A method for detecting cloaked human targets based on multi-angle millimeter-wave images

Abstract

This disclosure provides a method for cloaked human target detection based on multi-angle millimeter-wave images. An existing dataset is divided into training, validation, and test sets and preprocessed to construct a training image sequence set. Millimeter-wave images of the human body are acquired from multiple angles, continuously in time, with each capture around the body forming an image sequence. All images in the image sequence set are labeled. The image data from the image sequence set is input into a YOLOv5-CGRU network structure to iteratively update network parameters and complete the training of the network model. The preprocessed test set images are used to construct a millimeter-wave image sequence set. The millimeter-wave image sequence test set is input into the detection network of the trained network model, and the model outputs the detection results. The YOLO target detection algorithm is used for feature extraction and prediction of millimeter-wave images. Inserting CGRU into the YOLO target detection algorithm fuses feature information from multiple angles, improving detection accuracy and speed.

Description

Technical Field

[0001] This disclosure relates to the field of millimeter-wave image target detection technology, and in particular to a method for detecting cloaked human targets based on multi-angle millimeter-wave images. Background Technology

[0002] Millimeter-wave radar is harmless to humans and has strong penetrating power, enabling it to quickly and accurately detect various non-metallic, solid, liquid, and packaged hazardous gases, providing a new option for improving security inspection efficiency. Therefore, researching an automatic detection method based on millimeter-wave images is of great significance for improving the security of public places.

[0003] Currently, millimeter-wave cloaking object detection schemes mainly target single millimeter-wave images: Two-stage algorithms, such as R-CNN or its improved versions, are used for cloaking object detection in millimeter-wave images. The two-stage algorithm divides the detection problem into two stages: first generating candidate regions, then classifying them. This results in high accuracy but slow speed, unsuitable for real-time detection scenarios. One-stage algorithms, proposed by CHEN et al., use YOLOv3 as the main structure and add an attention mechanism for millimeter-wave image target detection. The one-stage algorithm uses a convolutional neural network to directly predict the category and location of different targets. This algorithm does not require generating pre-selected boxes to determine the object category probability and location coordinates, and the final result can be obtained directly in a single detection, thus its detection speed is faster than other detection networks, but its detection accuracy is poor.

[0004] Both methods detect concealed objects on a single millimeter-wave image. However, in real-world applications, millimeter-wave security inspection equipment can perform multi-angle imaging and has a temporal characteristic, meaning that a single concealed object has multiple millimeter-wave images from different angles. However, single-image-based target detection cannot utilize the feature information of the same concealed object from multiple angles, resulting in low detection accuracy and a high false alarm rate.

[0005] In summary, existing covert target detection methods based on millimeter-wave images are affected by the aforementioned detection methods, resulting in low detection accuracy, slow speed, and difficulty in meeting the needs of automation. Summary of the Invention

[0006] The purpose of this disclosure is to provide a method for detecting concealed human targets based on multi-angle millimeter-wave images, in order to solve the problems of low target detection accuracy and slow speed in the prior art.

[0007] The embodiments of this disclosure adopt the following technical solution: a method for detecting cloaked human targets based on multi-angle millimeter-wave images, comprising the following steps:

[0008] Step S1: Divide the existing dataset into training set, validation set and test set, and preprocess the images of training set, validation set and test set. The preprocessing process includes image noise reduction and non-linear gray level mapping to obtain the preprocessed training set, validation set and test set.

[0009] Step S2: Construct a training image sequence set and label all images in the image sequence set; specifically including:

[0010] Step S21: The millimeter wave acquires an image every time it passes through the human body by α degrees, for a total of m acquisitions, passing through β = mα degrees. Each acquisition is continuous in time, forming a set of image sequences; a single set of image sequences is X. i ={k α ,k 2α ,…k tα ,…,k mα}(1≤t≤m), the length of a single image sequence is m; construct the training set images into an image sequence set T={X1,X2,…X i}(i≤N);

[0011] Step S22: For the image sequence set T = {X1, X2, ... X...} i Label all images (i≤N);

[0012] Step S3: Set the image sequence set T = {X1, X2, ... X...} i Image data (i≤N) is input into the CGRU network structure of the YOLO V5-CGRU object detection algorithm and iteratively updated. The network weights are updated through backpropagation to obtain the first network parameters, thus completing the training of the network model.

