracy; adds a self-designed Ghost-IRB, replaces some of the depth-separable convolutions in MobileNetV3 with the Ghost module, that is, the shallow layer retains the MobileNetV3 structure, and the deep layer uses the Ghost module. And through the cheap feature generation mechanism of the Ghost module, the number of intermediate channels is compressed to reduce the amount of calculation, effectively reducing the model deployment cost; the neck part is embedded in the hybrid ECA-CBAM module, and ECA captures local cross-channel interactions through lightweight one-dimensional convolution, avoiding the high amount of calculation of the fully connected layer in CBAM. The spatial attention of CBAM makes up for the defect of ECA's lack of spatial positioning ability, improving the robustness of the road detection model when dealing with complex targets and further reducing redundant calculations. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 Flow chart of the method of the present invention;

[0032] Figure 2 This is the network structure diagram of MobileNetV3, MG and CSPBlock of the present invention;

[0033] Figure 3 This is a structural diagram of the Ghost-IRB module of the present invention;

[0034] Figure 4 This is the structure diagram of the improved YOLOv5 model of the present invention. DETAILED DESCRIPTION

[0035] To make the purpose, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts shall fall within the scope of protection of the present invention.

[0036] like Figure 1 As shown, the present invention discloses a millimeter wave radar road target detection method based on a lightweight deployment model, and the method steps are as follows:

[0037] S1 receives the data stream from the radar RF front end, uses Doppler FFT to obtain the target velocity and non-coherent accumulation method to obtain the RD map, and annotates the radar RD map as a data set for network training and verification;

[0038] Millimeter wave radar uses frequency modulated continuous wave (FMCW), and its transmission signal is:

[0039]

[0040] Where f0 is the starting frequency, (B is bandwidth, T c is the frequency modulation period); at the same time, the reflected echo signal can be obtained as:

[0041]

[0042] The delay (R is the initial distance to the target, v is the radial velocity of the target, and c is the speed of light); the transmitted signal is then mixed with the echo signal and passed through a low-pass filter to obtain an intermediate frequency signal:

[0043]

[0044] Where α is the FM slope, τ is the time delay, f0 is the starting frequency, j is a virtual unit, ατt represents the phase term generated by the combined effect of the FM slope and the time delay, corresponding to the frequency of the intermediate frequency signal; f0τ represents the fixed phase offset generated by the starting frequency and the time delay; The quadratic phase correction term of the FM signal. The distance to the reflection point is obtained through distance FFT, and the velocity of the reflection point is obtained through Doppler FFT. Finally, the two-dimensional information is mapped into an image to obtain the RD map.

[0045] In this embodiment, the radar RD map is labeled using the LabelImg tool, and is used as a data set for network training and verification.

[0046] S2 builds an improved YOLOv5 model. The input is a radar RD image scaled to a resolution of 640×640. The K-means++ algorithm is used to re-cluster the data. The detailed steps are as follows:

[0047] Extract the width and height information of all annotation boxes from the training set Normalize to the same height. Use the complement of the intersection-over-union ratio as the cluster distance to avoid the problem that the Euclidean distance is sensitive to the size ratio:

[0048] d(box,centroid)=1-IoU(box,centroid)

[0049] k-means++ is initialized, first from {(w i ,h i )} randomly select the first cluster center c1 and then for each sample x i Calculate the closest distance D(x i);then press the probability Select the next center, and repeat until k initial centers are selected.

[0050] In the assignment phase, each target box is assigned to the nearest cluster center:

[0051]

[0052] Among them, w i , h i are the original width (width) and height (height) of the i-th target box, Ci is the cluster center index to which the i-th sample belongs, IoU is the intersection over union ratio, c j Indicates the width and height of the j-th cluster center, box i is the width and height of the i-th labeled box. In the update phase, the cluster center is recalculated as the median of the width and height of all boxes in the cluster.

[0053] Finally, we get k cluster centers {(w c1 ,h c1 )....(w ck ,h ck )}, which is the optimal anchor box set that adapts to the target size.

[0054] The specific preset anchor box sizes obtained in this embodiment are (12, 18), (22, 25), (30, 16), (35, 28), (45, 33), (60, 45), (75, 50), (90, 65), and (120, 80).

