Personnel accurate positioning and sensing method based on Doppler radar and infrared camera

Abstract

The invention discloses a personnel accurate positioning and sensing method based on Doppler radars and infrared cameras. The method comprises the following steps: respectively obtaining an infrared data set containing coal mine underground operating personnel and a plurality of radar RD data sets; performing data preprocessing on the infrared and radar RD data sets; constructing a YOLOv8 network model by a target recognition algorithm based on radar-infrared fusion; a YOLO-T1 network model is constructed on the basis of a YOLOv8 network; carrying out the training of the YOLO-T1 network model; high-precision positioning and sensing can be carried out on underground coal mine operating personnel; real-time processing is carried out on infrared data including coal mine underground operating personnel and data of a plurality of radars RD, and then data preprocessing is carried out; and inputting the preprocessed infrared data and radar RD data as input data into the YOLO-T1 network model for processing, and outputting a positioning result of the operating personnel by the YOLO-T1 network model. According to the method, high-precision positioning and sensing of the underground coal mine operating personnel can be realized.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent sensing technology, and in particular relates to a personnel precise positioning and sensing method based on Doppler radar and infrared camera. Background Art

[0002] In the field of coal mine safety production, how to achieve accurate positioning of underground coal mine workers has always been a very challenging problem. Due to the extremely complex underground environment of coal mines, there are usually a large amount of floating dust gases such as coal dust and water mist, and the geological structure is changeable. In daily production and when accidents occur, the location of underground coal mine workers is often difficult to determine in a timely and accurate manner, which brings great difficulties to related work. Although the existing underground communication system and personnel positioning equipment provide a certain basis for the positioning of workers, in such a harsh and changeable environment, the performance of existing systems and equipment is often severely restricted. In actual application, problems such as signal interference and equipment failure occur frequently, which seriously affects the accuracy and reliability of positioning. More importantly, the traditional manual positioning method is not only time-consuming and labor-intensive, but also accompanied by potential safety hazards that cannot be ignored.

[0003] In view of this, it is particularly urgent to develop a technical solution that can achieve high-precision positioning and perception of underground coal mine workers. With the advancement of science and technology, the integrated application of cutting-edge technologies such as the Internet of Things, big data analysis, and artificial intelligence has provided a new solution to this problem. By integrating these advanced technologies into the underground precise positioning system, it is expected to greatly improve the positioning accuracy and perception efficiency of underground coal mine workers, thereby ensuring the life safety of underground coal mine workers to the greatest extent. Summary of the invention

[0004] In response to the problems existing in the above-mentioned prior art, the present invention provides a method for precise positioning and perception of personnel based on Doppler radar and infrared camera. Through the deep fusion of pulse Doppler radar and infrared camera, the method can effectively expand the target detection dimension and achieve high-precision positioning and perception of personnel working underground in coal mines.

[0005] In order to achieve the above object, the present invention provides a method for accurately locating and sensing personnel based on Doppler radar and infrared camera, comprising the following steps:

[0006] Step 1: Use infrared cameras and pulse Doppler radars to scan the direction of underground coal mine workers. Through a massive scanning process, infrared data sets and multiple radar RD data sets containing underground coal mine workers are obtained respectively;

[0007] Step 2: Preprocess the infrared data set according to S21, and preprocess the radar RD data set according to S22;

[0008] S21: Preprocessing of infrared dataset:

[0009] A1: Use the regularized median filtering method to denoise the salt and pepper noise in the infrared data;

[0010] A2: Super-resolution reconstruction based on bicubic interpolation algorithm is performed on the infrared data after removing salt and pepper noise;

[0011] Based on the surrounding n 2 pixels, by calculating n 2 The weight of each pixel is accumulated to obtain the pixel value I(x',y') of the super-resolution point;

[0012]

[0013] In the formula, I(i,j) represents the pixel value of point (i,j), and n represents the size of the window function, that is, according to the n near the super-resolution point 2 The super-resolution point value is calculated for each pixel, W(i,j) is the weight coefficient, W(i,j) = w(d i )*w(d j ), where d i =p(i,j).x-x',d j =p(i,j).y-y', Among them, the value range of a is [-1,0];