[0013] Step S4: Construct a millimeter-wave image sequence set from the preprocessed test set images according to step S21; input the millimeter-wave image sequence set into the detection network of the trained network model, and output the detection results through the model.

[0014] In some embodiments, the image sequence set T = {X1, X2, ... X} i All images (i≤N) are labeled, specifically including:

[0015] Step S221: Single-group image sequence labeling. Input the single-group image sequence X into the labeling software sequentially. i ;

[0016] Step S222: View the single image sequence X in sequence. isingle image x t =k tα (1≤t≤m);

[0017] Step S223: If there are hidden objects in the image, proceed to step S224; otherwise, proceed to step S222.

[0018] Step S224: Determine the location and type of the hidden object;

[0019] Step S225: If Proceed to step S226; otherwise, proceed to step S224.

[0020] Step S226: Label the results in a single image sequence. Zhang Tu Shang;

[0021] Step S227: Generate a file from the marking results. The file should contain at least the location coordinates and size information of the hidden object.

[0022] Step S228: For the image sequence set T = {X1, X2, ... X...} i (i≤N) All images are labeled.

[0023] In some embodiments, training the network model specifically includes:

[0024] Step S31: Feature Extraction: The Backbone feature extraction structure of the YOLO target detection algorithm is used to extract features from the millimeter-wave sequence image data and generate corresponding feature information for further processing by fusing context information.

[0025] Step S32: Contextual information fusion: Add a CGRU structure after the feature extraction structure in the YOLO object detection algorithm;

[0026] By adjusting the GRU network, the dot product operation between state transitions is transformed into a convolution; CGRU processes the current input feature x t The hidden layer state h after the input feature transformation at the previous time step t-1 To merge;

[0027] Step S33: Predict the fused feature information to form anchor boxes, and filter the anchor boxes to determine the final predicted boxes;

[0028] Step S34: Set the initial learning rate;

[0029] Step S35: Modify the maximum number of training rounds and hide the target label content;

[0030] Step S36: Set the validation rounds for the validation set;

[0031] Step S37: Start training the network. After repeating the training for a preset number of rounds, a trained human concealment detection network is obtained.

[0032] In some embodiments, the feature extraction specifically includes:

[0033] Step S311: Set the initial weights of YOLO V5 hyperparameters, including the depth_multiple and width_multiple parameters;

[0034] Step S312: Set the number of network detection categories (nc);

[0035] Step S313: Sequentially process the single image sequence set X i Each image x t =k tα (1≤t≤m) is input into the Backbone structure.

[0036] In some embodiments, the Backbone structure is configured to at least:

[0037] Backbone consists of a CBS layer, a CSP1_X layer, and an SPPF layer.

[0038] The CBS structure consists of convolutional layers, BN normalization layers, and SiLU activation function layers.

[0039] The CSP1_X layer includes at least the CSP1_1 layer, the CSP1_2 layer, and the CSP1_3 layer, wherein the CSP1_1 layer, the CSP1_2 layer, and the CSP1_3 layer include a corresponding number of CSP structures;

[0040] The SPPF layer consists of a CBS structure, a pooling layer Maxpool, and a splicing structure.

[0041] In some embodiments, the CSP structure is divided into two parts. The first part is a separate CBS structure that performs convolution operations. The second part introduces a residual structure on the basis of the CBS structure and concatenates it with the feature map after the convolution operation of the first part.

