An unmanned aerial vehicle detection system and method based on vision, lidar and sound waves

Abstract

The present application relates to the technical field of unmanned aerial vehicle detection and identification, in particular to an unmanned aerial vehicle detection system and method based on vision, laser radar and sound waves, which comprises a sensor unit, a data processing fusion unit and a target identification and tracking unit, the present application synchronously collects images, three-dimensional point clouds and sound wave signals through the sensor unit, carries out multi-source data alignment and weighted fusion on the images to extract texture, shape and color features through the data processing fusion unit, generates a depth map from the point clouds, maps the sound wave signals into a two-dimensional feature map, dynamically allocates weights based on sensor reliability to generate a unified feature map, the target identification and tracking unit identifies the type of unmanned aerial vehicle using an optimized deep learning model, realizes accurate prediction and update of the position and speed state in combination with Kalman filtering, the decision response unit triggers an audible and visual alarm and a wireless alarm in real time for non-cooperative unmanned aerial vehicles, and transmits target dynamic information to the ground station, thereby improving the target detection accuracy of unmanned aerial vehicles in complex environments.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of unmanned aerial vehicle detection and identification, in particular to an unmanned aerial vehicle detection system and method based on vision, laser radar and sound wave. BACKGROUND

[0002] Unmanned aerial vehicle detection and identification is an important technology. In the context of low-altitude safety management, an unmanned aerial vehicle detection system can perceive aerial targets in real time, provide early warning and prevention and control for threats such as illegal intrusion and privacy stealing, and is a key means to ensure the safety of sensitive areas such as airports and event venues. With the diversification of unmanned aerial vehicle application scenarios, there is an increasing demand for multi-target detection, cooperation and classification identification of cooperative and non-cooperative unmanned aerial vehicles in complex environments. Precise multi-source data fusion and target tracking technology has become the core of improving system reliability.

[0003] However, traditional unmanned aerial vehicle detection technology has the core problems of insufficient depth of multi-sensor data fusion and low target recognition accuracy. Existing solutions mostly use a single sensor to work independently, lack cross-modal correlation analysis of image texture, three-dimensional point cloud and sound wave signals, and rely solely on vision sensors, which can easily cause image blurring under complex lighting conditions, leading to target missed detection. A single laser radar cannot distinguish unmanned aerial vehicles from interference targets such as birds, and sound wave sensors cannot stably capture unmanned aerial vehicle engine characteristics due to environmental noise. In addition, traditional methods do not establish a dynamic data fusion mechanism, and when the reliability of single sensor data decreases in rain, fog, electromagnetic interference and other scenarios, the system cannot automatically switch to the dominant data source, resulting in poor target tracking continuity. This one-sidedness of data fusion and lack of environmental adaptability makes the target detection accuracy of traditional systems low in complex environments. To solve this problem, we provide an unmanned aerial vehicle detection system and method based on vision, laser radar and sound wave. SUMMARY

[0004] The purpose of the present application is to provide an unmanned aerial vehicle detection system and method based on vision, laser radar and sound wave to solve the problems raised in the background.

[0005] 1. Since traditional solutions use a single sensor to work independently, the depth of data fusion is insufficient, leading to target missed detection and misjudgment. Therefore, the present case fuses multi-source data of vision, laser radar and sound wave through the sensor unit, and the data processing and fusion unit performs feature extraction and weighted fusion, which can improve the data comprehensiveness and correlation analysis capability, and improve the target detection accuracy.

[0006] 2. Since traditional methods do not establish a dynamic data fusion mechanism, the environmental adaptability is insufficient, leading to poor tracking continuity. Therefore, the present case dynamically allocates weights according to the reliability of the sensor through the data processing and fusion unit, and the target recognition and tracking unit combines deep learning and Kalman filtering, which can dynamically adapt to complex environments and improve the stability of target tracking and recognition accuracy.

[0007] To achieve the above objectives, a drone detection system based on vision, lidar, and sound waves is provided, comprising:

[0008] The sensor unit includes a visual sensor, a lidar, and an acoustic sensor, which are used to collect image data, 3D point cloud data, and acoustic signals around the drone, respectively.

[0009] The data processing and fusion unit extracts the texture, shape, and color features of image data through the data preprocessing module, projects the point cloud data onto the image plane to generate a depth map, extracts time-domain, frequency-domain, and cepstral features of the acoustic signal, and maps them to a two-dimensional feature map for multi-source data alignment. Then, the feature fusion module performs weighted fusion of the preprocessed visual, lidar, and acoustic features to generate a unified feature map.