[0014] A3: Perform homomorphic filtering on the reconstructed infrared data;

[0015] A31: Formula (2) is used to represent an image as the product of high frequency and low frequency components;

[0016] f(x,y)=f i (x,y)f r (x,y) (2);

[0017] In the formula, f i Represents the light intensity component that varies with spatial position, which is characterized by slow changes and concentration in the low-frequency part of the image, f r Represents the reflection component reflected to the human eye, which contains various information about the scene and is rich in high-frequency components;

[0018] A32: Take a logarithmic transformation on formula (2) to obtain formula (3), and change the multiplicative component into an additive component;

[0019] lnf(x,y)=lnf i (x,y)+lnf r (x,y) (3);

[0020] A33: Use formula (4) to perform Fourier transform on the image to obtain the corresponding frequency domain;

[0021] DFT[lnf(x,y)]=DFT[lnf i (x,y)]+DFT[lnf r (x,y)] (4);

[0022] A34: Filter the image using a homomorphic filter according to formula (5);

[0023]

[0024] Where H(u,v) is the filter, Among them, γ H >1,γ L <1, control the range of filter amplitude;

[0025] A35: Perform inverse Fourier transform and exponential transform on the image in sequence to restore the image and obtain an image after homomorphic filtering;

[0026] S22: Preprocess the radar RD data set:

[0027] B1: Convert the radar data into the same image format as the infrared data according to formula (6), that is, scale the radar RD data to an integer between [0,255];

[0028]

[0029] Where V max is the global maximum value, p(i,j) and p'(i,j) are the points before and after scaling respectively;

[0030] B2: Binarization of radar RD data to remove environmental noise and improve the quality of radar RD images;

[0031] Step 3: Build the YOLOv8 network model based on the radar-infrared fusion target recognition algorithm;

[0032] S31: Build the YOLO-T1 network model based on the YOLOv8 network;

[0033] The YOLO-T1 network consists of four modules, namely the CBS module, C2f module, SPPF module and BiFPN module. In the backbone network of the YOLO-T1 network, CBS and C2f are used three times and an SPPF module is used at the end, and a total of 4 features are output at different layers. In the branch network of the YOLO-T1 network, the CBS module is used to downsample the feature map, and a C2f module is connected to further strengthen the features. Between CBS and C2f, the BiFPN module is introduced to weightedly fuse features to help the network capture features of different scales and output a total of one feature. Finally, the YOLO-T1 network outputs three detection heads to output the features and position of the target.

[0034] S32: Train the YOLO-T1 network model;

[0035] C1: For the preprocessed infrared dataset and radar RD dataset, the mosaic enhancement method is used to augment the data to obtain the augmented dataset;

[0036] C2: Use the augmented data set to train the YOLO-T1 network model, and use formula (7) as the loss function; iteratively train the YOLO-T1 network model parameters by minimizing the loss function to obtain the optimal YOLO-T1 network model after training;

[0037]

[0038] Where d is the distance between the two center points of the predicted box and the real box, c is the diagonal length of the minimum circumscribed rectangle of the predicted box and the real box; α is the correction factor, in, (ω G ,h G ) and (ω G ,h G ) are the width and height of the target box and the prediction box respectively;

[0039] Step 4: Achieve high-precision positioning and perception of underground coal mine workers;

[0040] S41: Use an infrared camera and a pulse Doppler radar to scan the direction of underground coal mine workers, and obtain infrared data including underground coal mine workers and multiple radar RD data respectively;

[0041] S42: performing data preprocessing on the infrared data and the radar RD data respectively;

[0042] S43: The pre-processed infrared data and radar RD data are input as input data into the YOLO-T1 network model for processing, and the YOLO-T1 network model outputs the positioning result of the operator.

[0043] As a preferred embodiment, in step 2 S21, the pixel points that need to be denoised are determined by traversing all pixel points of the image, and then the pixel points that need to be denoised are set to the median value of the grayscale values ​​of all pixel points in the neighborhood window within the filtering radius of the point, so that the pixel points that need to be denoised are close to the true value to eliminate isolated noise points.