[0042] In some embodiments, the CGRU applies the current input feature x t The hidden layer state h after the input feature transformation at the previous time step t-1 The integration process specifically includes:

[0043] Step S321: Initial weight parameter setting, input into the update gate weight parameter W z Hidden layer to update gate weight parameter U z The weight parameter W is input to the reset gate. rThe weight parameter U from the hidden layer to the reset gate r W h U is the weight parameter input to the hidden layer. h The weight parameters W are the weights from one hidden layer to another. o Set the initial time t = 1 and the initial state h. t-1 When t=1, h0=0;

[0044] Step S322: Input a single set of feature information sequence X sequentially. i ;

[0045] Step S323: Input a single set of feature information sequence X sequentially. i single image x t =k tα (1≤t≤m);

[0046] Step S324: via h t-1 With x t Calculate the gate control state r of the reset door t :

[0047] Step S325: By resetting the door r t h t-1 With x t Calculate the candidate state at the current time.

[0048] Step S326: via h t-1 With x t Calculate and update the gate's gate state z t :

[0049] Step S327: Calculate 1-z t ;

[0050] Step S328: Calculate the network update status:

[0051] Step S329: If t < m, calculate t = t + 1 and proceed to step S323; otherwise, proceed to step S3210.

[0052] Step S3210: via h t The final output feature y is calculated. i :

[0053] Step S3211: If i < N, calculate i = i + 1 and proceed to step S322; otherwise, proceed to step S3212.

[0054] Step S3212: Finally, N sets of fused feature information are obtained, represented as Y = {y1, y2, ..., y...} i}, where i represents the output of the i-th set of feature information sequences (i≤N).

[0055] In some embodiments, the step of predicting the fused feature information to form anchor boxes, and filtering the anchor boxes to determine the final predicted boxes, specifically includes:

[0056] Step S331: Input the N sets of feature information fused in step S32 into the prediction module;

[0057] Step S332: Generate a certain number of anchor boxes using the anchor mechanism;

[0058] Step S333: Calculate the similarity of the aspect ratio between each anchor box and the ground truth box:

[0059] Step S334: Calculate the IOU between each anchor box and the ground plane box. IOU is the overlap.

[0060] Step S335: Set the threshold for distinguishing positive and negative samples to a; when IOU > a, the anchor is judged as a positive sample, and when IOU < a, it is judged as a negative sample;

[0061] Step S336: Calculate the weighting coefficient α;

[0062] Step S337: Calculate the loss function Loss.

[0063] In some embodiments, calculating the loss function Loss specifically includes: calculating the CIOU between each anchor and the ground truth bounding box, where CIOU is the overlap adjusted for aspect ratio; and calculating the loss function Loss.

[0064] Loss CIOU =1-CIOU.

[0065] In some embodiments, forming a file from the tagging results specifically includes: the file being formed in at least one of the following forms: txt file, binary file, Word document, or Excel spreadsheet.

[0066] The beneficial effects of the embodiments disclosed herein are as follows:

[0067] The YOLO object detection algorithm's backbone structure is used to extract features from millimeter-wave images, and a prediction structure is used for prediction. A CGRU (CNN-Gated Recurrent Unit) is inserted into the YOLO object detection algorithm to fuse contextual information in the CGRU network structure, fully reflecting the temporal characteristics of millimeter-wave images. By fusing feature information from multiple angles, the accuracy of existing millimeter-wave image concealment detection is further improved, while also reducing detection time and increasing detection speed. Attached Figure Description

[0068] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0069] Figure 1 This is the overall flowchart of this disclosure;

[0070] Figure 2 This is a flowchart of the single-group sequence image labeling process disclosed in this publication;

[0071] Figure 3 This is a diagram showing the relationship between the feature extraction, context information, and prediction modules in this disclosure;

[0072] Figure 4 This is a schematic diagram of the CGRU network structure disclosed in this publication;

[0073] Figure 5 This is a schematic diagram showing the network structure and feature map size and channel number variations disclosed in this publication;

[0074] Figure 6 This is a flowchart of the public integration method;

[0075] Figure 7 This is a diagram showing the results of concealment detection in millimeter-wave images disclosed in this publication. Detailed Implementation

[0076] Various embodiments and features of this disclosure are described herein with reference to the accompanying drawings.

