A multi-modal based audio-video recognition system

Abstract

The application provides a kind of audio and video recognition system based on multi-modal, the system includes: data acquisition module: for collecting the voiceprint signal and unmanned aerial vehicle image of unmanned aerial vehicle in mining area;Voiceprint data processing module: for converting the voiceprint signal of unmanned aerial vehicle into voiceprint feature map;Image data processing module: for converting the image of unmanned aerial vehicle into the multi-dimensional image feature vector of unmanned aerial vehicle;Feature correlation network module: for building encoder-attention mechanism fusion network;Feature correlation optimization module: for training encoder-attention mechanism fusion network using contrast loss function, improve the mapping matching accuracy of unmanned aerial vehicle voiceprint and image feature;Early warning decision module: for real-time early warning to invading unmanned aerial vehicle, and push early warning information to mining area security terminal.The application realizes active perception and rapid disposal to unmanned aerial vehicle target through unmanned aerial vehicle voiceprint-image cross-modal technology deep fusion, and guarantees the safety of mining production.

Description

Technical Field

[0001] This invention relates to the field of audio and video recognition technology for unmanned aerial vehicles (UAVs), and more particularly to an audio and video recognition system based on multimodal operation. Background Technology

[0002] With the large-scale application of drones in civilian fields such as logistics, agricultural plant protection, geographic surveying, and aerial filming, their "low, slow, and small" characteristics and convenient take-off and landing capabilities have also given rise to the problem of "unauthorized flights." Incidents of unauthorized drones trespassing into no-fly zones and disrupting air traffic order are frequent, posing a continuous threat to public safety and privacy. Especially in mining environments with concentrated high-value infrastructure and intensive production activities, drone intrusions could lead to production data leaks, equipment malfunctions, illegal collection of mining area maps, and even extreme security incidents such as the dropping of explosives. However, traditional security systems centered on manual patrols and fixed monitoring are insufficient for proactive detection and rapid response to low-altitude drone targets, revealing inherent flaws such as large monitoring blind spots and long response chains.

[0003] Currently, detection and early warning of drones mainly rely on technologies such as radar, radio spectrum monitoring, acoustic sensing, and photoelectric identification. However, in typical complex scenarios like mining areas, all these technologies face significant limitations. Acoustic sensing is greatly affected by environmental noise, has a short effective range, and is difficult to achieve accurate positioning and target classification; while photoelectric identification can provide intuitive visual evidence, its monitoring range is limited, is significantly constrained by lighting and weather conditions, and its performance drops sharply at night or in foggy or hazy weather. Summary of the Invention

[0004] Therefore, the purpose of this invention is to provide a multimodal audio and video recognition system to solve or at least partially solve the above-mentioned problems existing in the prior art.

[0005] To achieve the above objectives, the present invention provides a multimodal audio and video recognition system, the system comprising: Data acquisition module: used to acquire the acoustic signature signals and images of drones in the mining area; Voiceprint data processing module: used to convert the voiceprint signal of the UAV into a voiceprint feature map; Image data processing module: used to convert UAV images into multi-dimensional image feature vectors of the UAV; Feature association network module: used to build encoder-attention mechanism fusion network to realize the mapping and matching of drone voiceprint and image features, and to identify intruding drones based on the mapping and matching of drone voiceprint and image features; Feature association optimization module: used to train the encoder-attention mechanism fusion network using the contrastive loss function, thereby improving the mapping and matching accuracy of drone voiceprints and image features and the accuracy of intrusion drone identification; Early warning decision module: used to provide real-time early warning of intruding drones and push the early warning information to the mine security terminal.

[0006] Furthermore, multiple data acquisition modules are deployed within the mining area. Each data acquisition module includes an acoustic fingerprint sensor and a high-definition camera. The acoustic fingerprint sensor is deployed at a preset interval around the boundary of the mining area and the perimeter of the work area of ​​the production area to collect acoustic fingerprint signals generated by drones during flight within the monitoring range. The high-definition camera's field of view covers the monitoring range of the acoustic fingerprint sensor and is used to collect drone images.