[0044] As a preferred embodiment, in step 3 S31;

[0045] The CBS module consists of a two-dimensional convolutional layer, a batch normalization layer and an activation layer, which is used to downsample the feature map;

[0046] The C2f module is used to pass the input data through the first CBS module, and then divide the output of the first CBS module into two parts, one of which is directly passed to the output, and the other part is passed through multiple Bottleneck depths to extract features. Finally, the results of the two parts are spliced ​​in the channel dimension and passed through the second CBS module to obtain the output of the C2f module, where Bottleneck is composed of two CBS modules;

[0047] The SPPF module is used to pass the input data through the CBS module once, and then through the maximum pooling three times. Each maximum pooling obtains a feature output. After splicing the three outputs, it passes through the second CBS to obtain the SPPF module output;

[0048] The BiFPN module is used to extract features from the backbone network and branch network of the YOLO-T1 network, generate a feature pyramid, and introduce bidirectional connections between adjacent levels of the feature pyramid. At the same time, it integrates information from different levels of the pyramid in two directions, and adopts a weighted feature fusion mechanism to combine features of different levels. The training weights are learned and optimized during the training process.

[0049] As a preferred embodiment, in step 2 A2, the value of n is 4.

[0050] As a preferred embodiment, in A2 of step 2, the value of a is -0.5.

[0051] In the present invention, for the infrared data set collected by the infrared camera, a large amount of salt and pepper noise therein greatly affects the quality of the image, and then causes serious interference to the accurate identification of the target. Since the identification target accounts for a relatively small proportion, global filtering will cause the identification target to be blurred. For this reason, the regularized median filter is innovatively used to denoise the salt and pepper noise in the infrared data, so that the noise points can be effectively removed and the clarity of the identification target can be ensured. Since the target of the coal mine underground operator in the infrared data is extremely small, it brings a great challenge to the recognition process. Even if the clarity of the operator's target is ensured, it is still difficult to accurately identify the position of the coal mine underground operator. For this reason, in the preprocessing part, the infrared data is reconstructed by super-resolution based on the bicubic interpolation algorithm, which can significantly improve the recognition accuracy of the operator. At the same time, since the contrast between the coal mine underground operator and the background is not high, by performing homomorphic filtering on the infrared data set, the image quality can be improved by compressing the brightness range and enhancing the contrast. In this way, the contrast between the coal mine underground operator and the background can be effectively enhanced, and the recognition accuracy of the coal mine underground operator target can be greatly improved. Scaling the radar RD data to [0,255] is conducive to the effective fusion of radar RD data and infrared data. In this way, the fused data can be used to significantly improve the recognition accuracy and efficiency. By binarizing the radar RD data, a large amount of clutter and environmental noise in the data can be effectively removed, and at the same time, the quality of the radar RD map can be effectively improved. Building the YOLO-T1 network model based on the YOLOv8 network can more effectively fuse radar RD data and infrared data, and can efficiently and accurately capture the features required for recognition at a low computational cost. In the training process of the YOLO-T1 network model, CIoU Loss is used. CIoU Loss can more accurately reflect the difference between the predicted box and the real box by introducing the center distance, aspect ratio and scale factor of the bounding box, thereby better optimizing the detection performance during the training process. At the same time, CIoU Loss only needs to set one hyperparameter, which simplifies the setting process of hyperparameters during training. It can converge faster during training, reduce training time and improve training efficiency. Therefore, by adopting CIoU Loss, the center point distance, aspect ratio and scale factor can be comprehensively considered, which can provide better performance and faster convergence speed in target detection tasks.

[0052] This method can effectively expand the target detection dimension through the deep fusion of pulse Doppler radar and infrared camera, and can achieve high-precision positioning and perception of underground coal mine workers. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 is a flow chart of the present invention;

[0054] Figure 2 It is a schematic diagram of the structure of the YOLO-T1 network model in the present invention. DETAILED DESCRIPTION

[0055] The present invention will be further described below in conjunction with the accompanying drawings.