[0077] It should be understood that various modifications can be made to the embodiments described herein. Therefore, the above description should not be considered as limiting, but merely as an example of embodiments. Other modifications within the scope and spirit of this disclosure will be apparent to those skilled in the art.

[0078] The accompanying drawings, which are included in and form part of this specification, illustrate embodiments of the present disclosure and, together with the general description of the disclosure given above and the detailed description of the embodiments given below, serve to explain the principles of the disclosure.

[0079] These and other features of this disclosure will become apparent from the following description of preferred forms of embodiments given as non-limiting examples, with reference to the accompanying drawings.

[0080] It should also be understood that although this disclosure has been described with reference to some specific examples, many other equivalent forms of this disclosure can be definitively implemented by those skilled in the art, which have the features of the claims and are therefore within the scope of protection defined herein.

[0081] The above and other aspects, features and advantages of this disclosure will become more apparent when taken in conjunction with the accompanying drawings and in view of the following detailed description.

[0082] Specific embodiments of this disclosure are described thereafter with reference to the accompanying drawings; however, it should be understood that the claimed embodiments are merely examples of this disclosure, which may be implemented in various ways. Well-known and / or repeated functions and structures are not described in detail to avoid unnecessary or redundant details that could obscure this disclosure. Therefore, the specific structural and functional details claimed herein are not intended to be limiting, but merely to serve as the basis and representative basis for the claims to teach those skilled in the art to use this disclosure in a variety of substantially any suitable detailed structures.

[0083] This specification may use the phrases “in one embodiment,” “in another embodiment,” “in yet another embodiment,” or “in still another embodiment,” all of which may refer to one or more of the same or different embodiments according to this disclosure.

[0084] To address the issues of slow speed and low accuracy in existing concealed target detection methods, this disclosure provides a method for detecting concealed human targets based on multi-angle millimeter-wave images. The overall flowchart of this invention is shown below. Figure 1 As shown, the method includes the following steps:

[0085] Step S1: Divide the dataset and perform data preprocessing: Divide the existing dataset into training set, validation set and test set, and preprocess the images of training set, validation set and test set. The preprocessing process includes image noise reduction and non-linear gray level mapping to obtain the preprocessed training set, validation set and test set.

[0086] Step S2: Construct a training image sequence set and label all images in the image sequence set; specifically including:

[0087] Step S21: The millimeter wave acquires an image every time it passes through the human body by α degrees, for a total of m acquisitions, passing through β = mα degrees. Each acquisition is continuous in time, forming a set of image sequences; a single set of image sequences is X. i ={k α ,k 2α ,…k tα ,…,k mα} (1≤t≤m), the length of a single image sequence is m; construct the entire training set of images into an image sequence set T={X1,X2,…X i} (i≤N), that is, a total of i sets of image sequences.

[0088] Step S22: For the image sequence set T = {X1, X2, ... X...} i All images (i≤N) are manually labeled. The labeling process for a single sequence of images is as follows: Figure 2 As shown.

[0089] Step S3: Model training, using the image sequence set T = {X1, X2, ... X} i Image data (i≤N) is input into the YOLOV5-CGRU network structure and iteratively updated. The network weights are updated through backpropagation to obtain the first network parameters, which are the optimal network parameters, thus completing the training of the network model.

[0090] Step S4: Hideout detection: Construct a millimeter-wave image sequence set from the preprocessed test set images according to step S21; input the millimeter-wave image sequence set into the detection network of the trained network model, and output the detection results through the model.

