A radar road target detection method based on a YOLO model

Abstract

The application belongs to the technical field of radar target detection, and particularly relates to a radar road target detection method based on a YOLO model. The method first needs to collect echo signals of road targets through a millimeter wave radar, and convert the echo signals into radar RD graphs to input into a detection model to complete detection of the road targets. The detection model is an improved YOLOv5, and the improvement mode comprises: using a K-means method to optimize preset anchor frames; using a SIoU to optimize a loss function; using a CA mechanism to strengthen the extraction capability of the model on features; and using a FasterNet as a backbone network to realize lightweight of the model. The application can realize high-performance detection of road targets through radar RD graphs.

Description

Technical Field

[0001] This invention belongs to the field of radar target detection technology, specifically relating to a method for radar road target detection based on the YOLO model. Background Technology

[0002] Optical cameras are the most widely used in target detection, providing rich image information such as color and contour. However, optical cameras are susceptible to lighting conditions and environmental factors such as rain, snow, and fog, resulting in poor detection performance. LiDAR uses lasers to acquire target information, offering high resolution and measurement accuracy, and enabling 3D imaging. However, its anti-interference capabilities are poor, and its high cost makes large-scale deployment difficult. Ultrasonic radar offers high accuracy for measuring close-range targets, but ultrasonic waves attenuate rapidly during propagation and are easily absorbed by other interfering objects, thus limiting its detection range. Millimeter-wave radar boasts strong anti-interference capabilities, can operate stably in poor lighting and weather conditions, and offers all-weather, all-day operation. Its lower cost and ease of deployment make it an important sensor for road target detection.

[0003] Constant false alarm rate (CFAR) detection is a commonly used algorithm in radar target detection. However, it is mainly based on physical models such as radar signals and environmental clutter. In complex and ever-changing real-world scenarios, model mismatch often occurs, leading to a decrease in detection capability. Deep learning-based detection methods typically use optical images as data sources, but processed radar images differ significantly from visual images, thus limiting the detection performance of these methods to some extent. Summary of the Invention

[0004] This invention proposes a millimeter-wave radar target detection method based on the YOLO model, mainly targeting road scenarios. The millimeter-wave radar data is processed and converted into a range Doppler (RD) map, and an improved YOLOv5 model is used to improve detection performance.

[0005] The technical solution of this invention is as follows:

[0006] A radar road target detection method based on the YOLO model includes the following steps:

[0007] S1. Use millimeter-wave radar to collect road target echo signals and obtain RD maps through range dimension fast Fourier transform (FFT), Doppler dimension FFT and noncoherent accumulation methods, which serve as data sources for deep learning models;

[0008] S2. Label the radar RD map to serve as the dataset for network training and validation.

[0009] S3. Build the improved YOLOv5 model, consisting of four parts: Input, Backbone, Neck, and Head. The Input is a radar RD image scaled to 640×640, and the dataset is re-clustered using the K-means algorithm with Intersection over Union (IoU) as the sample distance to obtain a new set of bounding boxes as the model's preset anchor boxes. The Backbone uses FasterNet as the backbone feature extraction network. The data input from the Input sequentially passes through Embedding, FasterNet Block, Merging, FasterNet Block×2, Merging, FasterNet Block×8, Merging, FasterNet Block×2, and Spatial Pyramid Pooling-Fast (SPPF). The FasterNet Block consists of partial convolutions, dotted convolutions, batch normalization (BN), and Gaussian Error Linear Units (GANs). The YOLOv5 model consists of a GELU activation function, with Embedding and Merging composed of downsampling convolutions and Batch Normalization (BN). SPPF consists of four parallel max pooling operations, concatenated at the end. The data from the first FasterNet Block×2, FasterNet Block×8, and SPPF outputs in the Backbone part are fused into the Neck part. The Neck part is used to fuse high and low layer features. It scales feature maps of different sizes through upsampling and downsampling convolution operations and then concatenates them along the channel dimension. The fusion path includes two paths: bottom-up and top-down. Furthermore, coordinate attention is added before the three outputs of the Neck in the original YOLOv5 model. The CA module enhances the model's feature extraction capabilities. The bottom-up path includes CBS, UpSample, concat, CSP, CBS, UpSample, concat, and CSP, while the top-down path includes CA, CBS, CSP, CA, CBS, concat, CSP, and CA. For the bottom-up path, after sorting the modules in the bottom-up order, the data output from SPPF in the Backbone part is output to the first CBS, the data output from FasterNet Block×8 is input to the first concat, and the data output from the first FasterNet Block×2 is input to the second concat.For the top-down path, after sorting the modules in top-down order, the data output from the first CBS in the bottom-up path is input into the second concat in the top-down path, the data output from the second CBS in the top-down path is input into the first concat in the top-down path, and the data output from the second CSP in the top-down path is input into the first CA in the top-down path. The three CAs in the top-down path output three streams of data to the Head section for predicting small, medium, and large targets respectively. The CBS structure consists of a convolutional layer, a BN layer, and a SiLU activation function connected in sequence. The CSP structure consists of multiple CBSs and residual connections. The CA is used to perform average pooling on the input feature matrix in both width and height dimensions to capture spatial and channel information simultaneously. The pooling results are then concatenated and dimensionality is reduced using convolution. Finally, after processing by BN and HSwish activation functions, the feature matrix is ​​segmented according to its height and width, and then normalized using the Sigmoid function after being increased to the initial dimension by convolution.