[0056] like Figure 1 and Figure 2 As shown, the present invention provides a method for accurate positioning and sensing of personnel based on Doppler radar and infrared camera, comprising the following steps:

[0057] Step 1: Use infrared cameras and pulse Doppler radars to scan the direction of underground coal mine workers. Through a massive scanning process, infrared data sets and multiple radar RD data sets containing underground coal mine workers are obtained respectively;

[0058] Step 2: Preprocess the infrared data set according to S21, and preprocess the radar RD data set according to S22;

[0059] S21: Preprocessing of infrared dataset:

[0060] A1: The large amount of salt and pepper noise in infrared data greatly affects the quality of the image, which in turn seriously interferes with the accurate recognition of the target. Since the target is small, global filtering will cause blurring of the target. Based on this, the regularized median filtering method is used to denoise the salt and pepper noise in the infrared data;

[0061] A2: The targets of coal mine workers in infrared data are extremely small, which brings great challenges to the recognition task. In order to accurately identify coal mine workers, super-resolution reconstruction based on the bicubic interpolation algorithm is performed on the infrared data after removing salt and pepper noise;

[0062] Based on the surrounding n 2 pixels, by calculating n 2 The weight of each pixel is accumulated to obtain the pixel value I(x',y') of the super-resolution point;

[0063]

[0064] In the formula, I(i,j) represents the pixel value of point (i,j), and n represents the size of the window function, that is, according to the n near the super-resolution point 2 The super-resolution point value is calculated for each pixel, W(i,j) is the weight coefficient, W(i,j) = w(d i )*w(d j ), where d i =p(i,j).x-x',d j =p(i,j).y-y', Among them, the value range of a is [-1,0];

[0065] A3: Since the targets of coal mine workers in the infrared image data are small and the contrast between coal mine workers and the background is not high, in order to enhance the contrast between coal mine workers and the background and improve the recognition accuracy of coal mine workers, the reconstructed infrared data is processed by homomorphic filtering; homomorphic filtering is an image processing method that combines frequency filtering and spatial grayscale transformation. It uses the illumination model of the image as the basis for frequency domain processing, and uses the compressed brightness range and enhanced contrast to improve the image quality. The specific process is as follows:

[0066] A31: Formula (2) is used to represent an image as the product of high frequency and low frequency components;

[0067] f(x,y)=f i (x,y)f r (x,y) (2);

[0068] In the formula, f i Represents the light intensity component that varies with spatial position, which is characterized by slow changes and concentration in the low-frequency part of the image, f r Represents the reflection component reflected to the human eye, which contains various information about the scene and is rich in high-frequency components;

[0069] A32: Take a logarithmic transformation on formula (2) to obtain formula (3), and change the multiplicative component into an additive component;

[0070] lnf(x,y)=lnf i (x,y)+lnf r (x,y) (3);

[0071] A33: Use formula (4) to perform Fourier transform on the image to obtain the corresponding frequency domain;

[0072] DFT[lnf(x,y)]=DFT[lnf i (x,y)]+DFT[lnf r (x,y)] (4);

[0073] A34: Filter the image using a homomorphic filter according to formula (5);

[0074]

[0075] Where H(u,v) is the filter, Among them, γ H >1,γ L <1, control the range of filter amplitude;

[0076] A35: Perform inverse Fourier transform and exponential transform on the image in sequence to restore the image and obtain an image after homomorphic filtering;

[0077] S22: Preprocess the radar RD data set:

[0078] B1: In order to effectively fuse infrared data and radar RD data, the radar RD data is converted into the same image format as the infrared data. According to formula (6), the radar data is converted into the same image format as the infrared data, that is, the radar RD data is scaled to an integer between [0, 255];

[0079]

[0080] Where V max is the global maximum value, p(i,j) and p'(i,j) are the points before and after scaling respectively;

[0081] B2: The quality of the converted radar RD image is not high. At the same time, there are a lot of clutter and environmental noise. In order to remove the environmental noise, the radar RD data is binarized to remove the environmental noise and improve the quality of the radar RD image.