[0091] This paper describes a method for feature extraction and prediction of millimeter-wave images using the YOLO object detection algorithm. A CGRU (CNN-Gated Recurrent Unit) is inserted into the YOLO algorithm to fuse contextual information within the CGRU network structure, fully reflecting the temporal characteristics of millimeter-wave images. By fusing feature information from multiple angles, the accuracy of concealed object detection in millimeter-wave images is further improved, while detection time is reduced, meeting the needs of practical applications. The CGRU network allows predicted information to be directly labeled on the millimeter-wave image, avoiding the difficulty of processing and judging large amounts of one-dimensional vector information. The network structure is simple, highly practical, and has broad application prospects with the increasing prevalence of millimeter-wave security inspection equipment.

[0092] In some embodiments, for the image sequence set T = {X1, X2, ... X... i All images (i≤N) are labeled, specifically including:

[0093] Step S221: Single-group image sequence labeling. Input the single-group image sequence X into the labeling software sequentially. i Labeling software includes Label me, Label img, etc. This software is used to open the corresponding image for manual review.

[0094] Step S222: View the single image sequence X in sequence. i single image x t =k tα (1≤t≤m).

[0095] Step S223: If there are hidden objects in the image, proceed to step S224; otherwise, proceed to step S222. Whether there are hidden objects in the image is determined by manual inspection and marking.

[0096] Step S224: Determine the location and type of the hidden object.

[0097] Step S225: If Proceed to step S226; otherwise, proceed to step S224.

[0098] Step S226: Label the results in a single image sequence. The picture is above.

[0099] Step S227: Create a file from the marking results. The file must be in at least one of the following formats: txt file, binary file, Word document, or Excel spreadsheet. The file must contain at least the location coordinates and size of the hidden object.

[0100] Step S228: For the image sequence set T = {X1, X2, ... X...} i (i≤N) All images are labeled.

[0101] Training the network model specifically includes:

[0102] Step S31: Feature Extraction: The YOLO V5 feature extraction structure Backbone is used to extract features from the millimeter-wave sequence image data and generate corresponding feature information for further processing by fusing context information.

[0103] Step S32: Contextual information fusion: such as Figure 3 As shown, CGRU, or gated neural unit, is added after the feature extraction structure in YOLO V5.

[0104] By adjusting the GRU network, the dot product operation between state transitions is transformed into a convolution, thus overcoming the problem of loss of two-dimensional spatial information during transmission. This allows the two-dimensional information output by the feature extraction module to be directly input into the GRU. Simultaneously, the GRU processes the current input feature x... t The hidden layer state h after the input feature transformation at the previous time step t-1 To achieve fusion, the CGRU network structure is as follows: Figure 4 As shown.

[0105] Step S33: Predict the fused feature information to form anchor boxes, and filter the anchor boxes to determine the final predicted boxes. The prediction module is used for prediction because the feature extraction step only outputs one set of feature information, which is then fused in step S32; therefore, the prediction module only has one set of prediction data.

[0106] Step S34: Set the initial learning rate.

[0107] Step S35: Modify the maximum number of training rounds and hide the target label content.

[0108] Step S36: Set the validation set validation rounds. Set the number of rounds as needed.

[0109] Step S37: Start training the network. After repeating the training for a preset number of rounds, a trained human concealment detection network is obtained.

[0110] In some embodiments, feature extraction specifically includes:

[0111] Step S311: Set the initial weights of YOLO V5 hyperparameters, including the depth_multiple and width_multiple parameters.

[0112] Step S312: Set the number of network detection categories (nc).

[0113] Step S313: Sequentially process the single image sequence set X i Each image x t =k tα (1≤t≤m) Input into the Backbone structure. The Backbone structure must be configured at least as follows:

[0114] The backbone consists of CBS layers, CSP1_X layers, and SPPF layers. The backbone comprises 10 layers in total: 5 CBS layers, 4 CSP1_X layers, and 1 SPPF layer. The specific network structure, feature map size, and channel number variations are as follows... Figure 5 As shown.