[0055] like Figure 2 As shown, the Backbone component uses a MobileNetv3-Ghost hybrid structure instead of the original CSPDarknet as the backbone feature extraction network. This MobileNetv3-Ghost hybrid replaces the convolution kernels in the original CSPDarknet's Stem layer with smaller convolution kernels and replaces the CSPBlock in the stage layer with Ghost-IRB. The Ghost module's inexpensive feature generation mechanism compresses the number of intermediate channels to reduce computational overhead. During the replacement process, all stride-2 convolution layers must be strictly aligned with the downsampling positions of the original CSPDarknet to ensure that the feature map resolution matches the detection head input. Secondly, skip connections are added only when the number of input and output channels is the same. Otherwise, they are replaced with a 1×1 GhostConv for channel alignment.

[0056] like Figure 3 As shown, the Ghost-IRB structure is:

[0057] X out=SEBlock(GhostConv(DepthwiseConv(Expand(X in ))))+X skip

[0058] The principle is as follows:

[0059] The expansion stage uses 1×1GhostConv to convert the input channels from C in Extended to t×C in ,

[0060] X expand =GhostConv 1×1 (X in ,C out =tC in ,groups=g)

[0061] in C in is the number of channels of the input feature map, t is the expansion factor, g is the number of groups, X in is the input feature map, X expand is the expanded feature map.

[0062] Then apply 3×3 depth-wise separable convolution, and use h-swish as activation function for depth-wise convolution:

[0063] X dw =h-swish(DepthwiseConv 3×3 (X expand ))

[0064] In the compression stage, 1×1 GhostConv is used to reduce the dimension to the target number of channels.

[0065] C out :X proj =GhostConv 1×1 (X dw ,C out ,groups=g')

[0066] in DepthwiseConv 3×3 represents a 3×3 depthwise separable convolution, N dw Denotes the number of parameters of depthwise convolution, C out is the target number of output channels, X proj is the compressed feature map.

[0067] The Ghost convolution formula is as follows: Given an input F∈R c×w×h(where c is the number of channels, h is the height, and w is the width), and the feature map F'∈R is obtained through the n×k×k convolution kernel. c×h'×w' For ordinary convolution, the number of parameters is: n×c×k×k, and the amount of calculation is h'×w'×n×c×k×k; while the number of parameters of ghostconv is The amount of calculation is Where d×d is the convolution kernel size of the linear operation, s is the number of linear transformations, and s< <c。 is the number of output channels in the first transformation, and s-1 is because the identity mapping does not need to be calculated, but it is also counted as part of the second transformation, so ghostconv can save calculations.

[0068] In this embodiment, Ghost-IRB introduces an inverted residual structure and an h-swish activation function, which improves the feature expression capability by 20%.

[0069] The Neck layer is used to fuse high- and low-level features. It scales feature maps of different sizes through upsampling and downsampling convolution operations and then concatenates them along the channel dimension. The FPN path upsamples high-level semantic features (small size) and concatenates them with low-level features (large size). The PAN path downsamples low-level detail features and concatenates them with high-level features.

[0070] The fusion path includes two paths: bottom-up and top-down. In the original YOLO model, a hybrid ECA-CBAM module is embedded in each feature fusion path of FPN+PAN. The input feature map passes through ECA (adaptive channel weight allocation) and CBAM (spatial attention weighting) in sequence; after the channel dimension is spliced, it is reduced by 1×1 convolution and outputs a weighted feature map. The output process is as follows: out =CBAM(ECA(X in ));

[0071] The steps of the ECA module are:

[0072] Global Averaging (GAP): Extract global information of each channel:

[0073] Among them, X c (i, j) is the position of the cth channel of the input feature map at position (i, j), and H and W are the height and width of the feature map respectively.

[0074] Adaptive one-dimensional convolution: Capture cross-channel interactions through one-dimensional convolution, and dynamically adjust the convolution kernel size k. (γ=2, b=1 are hyperparameters) Output channel weights: Where σ is the sigmoid function, Conv1D kIt is a one-dimensional convolution with a kernel size of k, acting on the channel dimension;

[0075] Channel reweighting, applying weights to the input feature maps.

[0076] The CBAM module is divided into two parts: channel attention and spatial attention. The channel attention steps are:

[0077] Dual pooling aggregation uses average pooling and maximum pooling at the same time: Z avg =GAP(X ECA ), Z MAX =GMP(X ECA );

[0078] Generate channel weight sharing MLP and weight fusion through the fully connected layer:

[0079] M c =σ(W1(δ(W0Z avg ))+W1(δ(W0Z max )))

[0080] Among them, δ is the ReLU activation function, σ is the sigmoid activation function, and channel reweighting is performed.

[0081] The spatial attention step is to concatenate the average and maximum pooling results along the channel dimension:

[0082] Z spatial =Concat(AvgPool(X CA ),MaxPool(X CA ))

[0083] Among them, Z spatial Represents the intermediate feature map after channel dimension pooling and splicing operations, Concat is the spatial dimension splicing operation, AvgPool is the average pooling of the output feature map in the channel dimension, and MaxPool is the maximum pooling of the output feature map in the channel dimension.