[0082] Step 3: Build the YOLOv8 network model based on the radar-infrared fusion target recognition algorithm;

[0083] S31: Build the YOLO-T1 network model based on the YOLOv8 network;

[0084] Aiming at the problem of target recognition of underground coal mine workers, this application innovatively designed the YOLO-T1 network based on the YOLOv8 network. Figure 2 As shown;

[0085] The YOLO-T1 network consists of four modules, namely the CBS module, C2f module, SPPF module and BiFPN module; in the backbone network of the YOLO-T1 network, CBS and C2f are used three times and an SPPF module is used at the end, and a total of 4 features are output at different layers; generally speaking, CBS uses a 3×3 convolution with a step size of 2 to downsample the feature map, and then connects a C2f module to further enhance the features. In the branch network of the YOLO-T1 network, the CBS module is used to downsample the feature map, and a C2f module is connected to further enhance the features. Between CBS and C2f, the BiFPN module is introduced to weightedly fuse features to help the network capture features of different scales and output a total of one feature; finally, the YOLO-T1 network outputs three detection heads, outputting the features and position of the target;

[0086] S32: Use YOLOv5 and YOLOv8 data enhancement and training loss strategies to train the YOLO-T1 network model;

[0087] C1: The YOLO-T1 network uses the data enhancement strategy of YOLOv5. For the preprocessed infrared dataset and radar RD dataset, mosaic enhancement is used to augment the data, that is, random cropping and scaling are used to splice and combine the input images to obtain the augmented dataset.

[0088] C2: The YOLO-T1 network uses the training loss strategy of YOLOv8. Like YOLOv8, it does not use the candidate box strategy, so the problem to be solved by the training loss is the multi-scale allocation of positive and negative sample matching. The training targets of YOLO-T1 include two categories, targetbbox and targetcls, that is, accurate recognition of position and classification. Correspondingly, the loss also includes two parts: position loss and category loss.

[0089] CIoU Loss is an improved version of YOLOv4 based on the original IOU loss function. It takes into account the complete intersection between target boxes and introduces a correction factor to more accurately measure the similarity between target boxes.

[0090] The YOLO-T1 network model is trained using the augmented data set, and formula (7) is used as the loss function. The YOLO-T1 network model parameters are iteratively trained by minimizing the loss function to obtain the optimal YOLO-T1 network model after training.

[0091]

[0092] Where d is the distance between the two center points of the predicted box and the real box, c is the diagonal length of the minimum circumscribed rectangle of the predicted box and the real box; α is the correction factor, It is used to accurately adjust the CIoU loss function and further consider the shape and direction of the target box, where (ω G ,h G ) and (ω G ,h G ) are the width and height of the target box and the prediction box respectively;

[0093] Step 4: Achieve high-precision positioning and perception of underground coal mine workers;

[0094] S41: Use an infrared camera and a pulse Doppler radar to scan the direction of underground coal mine workers, and obtain infrared data including underground coal mine workers and multiple radar RD data respectively;

[0095] S42: performing data preprocessing on the infrared data and the radar RD data respectively;

[0096] S43: The pre-processed infrared data and radar RD data are input as input data into the YOLO-T1 network model for processing, and the YOLO-T1 network model outputs the positioning result of the operator.

[0097] As a preferred embodiment, in step 2 S21, the pixel points that need to be denoised are determined by traversing all pixel points of the image, and then the pixel points that need to be denoised are set to the median value of the grayscale values ​​of all pixel points in the neighborhood window within the filtering radius of the point, so that the pixel points that need to be denoised are close to the true value to eliminate isolated noise points.

[0098] As a preferred embodiment, in step 3 S31;

[0099] The CBS module consists of a two-dimensional convolutional layer, a batch normalization layer and an activation layer, which is used to downsample the feature map;

[0100] The C2f module is used to pass the input data through the first CBS module, and then divide the output of the first CBS module into two parts, one of which is directly passed to the output, and the other part is passed through multiple Bottleneck depths to extract features. Finally, the results of the two parts are spliced ​​in the channel dimension and passed through the second CBS module to obtain the output of the C2f module, where Bottleneck is composed of two CBS modules;

[0101] The SPPF module is used to pass the input data through a CBS module once, and then through three maximum poolings. Each maximum pooling obtains a feature output. After splicing the three outputs, the SPPF module output is obtained after passing through a second CBS. The SPPF module obtains spatial information at multiple scales through maximum pooling, avoiding repeated extraction of graph-related features by the neural network and saving computing costs.