[0115] The CBS structure consists of convolutional layers, BN normalization layers, and SiLU activation function layers. The five CBS convolutional layers have sizes of 6*6, 3*3, 3*3, 3*3, 3*3, and 3*3, respectively. The SiLU function is calculated as follows:

[0116] f(x)=x·sigmoid(x) (1)

[0117] The CSP1_X structure mainly consists of two layers of CSP1_1, one layer of CSP1_2, and one layer of CSP1_3. CSP1_1, CSP1_2, and CSP1_3 contain 4, 8, and 12 CSP structures respectively. The CSP structure is divided into two parts: one part uses a separate CBS structure for convolutional operations, and the other part introduces a Resx residual structure based on the CBS structure and concatenates it with the feature map after the convolutional operation in the previous part. Introducing the residual structure reduces computation and solves the gradient vanishing problem. The SPPF structure consists of two CBS layers, three pooling layers (Maxpool), and a concatenation structure. Multiple pooling operations are used to extract more high-level semantic features.

[0118] CGRU for the current input feature x t The hidden layer state h after the input feature transformation at the previous time step t-1 The integration process is illustrated in the flowchart below. Figure 6 As shown, it specifically includes:

[0119] Step S321: Initial weight parameter setting, input into the update gate weight parameter W z Hidden layer to update gate weight parameter U z The weight parameter W is input to the reset gate. r The weight parameter U from the hidden layer to the reset gate r W h U is the weight parameter input to the hidden layer. h The weight parameters W are the weights from one hidden layer to another. o Set the initial time t = 1 and the initial state h. t-1 When t=1, h0=0.

[0120] Step S322: Input a single set of feature information sequence X sequentially. i .

[0121] Step S323: Input a single set of feature information sequence X sequentially. i single image x t =k tα (1≤t≤m);

[0122] Step S324: via h t-1 With x tCalculate the gate control state r of the reset door t :

[0123] r t =sigmoid(W r *x t +U r *h t-1 +b r (2)

[0124] Among them W z Update the gate z to the parameters to be trained. t The decision was made to h t-1 With the reset hidden layer state How much information will be retained as the final state?

[0125] Step S325: By resetting the door r t h t-1 With x t Calculate the candidate state at the current time.

[0126]

[0127] Among them W h These are the parameters to be trained.

[0128] Step S326: via h t-1 With x t Calculate and update the gate's gate state z t :

[0129] z t =sigmoid(W z *x t +U z *h t-1 +b z (4)

[0130] Among them W z Update the gate z to the parameters to be trained. t The decision was made to h t-1 With the reset hidden layer state How much information will be retained as the final state?

[0131] Step S327: Calculate 1-z t ;

[0132] Step S328: Calculate the network update status:

[0133]

[0134] Step S329: If t < m, calculate t = t + 1 and proceed to step S323; otherwise, proceed to step S3210.

[0135] Step S3210: via h t The final output feature y is calculated. i :

[0136] y t =sigmoid(W o *h t +b o (6)

[0137] Step S3211: If i < N, calculate i = i + 1 and proceed to step S322; otherwise, proceed to step S3212.

[0138] Step S3212: Finally, N sets of fused feature information are obtained, represented as Y = {y1, y2, ..., y...} i}, where i represents the output of the i-th set of feature information sequences (i≤N).

[0139] In some embodiments, the step of predicting the fused feature information to form anchor boxes, and filtering the anchor boxes to determine the final predicted boxes, specifically includes:

[0140] Step S331: Input the N sets of feature information fused in step S32 into the prediction module.

[0141] Step S332: Generate a certain number of anchor boxes using the anchor mechanism.

[0142] Step S333: Calculate the similarity of each anchor box with the aspect ratio of the ground truth (GT) box:

[0143]

[0144] Where w and h represent the length and width of the anchor box, w gt h gt This indicates the length and width of the GT box, where GT represents the actual label box.