[0084] Capturing spatial relationships through 7×7 convolution: M s =σ(Conv 7×7 (Z spatial )) performs spatial reweighting. ECA captures local cross-channel interactions through lightweight one-dimensional convolution, avoiding the high computational overhead of the fully connected layers in CBAM. CBAM's spatial attention compensates for ECA's lack of spatial localization capabilities.

[0085] S3 uses the dataset to train the improved YOLOv5 model, the structure is as follows Figure 4As shown in Figure 1, the radar data is converted into an RD map and input into the trained improved YOLOv5 model to obtain the target detection results.

[0086] During the training process, this embodiment introduces three loss functions including classification loss, positioning loss, and confidence loss.

[0087] The classification loss uses Focal Loss, which is used to deal with the problem of category imbalance. Its expression is:

[0088] FL(p t )=-α t (1-p t ) γ log(p t )

[0089] Where Pt is the model's predicted probability of the target class, αt is the balancing factor used to adjust the influence between positive and negative samples, and γ is the focus factor used to adjust the weight of difficult and easy samples.

[0090] The confidence loss uses binary cross entropy (BCE) to determine whether the anchor box contains a valid target and suppress the overfitting risk caused by a large number of negative samples. For N sample data, the loss expression is:

[0091]

[0092] Where y i represents the i-th true label, p i represents the probability that the i-th model predicts class 1.

[0093] The positioning loss function uses Alpha-IoU, which is used to accurately adjust the center coordinates (x, y) and size (w, h) of the anchor box to solve the problems of scale sensitivity and inaccurate positioning of rotated targets. Its expression is:

[0094]

[0095] Among them, ρ is the Euclidean distance between the center of the predicted box and the real box, d is the diagonal length of the minimum closed area containing the two boxes, is the adaptive weight coefficient.

[0096] The total loss function is L total =λ1L cls +λ2L conf +λ3L loc . Among them, λ1, λ2, and λ3 are the weight coefficients of classification loss, confidence loss, and positioning loss respectively.

[0097] As shown in Table 1, the present invention improves Yolov5. As the number of model layers increases, the number of parameters is reduced from 7235289 to 5561755, the number of model parameters is reduced by 20%, and the computational complexity is reduced from 16.5 to 11.6, achieving a lightweight effect. The addition of the attention mechanism enables the model to maintain its lightweight characteristics while only slightly reducing the detection accuracy, achieving the best balance between accuracy and speed.

[0098] Table 1

[0099] Methods parameterS GFLOPs mAP@0.5(%) layers Yolov5s 7235289 16.5 0.982 240 The present invention 5561755 11.6 0.978 297

[0100] The above description of the embodiments enables one skilled in the art to implement or use the present invention. Various modifications to the embodiments will be readily apparent to those skilled in the art. The general principles of the present invention may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention should not be limited to the embodiments shown herein, but should encompass the widest range consistent with the principles and novel features disclosed herein.

Claims

1. A millimeter-wave radar road target detection method based on a lightweight deployment model, characterized in that: The method comprises the following steps: S1 receives the data stream from the radar RF front end, uses Doppler FFT to obtain the target velocity and non-coherent accumulation method to obtain the RD map, and annotates the radar RD map as a data set for network training and verification; S2 builds an improved YOLOv5 model. The input is the radar RD image scaled to 640×640 resolution. The K-means++ algorithm is used to re-cluster the data. The Backbone part uses a MobileNetv3-Ghost hybrid structure instead of the original CSPDarknet as the backbone feature extraction network. The MobileNetv3-Ghost hybrid structure replaces the convolution kernel of the Stem layer of the original CSPDarknet with a smaller convolution kernel, and replaces the CSPBlock of the stage layer with Ghost-IRB. The Neck part embeds a hybrid ECA-CBAM module in each feature fusion path of FPN+PAN in the original YOLOv5; S3 uses the dataset to train the improved YOLOv5 model, converts the radar data into RD graphs, and inputs them into the trained improved YOLOv5 model to obtain target detection results.

2. The method for detecting road targets using a millimeter-wave radar based on a lightweight deployment model according to claim 1, wherein: The detailed steps of S1 are as follows: Millimeter-wave radar uses frequency-modulated continuous waves to obtain reflected echo signals. The transmitted signal and the echo signal are mixed and passed through a low-pass filter to obtain an intermediate frequency signal. The distance to the reflection point is then obtained through distance FFT, and the velocity of the reflection point is obtained through Doppler FFT. Finally, the two-dimensional information is mapped to an image to obtain a radar RD map. The LabelImg tool is used to annotate the radar RD map, which is used as a dataset for network training and validation.