[0102] The BiFPN module is used to extract features from the backbone network and branch network of the YOLO-T1 network, output 4 features in different layers of the backbone network of the YOLO-T1 network, and output one feature in the branch network of the YOLO-T1 network. After being downsampled by the CBS module, these features are exchanged and transmitted through the BiFPN module to generate a feature pyramid, and a bidirectional connection is introduced between adjacent levels of the feature pyramid. At the same time, information from different levels of the pyramid is integrated in two directions, and a weighted feature fusion mechanism is adopted to combine features of different levels, and the training weights are learned and optimized during the training process.

[0103] As a preferred embodiment, in step 2 A2, the value of n is 4.

[0104] As a preferred embodiment, in A2 of step 2, the value of a is -0.5.

[0105] The present invention constructs a YOLO-T1 network model based on the YOLOv8 network, which can more effectively fuse radar RD data and infrared data, and can efficiently and accurately capture the features required for identification on the basis of low computing cost. In the training process of the YOLO-T1 network model, CIoU Loss is adopted. CIoU Loss can more accurately reflect the difference between the predicted box and the real box by introducing the center distance, aspect ratio and scale factor of the bounding box, so as to better optimize the detection performance during the training process. At the same time, CIoU Loss only needs to set one hyperparameter, which simplifies the setting process of the hyperparameter during the training process. It can converge faster during the training process, reduce training time and improve training efficiency. Therefore, by adopting CIoULoss, the center point distance, aspect ratio and scale factor can be comprehensively considered, and better performance and faster convergence speed can be provided in the target detection task.

[0106] This method can effectively expand the target detection dimension through the deep fusion of pulse Doppler radar and infrared camera, and can achieve high-precision positioning and perception of underground coal mine workers.

Claims

1. A method for accurate positioning and sensing of personnel based on Doppler radar and infrared camera, characterized in that: The steps include: Step 1: Use infrared cameras and pulse Doppler radars to scan the direction of underground coal mine workers. Through a massive scanning process, infrared data sets and multiple radar RD data sets containing underground coal mine workers are obtained respectively; Step 2: Preprocess the infrared data set according to S21, and preprocess the radar RD data set according to S22; S21: Preprocessing of infrared dataset: A1: Use the regularized median filtering method to denoise the salt and pepper noise in the infrared data; A2: Super-resolution reconstruction based on bicubic interpolation algorithm is performed on the infrared data after removing salt and pepper noise; Based on the surrounding n 2 pixels, by calculating n 2 The weight of each pixel is accumulated to obtain the pixel value I(x',y') of the super-resolution point; In the formula, I(i,j) represents the pixel value of point (i,j), and n represents the size of the window function, that is, according to the n near the super-resolution point 2 The super-resolution point value is calculated for each pixel, W(i,j) is the weight coefficient, W(i,j) = w(d i )*w(d j ), where d i =p(i,j).x-x',d j =p(i,j).y-y', Among them, the value range of a is [-1,0]; A3: Perform homomorphic filtering on the reconstructed infrared data; A31: Formula (2) is used to represent an image as the product of high frequency and low frequency components; f(x,y)=f i (x,y)f r (x,y) (2); In the formula, f i Represents the light intensity component that varies with spatial position, which is characterized by slow changes and concentration in the low-frequency part of the image, f r Represents the reflection component reflected to the human eye, which contains various information about the scene and is rich in high-frequency components; A32: Take a logarithmic transformation on formula (2) to obtain formula (3), and change the multiplicative component into an additive component; lnf(x,y)=lnf i (x,y)+lnf r (x,y) (3); A33: Use formula (4) to perform Fourier transform on the image to obtain the corresponding frequency domain; DFT[lnf(x,y)]=DFT[lnf i (x,y)]+DFT[lnf r (x,y)] (4); A34: Filter the image using a homomorphic filter according to formula (5); Where H(u,v) is the filter, Among them, γ H >1,γ L <1, control the range of filter amplitude; A35: Perform inverse Fourier transform and exponential transform on the image in sequence to restore the image and obtain an image after homomorphic filtering; S22: Preprocess the radar RD data set: B1: Convert the radar data into the same image format as the infrared data according to formula (6), that is, scale the radar RD data to an integer between [0,255]; Where V max is the global maximum value, p(i,j) and p'(i,j) are the points before and after scaling respectively; B2: Binarization of radar RD data to remove environmental noise and improve the quality of radar RD images; Step 3: Build the YOLOv8 network model based on the radar-infrared fusion target recognition algorithm; S31: Build the YOLO-T1 network model based on the YOLOv8 network; The YOLO-T1 network consists of four modules, namely the CBS module, C2f module, SPPF module and BiFPN module. In the backbone network of the YOLO-T1 network, CBS and C2f are used three times and an SPPF module is used at the end, and a total of 4 features are output at different layers. In the branch network of the YOLO-T1 network, the CBS module is used to downsample the feature map, and a C2f module is connected to further strengthen the features. Between CBS and C2f, the BiFPN module is introduced to weightedly fuse features to help the network capture features of different scales and output a total of one feature. Finally, the YOLO-T1 network outputs three detection heads to output the features and position of the target. S32: Train the YOLO-T1 network model; C1: For the preprocessed infrared dataset and radar RD dataset, the mosaic enhancement method is used to augment the data to obtain the augmented dataset; C2: Use the augmented data set to train the YOLO-T1 network model, and use formula (7) as the loss function; iteratively train the YOLO-T1 network model parameters by minimizing the loss function to obtain the optimal YOLO-T1 network model after training; Where d is the distance between the two center points of the predicted box and the real box, c is the diagonal length of the minimum circumscribed rectangle of the predicted box and the real box; α is the correction factor, in, (ω G ,h G ) and (ω G ,h G ) are the width and height of the target box and the prediction box respectively; Step 4: Achieve high-precision positioning and perception of underground coal mine workers; S41: Use an infrared camera and a pulse Doppler radar to scan the direction of underground coal mine workers, and obtain infrared data including underground coal mine workers and multiple radar RD data respectively; S42: performing data preprocessing on the infrared data and the radar RD data respectively; S43: The pre-processed infrared data and radar RD data are input as input data into the YOLO-T1 network model for processing, and the YOLO-T1 network model outputs the positioning result of the operator.