[0145] Step S334: Calculate the IOU between each anchor box and the ground plane box. IOU is the overlap.

[0146]

[0147] C represents the area of ​​the anchor box. gt This represents the area of ​​the GT frame.

[0148] Step S335: Set the threshold 'a' to be used to distinguish between positive and negative samples; when IOU > a, the anchor is judged as a positive sample, and when IOU < a, it is judged as a negative sample.

[0149] Step S336: Calculate the weighting coefficient α;

[0150]

[0151] Step S337: Calculate the CIOU of each anchor and the ground truth bounding box. CIOU is the overlap with aspect ratio correction.

[0152]

[0153] Where ρ 2 (b,b gt ) represents the Euclidean distance between the anchor box and the ground truth box, and b represents the center coordinate parameter of the anchor box. gt The center coordinate parameter of the ground truth (GT) box is represented by , and c is the diagonal distance between the anchor and the smallest bounding rectangle of the GT box.

[0154] Calculate the loss function:

[0155] Loss CIOU =1-CIOU (11)

[0156] The improved algorithm of this invention is as follows regarding the detection results of hidden objects in millimeter-wave images: Figure 7 As shown, it is possible to Figure 7 The markers of the concealed objects can be seen in the human body image on the right, indicating good detection performance.

[0157] The foregoing has provided a detailed description of several embodiments of this disclosure. However, this disclosure is not limited to these specific embodiments. Those skilled in the art can make various variations and modifications based on the concept of this disclosure, and all such variations and modifications should fall within the scope of protection claimed by this disclosure.

Claims

1. A method for detecting cloaked human targets based on multi-angle millimeter-wave images, characterized in that, Includes the following steps: Step S1: Divide the existing dataset into training set, validation set and test set, and preprocess the images of training set, validation set and test set. The preprocessing process includes image noise reduction and non-linear gray level mapping to obtain the preprocessed training set, validation set and test set. Step S2: Construct a training image sequence set and label all images in the image sequence set; specifically including: Step S21: Millimeter waves pass through the human body A single image was acquired, and a total of [number] images were acquired. Next, after Each acquisition is continuous in time, forming a set of image sequences; a single set of image sequences is... The length of a single image sequence is ; Construct an image sequence set from the training set images ; Step S22: For the image sequence set All images are labeled; Step S3: Set the image sequence Image data is input into the YOLO V5-CGRU network structure for iterative updates. The network weights are updated through backpropagation to obtain the first network parameters, thus completing the training of the network model. Step S4: Construct a millimeter-wave image sequence set from the preprocessed test set images according to step S21; input the millimeter-wave image sequence set into the detection network of the trained network model, and output the detection results through the model; Among them, the image sequence set All images are tagged, including: Step S221: Single-group image sequence labeling, input single-group image sequences sequentially into the labeling software. ; Step S222: Review the single image sequence in sequence. single image ( ); Step S223: If there are hidden objects in the image, proceed to step S224; otherwise, proceed to step S222. Step S224: Determine the location and type of the hidden object; Step S225: If If yes, proceed to step S226; otherwise, proceed to step S224. Step S226: Label the results in a single image sequence. Zhang Tu Shang; Step S227: Generate a file from the marking results. The file should contain at least the location coordinates and size information of the hidden object. Step S228: For the image sequence set All images are labeled.

2. The detection method according to claim 1, characterized in that, The training of the network model specifically includes: Step S31: Feature Extraction: The Backbone feature extraction structure of the YOLO target detection algorithm is used to extract features from the millimeter-wave sequence image data and generate corresponding feature information for further processing by fusing context information. Step S32: Contextual information fusion: Add a CGRU structure after the feature extraction structure in the YOLO object detection algorithm; By adjusting the GRU network, the dot product operation between state transitions is transformed into a convolution; CGRU processes the current input features Hidden layer state after input feature transformation at the previous time step To merge; Step S33: Predict the fused feature information to form anchor boxes, and filter the anchor boxes to determine the final predicted boxes; Step S34: Set the initial learning rate; Step S35: Modify the maximum number of training rounds and hide the target label content; Step S36: Set the validation rounds for the validation set; Step S37: Start training the network. After repeating the training for a preset number of rounds, a trained human concealment detection network is obtained.