3. The millimeter wave radar road target detection method based on a lightweight deployment model according to claim 1 is characterized in that: The steps for re-clustering the data using the K-means++ algorithm described in S2 are as follows: Extract the width and height information of all annotation boxes from the training set Normalized to the same height, the complement of the intersection ratio is used as the clustering distance, from {(w i ,h i )} randomly select the first cluster center c1 and then for each sample x i Calculate the closest distance D(x i ), then by probability Select the next center and repeat until k initial centers are selected; in the assignment phase, each target box is assigned to the nearest cluster center: And in the update phase, the cluster center is recalculated as the median of the width and height of all boxes in the cluster, and finally k cluster centers {(w c1 ,h c1 )....(w ck ,h ck )}, which is the optimal anchor box set that adapts to the target size, where w i , h i are the original width and height of the i-th target box, C i is the cluster center index to which the i-th sample belongs, IoU is the intersection over union ratio, c j Indicates the width and height of the j-th cluster center, box i is the width and height of the i-th annotation box.

4. The method for detecting road targets using a millimeter-wave radar based on a lightweight deployment model according to claim 1, wherein: The MobileNetv3-Ghost hybrid structure described in S2 replaces the original CSPDarknet as the backbone feature extraction network. By integrating lightweight convolution and feature reuse technology, it reduces the number of parameters while maintaining multi-scale feature extraction capabilities. The MG hybrid structure replaces the convolution kernel of the Stem layer of the original CSPDarknet with a smaller convolution kernel, and replaces the CSPBlock of the stage layer with Ghost-IRB. During the replacement process, all convolution layers with stride = 2 must be strictly aligned with the downsampling position of the original CSPDarknet to ensure that the feature map resolution matches the detection head input. When the number of input and output channels is the same, a jump connection is added, otherwise it is replaced with 1×1 GhostConv for channel alignment.

5. The method for detecting road targets using a millimeter-wave radar based on a lightweight deployment model according to claim 4, wherein: The specific structure of Ghost-IRB is as follows: In the expansion stage, 1×1 GhostConv is used to expand the input channel from Cin to t×Cin, and 3×3 depth-separable convolution is applied. The activation function is h-swish for depth convolution. In the compression stage, 1×1 GhostConv is used to reduce the dimension to the target number of channels C. out ;Given an input F∈R c×w×h , where c is the number of channels, h is the height, and w is the width. After the n×k×k convolution kernel, the feature map F'∈R is obtained. c×h'×w' , the parameters of ghostconv are: The amount of calculation is: Among them, d×d is the convolution kernel size of the linear operation, s is the number of linear transformations, s< <c, is the number of output channels at the first transformation, and s-1 is the identity mapping.

6. The method for detecting road targets using a millimeter-wave radar based on a lightweight deployment model according to claim 1, wherein: The Neck part described in S2 is used to fuse high- and low-level features. The feature maps of different sizes are scaled by upsampling and downsampling convolution operations and then spliced in the channel dimension. The FPN path upsamples the high-level semantic features and splices them with the bottom-level features, while the PAN path downsamples the bottom-level detail features and splices them with the high-level features. The fusion paths include bottom-up and top-down paths. In the original YOLOv5 model, a hybrid ECA-CBAM module is embedded in each feature fusion path of FPN+PAN, and the input feature map passes through ECA and CBAM in turn. After the channel dimension is spliced, the dimension is reduced by 1×1 convolution to output the weighted feature map.

7. The method for detecting road targets using a millimeter-wave radar based on a lightweight deployment model according to claim 6, wherein: The hybrid ECA-CBAM module operates as follows: The steps of the ECA module are as follows: global averaging is used to extract global information of each channel; cross-channel interactions are captured through one-dimensional convolution, the convolution kernel size k is dynamically adjusted, the channel weights are output, and after channel reweighting, the weights are applied to the input feature map; The CBAM module is divided into two parts: channel attention and spatial attention: The channel attention step is: using average pooling and maximum pooling to aggregate, generating channel weight sharing MLP through the fully connected layer and fusing it with the output weight to perform channel reweighting; The spatial attention steps are as follows: concatenating the average and maximum pooling results along the channel dimension, capturing spatial relationships through 7×7 convolution for spatial reweighting, and ECA capturing local cross-channel interactions through lightweight one-dimensional convolution to avoid the high computational complexity of the fully connected layer in CBAM.

Images