[0010] S4. Train the improved YOLOv5 network using the dataset. The loss function includes confidence loss, classification loss, and localization loss, with a total loss of [missing information].

[0011]

[0012] in, , and These are the weighting coefficients for confidence loss, classification loss, and localization loss, respectively. SIoU was used for localization loss to increase consideration of the regression angle of the predicted bounding box.

[0013] S5. Convert the measured radar data into an RD map and feed it into the trained improved YOLOv5 model to obtain the target detection results.

[0014] The beneficial effects of this invention are as follows: (1) It uses a data-driven deep learning method, which has more stable detection performance than the traditional model-driven constant false alarm rate (CFAR) detection method and does not have the problem of model mismatch. (2) To address the sparsity of radar RD maps, a lightweight backbone network FasterNet is used, which speeds up the detection process while reducing the number of parameters. (3) The overall detection performance of the YOLOv5 model is improved by redesigning the preset anchor boxes, adding an attention mechanism, and optimizing the loss function. Attached Figure Description

[0015] Figure 1 This is a flowchart of a radar road target detection method based on the YOLO model.

[0016] Figure 2The structure diagram for improving the YOLOv5 model.

[0017] Figure 3 This is a detection result diagram of the radar RD map. Detailed Implementation

[0018] The specific technical solution of the present invention will be described below with reference to the accompanying drawings.

[0019] This invention relates to a radar road target detection method based on the YOLO model. The overall process is as follows: Figure 1 As shown, it specifically includes the following steps.

[0020] S1. Use millimeter-wave radar to collect road target echo signals and obtain RD maps through range-dimensional FFT, Doppler-dimensional FFT and non-coherent accumulation methods;

[0021] S2. Label the radar RD map to serve as the dataset for network training and validation.

[0022] S3. Build the improved YOLOv5 model. The YOLOv5 model consists of four parts: Input, Backbone, Neck, and Head. The Input is a radar RD image scaled to 640×640. The dataset is re-clustered using the K-means algorithm with IoU as the sample distance to obtain a new set of bounding box sizes as the model's preset anchor boxes. The specific preset anchor box sizes are (15,21), (24,21), (23,30), (33,22), (47,22), (36,34), (25,64), (59,35), and (63,87). The Backbone uses FasterNet as the main feature extraction network, which mainly consists of the FasterNet Block, Embedding, and Merging modules. The FasterNet Block consists of partial convolutions, pointwise convolutions, BN, and the GELU activation function. Embedding and Merging consist of downsampling convolutions and BN. The SPPF after FasterNet consists of four parallel max pooling operations, concatenated at the end using concat. The Neck section fuses high- and low-level features, scaling feature maps of different sizes through upsampling and downsampling convolutions before concatenating them along the channel dimension. The fusion path includes both bottom-up and top-down paths. A CA module is added before the three outputs of the Neck to enhance the model's feature extraction capability. CA performs average pooling on the input feature matrix in both width and height dimensions to capture both spatial and channel information. The pooling results are then concatenated and dimensionality-reduced using convolution. After processing with BN and HSwish activation functions, the feature matrix is ​​split according to its height and width, then increased to the initial dimension through convolution and normalized using the Sigmoid function. Finally, the feature maps from the two branches are used as weights to weight the initial feature matrix to obtain the result. The Head section predicts target information, and the three outputs of the Neck section are used to predict small, medium, and large targets, respectively. The improved YOLOv5 model network structure is as follows: Figure 2 As shown;