3. The detection method according to claim 2, characterized in that, The feature extraction specifically includes: Step S311: Set the initial weights of YOLO V5 hyperparameters, including the depth_multiple and width_multiple parameters; Step S312: Set the number of network detection categories (nc); Step S313: Sequentially process the single image sequence sets Each image Input into the Backbone structure.

4. The detection method according to claim 3, characterized in that, The Backbone structure is configured at least as follows: Backbone consists of a CBS layer, a CSP1_X layer, and an SPPF layer. The CBS structure consists of convolutional layers, BN normalization layers, and SiLU activation function layers. The CSP1_X layer includes at least the CSP1_1 layer, the CSP1_2 layer, and the CSP1_3 layer, wherein the CSP1_1 layer, the CSP1_2 layer, and the CSP1_3 layer include a corresponding number of CSP structures; The SPPF layer consists of a CBS structure, a pooling layer Maxpool, and a splicing structure.

5. The detection method according to claim 4, characterized in that, The CSP structure is divided into two parts. The first part is a separate CBS structure that performs convolution operations. The second part introduces a residual structure on the basis of the CBS structure and concatenates it with the feature map after the convolution operation of the first part.

6. The detection method according to claim 2, characterized in that, The CGRU uses the current input features Hidden layer state after input feature transformation at the previous time step The integration process specifically includes: Step S321: Initial weight parameter setting, input into update gate weight parameters Hidden layer to update gate weight parameters The weight parameters input to the reset gate Weight parameters from hidden layer to reset door , These are the weight parameters input to the hidden layer. The output weight parameters are the weight parameters from one hidden layer to another. Set initial time initial state ,when time ; Step S322: Input single-group feature information sequences sequentially. ; Step S323: Input single-group feature information sequences sequentially. single image ( ); Step S324: Through and Calculate and reset the gate control status ; Step S325: Reset the door , and Calculate the candidate state at the current time. ; Step S326: Through and Calculate and update the gate's gate control state ; Step S327: Calculation ; Step S328: Calculate the network update state; Step S329: If ,calculate If yes, proceed to step S323; otherwise, proceed to step S3210. Step S3210: Through The final output features are calculated. ; Step S3211: If ,calculate If yes, proceed to step S322; otherwise, proceed to step S3212. Step S3212: Finally, N sets of fused feature information are obtained, represented as , Indicates the first Output of group feature information sequence .

7. The detection method according to claim 2, characterized in that, The process of predicting the fused feature information to form anchor boxes, and then filtering the anchor boxes to determine the final predicted boxes, specifically includes: Step S331: Input the N sets of feature information fused in step S32 into the prediction module; Step S332: Generate a certain number of anchor boxes using the anchor mechanism; Step S333: Calculate the similarity of the aspect ratio between each anchor box and the ground truth box; Step S334: Calculate the relationship between each anchor box and the ground truth box. IOU, IOU Overlap degree; Step S335: Set the threshold used to distinguish between positive and negative samples. a ;when When, the anchor is judged as a positive sample, when It was determined to be a negative sample; Step S336: Calculate the weighting coefficient α; Step S337: Calculate the loss function Loss .

8. The detection method according to claim 7, characterized in that, The calculation loss function Loss Specifically, this includes: calculating the relationship between each anchor and the ground truth bounding box. CIOU, CIOU To increase the overlap of aspect ratio corrections: calculate the loss function. Loss : 。 9. The detection method according to claim 1, characterized in that, The step of forming a file from the marking results specifically includes: the file being formed in at least one of the following formats: txt file, binary file, Word document, or Excel spreadsheet.

Images