[0007] Furthermore, the conversion of the drone's voiceprint signal into a voiceprint feature map specifically includes the following steps: S11. Use short-time Fourier transform to convert the acoustic signature signal into a two-dimensional spectrum. S12. Calculate the power spectrum of the two-dimensional spectrum and convert the complex spectrum into an energy representation; S13. Remove low-frequency noise from the mining area in the two-dimensional spectrum using Mel filter; S14. Perform Mel spectrum calculation on the two-dimensional spectrogram; S15. Adjust the size of the Mel spectrum of the two-dimensional spectrogram and then normalize it to generate a voiceprint feature map.

[0008] Furthermore, the conversion of the drone image into a multi-dimensional image feature vector of the drone specifically includes the following steps: S21. Use the data acquisition module to collect drone images of different types of intrusive drones in different scenarios in the mining area, and divide the drone images into training samples and verification samples according to a preset ratio. S22. Combine a lightweight neural network model with a multi-scale feature path aggregation network to construct a UAV image feature recognition model, and replace the original classification head of the model with a target detection branch and a feature extraction branch. S23. Input the training samples into the UAV image feature recognition model for training. Output the UAV bounding box through the target detection branch. Output the multi-dimensional image feature vector of the UAV based on the UAV bounding box through the feature extraction branch. S24. The UAV image feature recognition model is trained using cross-entropy loss and smooth L1 loss. A total loss function is constructed using weighted coefficients. The parameters of the UAV image feature recognition model are updated by backpropagation based on the total loss function to complete the training of the UAV image feature recognition model. Validation samples are input into the trained UAV image feature recognition model to verify the accuracy of the UAV image feature recognition model in detecting UAVs.

[0009] Furthermore, the drone images are filtered to remove image noise caused by mining dust and haze, data enhancement is performed on the drone images, the image size is scaled to a preset size, and the drone and images are standardized.

[0010] Furthermore, the feature association network module is specifically used to perform the following steps: S31. Construct an encoder-attention mechanism fusion network, including an encoder and an attention mechanism. The encoder includes a multi-layer convolutional neural network. The multi-layer convolutional neural network is used to extract features from the voiceprint feature map and output a multi-dimensional voiceprint feature vector. S32. Calculate the cosine similarity matrix between the multidimensional voiceprint feature vector and the multidimensional image feature vector. Using an attention mechanism, assign weights to the multidimensional voiceprint feature vector and the multidimensional image feature vector based on their similarity to generate a fused feature vector, as shown below:

[0011] in, To fuse feature vectors, For attention weights, For multidimensional image feature vectors, This is a multidimensional voiceprint feature vector; S33. The feature association network module is also equipped with a UAV feature vector. The similarity between the UAV feature vector and the fused feature vector is calculated. If the similarity is greater than or equal to the similarity threshold, it is determined to be a valid intrusion UAV; otherwise, it is determined to be a verification error.

[0012] Furthermore, the feature association optimization module is specifically used to perform the following steps: S41. Construct a voiceprint-image pairing dataset, including multiple sets of positive samples and multiple sets of negative samples. Each set of positive samples consists of the voiceprint and image of a different UAV, and the negative samples consist of mining area noise and non-UAV images. S42. Input the voiceprint-image pairing dataset into the encoder-attention mechanism fusion network and train it using the contrastive loss function value. The contrastive loss function is trained with the goal of minimizing the Euclidean distance of positive sample pairs and maximizing the Euclidean distance of negative sample pairs. S43. Update the parameters of the encoder-attention mechanism fusion network by backpropagation based on the contrast loss function value to improve the mapping and matching accuracy of UAV voiceprints and image features.

[0013] Furthermore, the early warning decision module is specifically used to perform the following steps: S51. Install audible and visual alarm devices at the boundary of the mining area and deploy anti-drone jammers in the production area; S52. If the intrusion drone is determined to be valid, a warning message is pushed to the mine security terminal in real time. The warning message includes the location, flight direction and real-time image of the intrusion drone. At the same time, the security terminal links the sound and light alarm device at the boundary of the mine to issue an alarm signal and triggers the anti-drone jammer in the production area. S53. Store the drone voiceprint signals and image data of valid drone intrusion incidents into the database for source tracing and model optimization.