[0010] The target recognition and tracking unit uses a deep learning model to detect and classify targets on a unified feature map, identifies drones and distinguishes between cooperative and non-cooperative targets, and uses a Kalman filter module to combine the target dynamic model with the deep learning detection results to predict and update the drone's position, velocity and acceleration.

[0011] The decision response unit identifies non-cooperative drones based on classification results and tracking information, and triggers an alarm signal when a non-cooperative drone is detected. At the same time, it outputs the target position, speed and category information to the ground station display.

[0012] As a further improvement to this technical solution, the data processing and fusion unit extracts the texture, shape, and color features of the image data, including the following steps:

[0013] By removing noise through Gaussian filtering, the image pixel values ​​are normalized to the range of [0,1], and edge features, texture features, and color features are extracted.

[0014] As a further improvement to this technical solution, the data processing and fusion unit projects point cloud data onto an image plane to generate a depth map, including the following steps:

[0015] By removing noise from the point cloud through statistical filtering, the 3D point cloud is projected onto the 2D image plane to generate a 2D depth map containing depth information.

[0016] As a further improvement to this technical solution, the data processing and fusion unit extracts time-domain, frequency-domain, and cepstral features of the acoustic signal and maps them to a two-dimensional feature map, including the following steps:

[0017] Noise is removed by frequency domain filtering or wavelet transform, and the time-domain amplitude features, frequency-domain spectrum features, and cepstral coefficient features of the acoustic signal are extracted. These features are then mapped into a two-dimensional matrix to align with visual and lidar data.

[0018] As a further improvement to this technical solution, in the data processing fusion unit, the feature fusion module performs weighted fusion including the following steps:

[0019] Based on the dynamic allocation of weights according to the reliability of different sensors, the visual feature map, the lidar depth map feature map, and the acoustic two-dimensional feature map are weighted and summed to generate a unified multi-source fusion feature map.

[0020] As a further improvement to this technical solution, in the target recognition and tracking unit, the deep learning model performs target detection and classification on the unified feature map, including the following steps:

[0021] An optimized deep learning model is used as the basic architecture. High-level features of the fused data are extracted through convolutional layers to identify target locations and classify them as cooperative or non-cooperative drones.

[0022] As a further improvement to this technical solution, in the target recognition and tracking unit, the Kalman filter module combines the target dynamic model with the deep learning detection results to perform prediction and state update, including the following steps:

[0023] The target dynamic model predicts the position, velocity, and acceleration at the next moment, and the deep learning detection results are used as measurements to update the Kalman filter state estimate.

[0024] As a further improvement to this technical solution, the target identification and tracking unit includes the following steps for identifying non-cooperative drones:

[0025] Analyze whether the target's flight path is unknown or unauthorized, detect whether the signal characteristics are abnormal, and combine the classification results and tracking information to make a comprehensive judgment on the non-cooperative status.

[0026] As a further improvement to this technical solution, the decision response unit triggers the alarm signal by including the following steps:

[0027] When a non-cooperative drone is identified, at least one of the following is activated: an audible alarm, a visual alarm, or a wireless signal alarm. Simultaneously, the alarm information is transmitted wirelessly to the ground station. The ground station outputs the target's location, speed, and category information, including the following steps:

[0028] Target status data packets are generated in real time and transmitted to the ground station display via the communication interface. The position coordinates, velocity vector, and target category identifier are dynamically displayed on the ground station interface.

[0029] The second objective of this invention is to provide a method for implementing a drone detection system based on vision, lidar, and sound waves, comprising the following steps:

[0030] S1. The device collects image data, 3D point cloud data, and sound signals around the drone using visual sensors, lidar, and acoustic sensors.

[0031] S2. Perform Gaussian filtering on the image data to remove noise, normalize the pixel values ​​to the range of [0,1], extract edge, texture and color features, perform statistical filtering on the point cloud data to remove noise, and project the three-dimensional point cloud onto the two-dimensional image plane to generate a depth map. Perform frequency domain filtering on the acoustic signal to remove noise, extract the time domain amplitude, frequency domain spectrum and cepstral coefficient features, and map them into a two-dimensional matrix feature map.

[0032] S3. Based on the dynamic allocation of weights according to the reliability of each sensor, the visual feature map, depth map feature map and acoustic two-dimensional feature map are weighted and summed to generate a unified multi-source fusion feature map.