[0023] S4. Train the improved YOLOv5 network using the dataset. The training parameters are configured as follows: batch size of 16, training epochs of 300, using stochastic gradient descent, initial learning rate of 0.01, and momentum parameter of 0.937. The loss mechanism includes confidence loss, classification loss, and localization loss, with a total loss of [missing value].

[0024]

[0025] The localization loss uses SIoU to incorporate consideration of the regression angle of the predicted bounding box.

[0026] S5. Convert the measured radar data into an RD map and feed it into the trained improved YOLOv5 model to obtain the target detection results, such as... Figure 3 As shown, the proposed solution can achieve high-performance detection of road targets and can be applied to the field of radar target detection technology.

Claims

1. A radar road target detection method based on the YOLO model, characterized in that, Includes the following steps: S1. Use millimeter-wave radar to collect road target echo signals and obtain RD maps through range dimension FFT, Doppler dimension FFT and noncoherent accumulation, which serve as data sources for deep learning models. S2. Label the radar RD map to obtain the dataset used for network training; S3. Construct the improved YOLOv5 model, consisting of four parts: Input, Backbone, Neck, and Head. The Input is a radar RD image scaled to 640×640, and the dataset is re-clustered using the K-means algorithm with intersection-over-union ratio (IoU) as the sample distance to obtain a new set of bounding boxes as the model's preset anchor boxes. The Backbone uses FasterNet as the main feature extraction network. The input data sequentially passes through Embedding, FasterNetBlock, Merging, FasterNet Block×2, Merging, FasterNet Block×8, Merging, FasterNetBlock×2, and SPPF. The FasterNet Block consists of partial convolutions, dotted convolutions, Batch Normalization (BN), and Gaussian error linear unit activation functions. Embedding and Merging consist of downsampling convolutions and BN. SPPF consists of four parallel max pooling operations, concatenated at the end using concat. The first FasterNet Block×2 in the Backbone... The data output from Block×8 and SPPF are fed into the Neck section for fusion. The Neck section is used to fuse high- and low-level features. It scales feature maps of different sizes using upsampling and downsampling convolution operations and then concatenates them along the channel dimension. The fusion path includes two paths: bottom-up and top-down. The bottom-up path includes CBS, UpSample, concat, CSP, CBS, UpSample, concat, CSP. The top-down path includes CA, CBS, CSP, CA, CBS, concat, CSP, CA. For the bottom-up path, after sorting the modules in bottom-up order, the SPPF output from the Backbone section is output to the first CBS, the FasterNet Block×8 output is input to the first concat, and the first FasterNet... The data output from Block×2 is input into the second concat. For the top-down path, after sorting the modules in top-down order, the data output from the first CBS in the bottom-up path is input into the second concat in the top-down path, the data output from the second CBS in the top-down path is input into the first concat in the top-down path, and the data output from the second CSP in the top-down path is input into the first CA in the top-down path. The three CAs in the top-down path output three data paths to the Head section for predicting small, medium, and large targets. The structure of CBS consists of a convolutional layer, a BN layer, and a SiLU activation function connected in sequence.The CSP structure consists of multiple CBSs and residual connections; CA is used to perform average pooling on the input feature matrix in both width and height dimensions to capture spatial and channel information simultaneously. The pooling results are then concatenated and dimensionality is reduced using convolution. Finally, after processing by BN and HSwish activation functions, the feature matrix is ​​segmented according to its height and width. After being increased to the initial dimension by convolution, it is normalized using the Sigmoid function. S4. Train the improved YOLOv5 network using the dataset. The loss mechanism includes confidence loss, classification loss, and localization loss, with a total loss of [missing value]. ， in, , and These are the weighting coefficients for confidence loss, classification loss, and localization loss, respectively. SIoU is used for localization loss to increase consideration of the regression angle of the predicted box. S5. Convert the measured radar data into an RD map and feed it into the trained improved YOLOv5 model to obtain the target detection results.

Images