[0014] Compared with the prior art, the beneficial effects of the present invention are: This invention proposes a multimodal audio and video recognition system. It removes low-frequency noise from mining areas using a voiceprint data processing module to generate a voiceprint feature map, removes image noise from mining areas using an image data processing module to generate a multidimensional image feature vector, maps and matches the drone's voiceprint with image features using a feature association network module, improves the accuracy of the mapping and matching using a feature association optimization module, and provides real-time early warning for the drone using an early warning decision module. This invention achieves proactive perception and rapid response to drone targets through deep fusion of drone voiceprint and image cross-modal technology, ensuring the safety of mining production. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only preferred embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of the overall structure of a multimodal audio and video recognition system provided in an embodiment of the present invention. Detailed Implementation

[0017] The principles and features of the present invention are described below with reference to the accompanying drawings. The listed embodiments are only used to explain the present invention and are not intended to limit the scope of the present invention.

[0018] Reference Figure 1 This embodiment provides a multimodal audio and video recognition system, the system comprising: Data acquisition module: used to acquire the acoustic signature signals and images of drones in the mining area, specifically including: Multiple data acquisition modules are deployed within the mining area. Each data acquisition module includes an acoustic fingerprint sensor and a high-definition camera. The acoustic fingerprint sensor is deployed at a preset interval around the boundary of the mining area and the perimeter of the work area of ​​the production area to collect acoustic fingerprint signals generated by drones during flight within the monitoring range. The high-definition camera has a field of view that covers the monitoring range of the acoustic fingerprint sensor and is used to collect drone images. Based on changes in the mining environment, such as day / night, strong winds / no wind, the voiceprint detection frequency band (e.g., reducing the background noise threshold at night) and cross-modal matching threshold can be adjusted in real time to reduce false detections caused by environmental interference.

[0019] By combining the drone's acoustic signature with a high-definition camera, industrial noise and dust interference in the mining area can be filtered out, intruding drones can be quickly located, and full coverage can be achieved day and night and in complex terrain, eliminating blind spots for manual inspections.

[0020] By combining the voiceprint signal of drones with high-definition cameras, deployment and maintenance costs are reduced, and installation and maintenance are convenient. Compared with traditional security solutions, it saves manpower costs for deployment and maintenance and is suitable for large-scale applications in mining areas.

[0021] Voiceprint data processing module: Used to convert the drone's voiceprint signal into a voiceprint feature map. This includes using short-time Fourier transform to convert the voiceprint signal into a two-dimensional spectrogram (dimension 256×256), removing low-frequency noise such as mine fan and truck noise through Mel filtering, and then normalizing (mapping pixels to 0-1) to generate a voiceprint feature map, which serves as the input basis for cross-modal matching. Specifically, it includes the following steps: S11. The short-time Fourier transform is used to convert the acoustic signature signal into a two-dimensional spectrum, which can completely preserve the frequency composition of the acoustic signature signal at different times, accurately capture the time-frequency characteristic changes during the flight of the UAV, and when there are multiple UAVs, the acoustic signature spectrum characteristics of different UAVs will be represented as energy bands in different frequency ranges in the two-dimensional spectrum. By separating the energy distribution of each frequency band, the number of multiple targets and the model differentiation can be realized, as shown below:

[0022] in, For the first Frame number Complex spectrum at each frequency point For the first The audio signal of the frame, For Hamming window functions, For FFT points, The imaginary unit, It is a natural constant. This is the sampling point index of the voiceprint signal, corresponding to the discrete sampling position of the voiceprint signal in the time domain, with a value range from 0 to N-1; S12. Calculate the power spectrum of the two-dimensional spectrogram and convert the complex spectrum into an energy representation. The energy representation will form significant peaks at the corresponding time-frequency positions, creating a clear difference from the low-energy distribution of the background noise, which facilitates subsequent threshold segmentation and peak extraction, as shown below:

[0023] in, For the first Frame number Power spectrum at each frequency point; S13. Low-frequency noise from the mining area in the two-dimensional spectrum is removed by Mel filtering. Mel filtering provides finer segmentation of the low-frequency band, accurately identifying and filtering out typical low-frequency noise in the mining area, thus avoiding damage to the effective mid-to-high frequency band of the acoustic signature signal. This is illustrated below:

[0024] in, For the first The frequency response function of a Mel filter For voiceprint signals, (This refers to the continuous frequency variables.) For the first The center frequency of the Mel filter For the first The center frequency of the Mel filter, For the first The center frequency of the Mel filter; S14. Perform Mel-frequency spectrum calculation on the two-dimensional spectrogram to achieve fine low-frequency resolution and high-frequency compression and merging of the UAV's voiceprint signal, thereby improving the recognition and anti-interference capability of the UAV's voiceprint signal, as shown below:

[0025] in, For the first The frame of the voiceprint signal passes through the first Mel spectrum value after Mel filter Minimum value ; S15. The Mel spectrum of the two-dimensional spectrogram is resized and then normalized. Bilinear interpolation is used to adjust the Mel spectrum of any size to 256×256, generating a voiceprint feature map, as shown below:

[0026]

[0027] in, After bilinear interpolation adjustment, the size is The first in the Mel spectrum matrix Line number The element values ​​of the column, For the original Mel spectrum, For the final generated voiceprint feature map, the first... Line number The normalized eigenvalues ​​of the column.

[0028] Image data processing module: Used to convert UAV images into multi-dimensional image feature vectors of the UAV, specifically including the following steps: S21. Use the data acquisition module to collect drone images of different types of intrusion drones in different scenarios in the mining area. The different scenarios in the mining area include sunny days, cloudy days, dusty weather, etc. Divide the drone images into 10,000 training samples and 2,000 verification samples according to a preset ratio. The collected drone images are filtered to remove image noise caused by mining dust, haze and other factors. Data augmentation (random horizontal flip and ±15° rotation) is performed on the drone images to avoid the model's over-reliance on the drone's flight angle. At the same time, the training sample size is expanded, the image size is scaled to 640×640 (to adapt to the input requirements of the drone image feature recognition model), and the drone images are standardized to improve the model's convergence speed. S22. Combine a lightweight neural network model with a multi-scale feature path aggregation network to construct a UAV image feature recognition model, and replace the original classification head of the model with a target detection branch and a feature extraction branch. S23. Input the training samples into the UAV image feature recognition model. Output the UAV bounding box (positioning accuracy ≤ 5 pixels) through the target detection branch. Output the 2048-dimensional image feature vector of the UAV based on the UAV bounding box through the feature extraction branch. It can adapt to the task requirements of UAV image recognition, accurately output target classification and detection results, and improve the model's ability to distinguish UAV images. S24. The UAV image feature recognition model is trained using cross-entropy loss and smooth L1 loss. A total loss function is constructed using weighted coefficients. The parameters of the UAV image feature recognition model are updated by backpropagation based on the total loss function to complete the training of the UAV image feature recognition model. Validation samples are input into the trained UAV image feature recognition model to verify the accuracy of the UAV image feature recognition model in detecting UAVs.

[0029] The cross-entropy loss is used for classification tasks, corresponding to the distinction between drones and non-drones in the feature extraction branch, as shown below:

[0030] in, For real labels, To predict class probabilities for the model, To represent row indices in the feature map / matrix, Total number of categories; The smooth L1 loss is used for the regression task, corresponding to the coordinate regression of the UAV bounding box in the object detection branch, as shown below:

[0031] in, To calculate the difference between the predicted coordinates and the actual coordinates; The total loss function is expressed as follows:

[0032] in, For the total loss function, For the weight of the classification task, To regress task weights, =0.4, By validating and debugging the samples, the weights of the classification and regression tasks were determined and balanced.

[0033] Feature Association Network Module: Used to construct an encoder-attention mechanism fusion network to achieve mapping and matching of drone voiceprints and image features. Based on this mapping and matching, it identifies intruding drones, focusing on key cross-modal features, establishing accurate associations between drone voiceprints and images, improving the robustness and accuracy of drone identification in complex environments, and generating fused feature vectors with richer discriminative power compared to single-modal features. It can effectively distinguish drones of similar models or with similar voiceprint or image features, improving classification and detection accuracy. Specifically, it includes the following steps: S31. Construct an encoder-attention mechanism fusion network, including an encoder and an attention mechanism. The encoder includes a multi-layer convolutional neural network. A 3-layer convolutional neural network is used to extract features from the voiceprint feature map and output a 2048-dimensional voiceprint feature vector. S32. Calculate the cosine similarity matrix between the 2048-dimensional voiceprint feature vector and the 2048-dimensional image feature vector. Using a soft attention mechanism, assign weights to the 2048-dimensional voiceprint feature vector and the 2048-dimensional image feature vector based on their similarity, generating a fused feature vector as follows:

[0034] in, To fuse feature vectors, Attention weights (dynamically adjustable from 0.4 to 0.6). For multidimensional image feature vectors, This is a multidimensional voiceprint feature vector; S33. The feature association network module is also equipped with a drone feature vector. It calculates the similarity between the drone feature vector and the fused feature vector. If the similarity is greater than or equal to a similarity threshold, it is determined to be a valid intrusion drone; otherwise, it is determined to be a verification error (e.g., birds, fallen plastic bags, etc.).