[0033] S4. Input the fused feature map into the optimized deep learning model, extract high-level features through convolutional layers to identify the target location, and classify it into cooperative and non-cooperative drones. Based on the target dynamic model, predict the position, velocity and acceleration. Input the deep learning detection results into the Kalman filter module to update the target state estimate.

[0034] S5 analyzes whether the target's flight path is abnormal and whether the signal characteristics are unauthorized. Combining the classification results and tracking information, it determines the non-cooperative drone, triggers an alarm, and generates a target status data packet in real time. Finally, it transmits the data to the ground station display via the communication interface to dynamically display the target's real-time status.

[0035] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0036] This invention significantly improves the reliability of UAV detection through multi-sensor fusion and intelligent processing technology. The sensor unit integrates visual, lidar, and acoustic data, covering image texture, 3D point cloud, and acoustic features, solving the problem of one-sided information from a single sensor. The data processing and fusion unit achieves multi-source data alignment through Gaussian filtering, point cloud projection, and feature mapping, and dynamically weights the fusion based on sensor reliability to improve data credibility in complex environments. The target recognition and tracking unit uses a deep learning model to extract fused features to distinguish between cooperative and non-cooperative targets, and combines Kalman filtering to predict target status, achieving accurate detection and continuous tracking. The decision response unit triggers alarms and transmits target information in real time, ensuring the timeliness of low-altitude safety control and improving the accuracy of target detection in complex environments. Attached Figure Description

[0037] Figure 1 This is an overall block diagram of the present invention;

[0038] Figure 2 This is the overall flowchart of the present invention.

[0039] The meanings of the labels in the diagram are as follows:

[0040] 1. Sensor unit; 2. Data processing and fusion unit; 3. Target recognition and tracking unit; 4. Decision response unit. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] Example 1

[0043] One objective of this invention is to provide a drone detection system based on vision, lidar, and sound waves. (See also...) Figure 1 As shown, it includes:

[0044] Sensor unit 1 includes a visual sensor, a lidar and an acoustic sensor, which are used to collect image data, 3D point cloud data and acoustic signals around the UAV, respectively.

[0045] The data processing and fusion unit 2 extracts the texture, shape and color features of the image data through the data preprocessing module, projects the point cloud data onto the image plane to generate a depth map, extracts the time domain, frequency domain and cepstral features of the acoustic signal, and maps them to a two-dimensional feature map for multi-source data alignment. Then, the feature fusion module performs weighted fusion of the preprocessed visual, lidar and acoustic features to generate a unified feature map.

[0046] The extraction of texture, shape, and color features from image data in data processing and fusion unit 2 includes the following steps:

[0047] During image acquisition, images are susceptible to sensor noise and ambient light interference. Filtering is necessary to improve image quality and prevent noise from affecting the accuracy of subsequent feature extraction. Gaussian filtering removes noise and normalizes image pixel values ​​to the [0,1] range, extracting edge features, texture features, and color features, enhancing the reliability of feature extraction, and improving the overall detection accuracy of UAV targets.

[0048] The data processing and fusion unit 2 projects point cloud data onto an image plane to generate a depth map, including the following steps:

[0049] Point clouds acquired by lidar contain outliers and redundant data, requiring filtering to improve point cloud quality and prevent noise from distorting the depth map. Statistical filtering or voxel filtering is used to remove point cloud noise. The distance between each point and its neighbors is calculated, and outliers with distances exceeding the mean ± 2 standard deviations are removed. The 3D point cloud is then projected onto a 2D image plane for pixel-level alignment with the visual image, achieving spatial registration of multi-source data. The Zhang Zhengyou calibration method is used to obtain the intrinsic parameter matrix of the visual camera, including focal length and optical center coordinates. Perspective projection is performed on each point cloud coordinate to generate a 2D depth map containing depth information, providing a spatial alignment basis for cross-modal feature fusion.

[0050] In data processing and fusion unit 2, the feature fusion module performs weighted fusion, including the following steps:

[0051] Sensor reliability varies under different environments, requiring dynamic weight adjustment to highlight reliable data sources. Weights are dynamically allocated based on the reliability of different sensors. Reliability indices are calculated for visual sensors, including image entropy (higher entropy indicates richer texture and higher reliability). Point cloud density is statistically analyzed for lidar (higher density indicates higher reliability). Signal-to-noise ratio (SNR) is calculated for acoustic sensors; a SNR greater than 20dB is considered reliable. Weighted summation of visual feature maps, lidar depth maps, and acoustic 2D feature maps enhances the environmental adaptability of the fused features. Weighted fusion integrates multimodal features, avoiding the one-sidedness of single-sensor information and improving the completeness and recognizability of target features. Visual feature maps, lidar depth maps, and acoustic feature maps are normalized to the [0, 1] range and weighted fusion is performed to generate a unified multi-source fusion feature map, providing a highly reliable unified feature representation for subsequent target recognition and tracking.