2. The method for accurate positioning and sensing of personnel based on Doppler radar and infrared camera according to claim 1 is characterized in that: In step 2 S21, the pixel points that need to be denoised are determined by traversing all the pixel points of the image, and then the pixel points that need to be denoised are set to the median of the grayscale values ​​of all the pixel points in the neighborhood window within the filtering radius of the point, so that the pixel points that need to be denoised are close to the true value to eliminate isolated noise points.

3. A method for accurate positioning and sensing of personnel based on Doppler radar and infrared camera according to claim 1 or 2, characterized in that: In step 3 S31; The CBS module consists of a two-dimensional convolutional layer, a batch normalization layer and an activation layer, which is used to downsample the feature map; The C2f module is used to pass the input data through the first CBS module, and then divide the output of the first CBS module into two parts, one of which is directly passed to the output, and the other part is passed through multiple Bottleneck depths to extract features. Finally, the results of the two parts are spliced ​​in the channel dimension and passed through the second CBS module to obtain the output of the C2f module, where Bottleneck is composed of two CBS modules; The SPPF module is used to pass the input data through the CBS module once, and then through the maximum pooling three times. Each maximum pooling obtains a feature output. After splicing the three outputs, it passes through the second CBS to obtain the SPPF module output; The BiFPN module is used to extract features from the backbone network and branch network of the YOLO-T1 network, generate a feature pyramid, and introduce bidirectional connections between adjacent levels of the feature pyramid. At the same time, it integrates information from different levels of the pyramid in two directions, and adopts a weighted feature fusion mechanism to combine features of different levels. The training weights are learned and optimized during the training process.

4. The method for accurate positioning and sensing of personnel based on Doppler radar and infrared camera according to claim 3 is characterized in that: In step 2 A2, the value of n is 4.

5. The method for accurate positioning and sensing of personnel based on Doppler radar and infrared camera according to claim 4 is characterized in that: In step 2, A2, the value of a is -0.5.

Images