[0035] Feature association optimization module: This module is used to train an encoder-attention mechanism fusion network using a contrastive loss function to improve the mapping and matching accuracy of drone voiceprints and image features, as well as the accuracy of intrusion drone identification. Specifically, it includes the following steps: S41. Construct a voiceprint-image pairing dataset, including multiple sets of positive samples and multiple sets of negative samples, totaling 8000 sets of positive samples and 2000 sets of negative samples; Voiceprint and image data of common interfering targets in mining areas, such as birds, kites, and mining machinery parts, are collected to expand the negative sample set and reduce the misjudgment rate of interfering targets. S42. Input the voiceprint-image pairing dataset into the encoder-attention mechanism fusion network and train it using a contrastive loss function. The contrastive loss function aims to minimize the Euclidean distance between positive sample pairs and maximize the Euclidean distance between negative sample pairs, so that similar samples are clustered and dissimilar samples are separated in the cross-modal feature space. Each positive sample pair consists of the voiceprint feature vector and image feature vector of different UAVs, and the negative sample pair consists of the voiceprint feature vector of mining noise and the image feature vector of non-UAVs, as shown below:

[0036] in, The Euclidean distance between positive sample pairs is given. The Euclidean distance between negative sample pairs. To define the distance boundaries between positive and negative sample pairs and ensure that the distance between the two types of samples has clear distinguishability; S43. Update the parameters of the encoder-attention mechanism fusion network by backpropagation based on the contrast loss function value to improve the mapping and matching accuracy of UAV voiceprints and image features.

[0037] Early warning decision module: Used to provide real-time early warning of intruding drones and push the warning information to the mine security terminal. Specifically, it includes the following steps: S51. Install audible and visual alarm devices at the boundary of the mining area and deploy anti-drone jammers in the production area; S52. If the intrusion drone is determined to be valid, a warning message is pushed to the mine security terminal in real time. The warning message includes the location, flight direction and real-time image of the intrusion drone. At the same time, the security terminal links the sound and light alarm device at the boundary of the mine to issue an alarm signal and triggers the anti-drone jammer in the production area. S53. Store the drone voiceprint signals and image data of valid drone intrusion incidents into the database for source tracing and model optimization.

[0038] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multimodal audio and video recognition system, characterized in that, The system includes: Data acquisition module: used to acquire the acoustic signature signals and images of drones in the mining area; Voiceprint data processing module: used to convert the voiceprint signal of the UAV into a voiceprint feature map; Image data processing module: used to convert UAV images into multi-dimensional image feature vectors of the UAV; Feature association network module: used to build encoder-attention mechanism fusion network to realize the mapping and matching of drone voiceprint and image features, and to identify intruding drones based on the mapping and matching of drone voiceprint and image features; Feature association optimization module: used to train the encoder-attention mechanism fusion network using the contrastive loss function to improve the mapping and matching accuracy of drone voiceprints and image features, as well as the accuracy of intrusion drone identification; Early warning decision module: used to provide real-time early warning of intruding drones and push the early warning information to the mine security terminal.

2. The multimodal audio and video recognition system according to claim 1, characterized in that, Multiple data acquisition modules are deployed within the mining area. Each data acquisition module includes an acoustic fingerprint sensor and a high-definition camera. The acoustic fingerprint sensor is deployed at a preset interval around the boundary of the mining area and the perimeter of the work area of ​​the production area to collect acoustic fingerprint signals generated by drones during flight within the monitoring range. The high-definition camera's field of view covers the monitoring range of the acoustic fingerprint sensor and is used to collect drone images.

3. The multimodal audio and video recognition system according to claim 1, characterized in that, The process of converting the drone's voiceprint signal into a voiceprint feature map specifically includes the following steps: S11. Use short-time Fourier transform to convert the acoustic signature signal into a two-dimensional spectrum. S12. Calculate the power spectrum of the two-dimensional spectrum and convert the complex spectrum into an energy representation; S13. Remove low-frequency noise from the mining area in the two-dimensional spectrum using Mel filter; S14. Perform Mel spectrum calculation on the two-dimensional spectrogram; S15. Adjust the size of the Mel spectrum of the two-dimensional spectrogram and then normalize it to generate a voiceprint feature map.

4. The multimodal audio and video recognition system according to claim 3, characterized in that, The process of converting drone images into multidimensional image feature vectors for drones specifically includes the following steps: S21. Use the data acquisition module to collect drone images of different types of intrusive drones in different scenarios in the mining area, and divide the drone images into training samples and verification samples according to a preset ratio. S22. Combine a lightweight neural network model with a multi-scale feature path aggregation network to construct a UAV image feature recognition model, and replace the original classification head of the model with a target detection branch and a feature extraction branch. S23. Input the training samples into the UAV image feature recognition model, output the UAV bounding box through the target detection branch, and output the multi-dimensional image feature vector of the UAV based on the UAV bounding box through the feature extraction branch; S24. Train the UAV image feature recognition model using cross-entropy loss and smooth L1 loss, and construct a total loss function through weighted coefficients. Update the parameters of the UAV image feature recognition model through backpropagation based on the total loss function to complete the training of the UAV image feature recognition model. Input the validation samples into the trained UAV image feature recognition model to verify the accuracy of the UAV image feature recognition model in detecting UAVs.

5. The multimodal audio and video recognition system according to claim 4, characterized in that, The drone images are filtered to remove image noise caused by mining dust and haze, data enhancement is performed on the drone images, the image size is scaled to a preset size, and the drone and images are standardized.

6. The multimodal audio and video recognition system according to claim 4, characterized in that, The feature association network module is used to specifically perform the following steps: S31. Construct an encoder-attention mechanism fusion network, including an encoder and an attention mechanism. The encoder includes a multi-layer convolutional neural network. The multi-layer convolutional neural network is used to extract features from the voiceprint feature map and output a multi-dimensional voiceprint feature vector. S32. Calculate the cosine similarity matrix between the multidimensional voiceprint feature vector and the multidimensional image feature vector. Using an attention mechanism, assign weights to the multidimensional voiceprint feature vector and the multidimensional image feature vector based on their similarity to generate a fused feature vector, as shown below: in, To fuse feature vectors, For attention weights, For multidimensional image feature vectors, This is a multidimensional voiceprint feature vector; S33. The feature association network module is also equipped with a UAV feature vector. The similarity between the UAV feature vector and the fused feature vector is calculated. If the similarity is greater than or equal to the similarity threshold, it is determined to be a valid intrusion UAV; otherwise, it is determined to be a verification error.

7. The multimodal audio and video recognition system according to claim 6, characterized in that, The feature association optimization module is used to specifically perform the following steps: S41. Construct a voiceprint-image pairing dataset, including multiple sets of positive samples and multiple sets of negative samples. Each set of positive samples consists of voiceprints and images from different drones, and the negative samples consist of mining area noise and non-drone images. S42. Input the voiceprint-image pairing dataset into the encoder-attention mechanism fusion network and train it using a contrastive loss function. The contrastive loss function aims to minimize the Euclidean distance between positive sample pairs and maximize the Euclidean distance between negative sample pairs. S43. Update the parameters of the encoder-attention mechanism fusion network by backpropagation based on the contrast loss function to improve the mapping and matching accuracy of UAV voiceprints and image features.

8. The multimodal audio and video recognition system according to claim 6, characterized in that, The early warning decision module is used to execute the following steps: S51. Install audible and visual alarm devices at the boundary of the mining area and deploy anti-drone jammers in the production area; S52. If the intrusion drone is determined to be valid, a warning message is pushed to the mine security terminal in real time. The warning message includes the location, flight direction and real-time image of the intrusion drone. At the same time, the security terminal links the sound and light alarm device at the boundary of the mine to issue an alarm signal and triggers the anti-drone jammer in the production area. S53. Store the drone voiceprint signals and image data of valid drone intrusion incidents into the database for source tracing and model optimization.

Images