[0052] The target recognition and tracking unit 3 uses a deep learning model to perform target detection and classification on the unified feature map, identifies the drone and distinguishes between cooperative and non-cooperative targets, and uses a Kalman filter module to combine the target dynamic model and the deep learning detection results to predict and update the drone's position, speed and acceleration.

[0053] In target recognition and tracking unit 3, the deep learning model performs target detection and classification on the unified feature map, including the following steps:

[0054] Traditional general-purpose deep learning models suffer from problems such as large parameter count, low computational efficiency, and insufficient classification accuracy in UAV target detection scenarios. Therefore, models optimized for small target detection are needed to meet practical requirements. An optimized deep learning model is used as the basic architecture. High-level features from fused data are extracted through convolutional layers. Since the unified feature map contains multi-source information, further discriminative high-level semantic features are extracted through convolutional layers to accurately distinguish target categories and locations. Twelve convolutional layers are set in the backbone network, using convolutional kernels of different sizes for parallel convolution operations. Each convolutional layer is followed by a normalization layer and an activation function. The convolutional layers continuously extract features through local receptive fields, gradually generating deeper layers such as the overall outline and structure of the UAV from shallow edge and texture features. Advanced semantic features can effectively separate target features, reduce false detection rates, improve the model's feature extraction efficiency, identify target locations, and classify them as cooperative or non-cooperative drones. Based on the extracted advanced features, it is necessary to accurately determine the target's location in the image and distinguish between cooperative and non-cooperative drones to provide a basis for subsequent tracking and decision-making. Two parallel output branches are set in the model head: one branch outputs the target bounding box coordinates and uses the mean squared error loss function for regression training; the other branch outputs the target class probability and uses the cross-entropy loss function for classification training. The non-maximum suppression algorithm removes duplicate detection boxes and retains the bounding box with the highest confidence as the final detection result, reducing missed detections and misclassifications, and providing reliable data support for subsequent drone management.

[0055] In the target recognition and tracking unit 3, the Kalman filter module combines the target dynamic model with the deep learning detection results to perform prediction and state update, including the following steps:

[0056] The motion of drones is continuous, requiring dynamic models to predict their state at the next moment. This compensates for the instantaneous limitations of deep learning detection and improves tracking continuity. A second-order motion model is used to define the state vector. ,in, Current position For velocity components, The acceleration components are used, and the state transition matrix is ​​calculated based on the state vector. Based on the target dynamic model, the system predicts the position, velocity, and acceleration at the next moment. It uses deep learning detection results as measurements to update the Kalman filter state estimate and uses the previous moment's state to predict the next moment's position, providing prior estimates for the deep learning detection results and enhancing tracking robustness. The prediction equation is as follows: , ;in To predict the state vector, This is the optimal estimated state from the previous moment. To predict the covariance matrix, The covariance matrix at the previous time step is... The process noise covariance matrix characterizes model uncertainty. In deep learning frame detection, it can continuously extrapolate the target state, avoiding tracking interruptions and improving tracking continuity. Deep learning detection provides the current target position observation, which needs to be fused with the predicted value to correct model bias and improve positioning accuracy. (Measurement vector) Take the center coordinates of the bounding box detected by deep learning The measurement matrix is Measure the noise covariance matrix Take a diagonal matrix, with elements representing the detection error variance. By fusing predicted and measured values, an optimal state estimate is generated, noise is suppressed, and tracking accuracy is improved. The Kalman gain is then calculated. Update state estimation Update the covariance matrix ;in, As the identity matrix, iterative calculations gradually converge to the optimal estimate. During human-machine maneuvers, the model error increases, requiring dynamic adjustment of process noise to adapt to changes in motion patterns. When the target acceleration change rate is detected... Automatically increase The diagonal elements, the formula is ; By employing a Kalman filter process of "dynamic model prediction - detection data correction - noise adaptive adjustment," the transient nature and noise defects of deep learning detection are resolved, enabling continuous and accurate tracking of UAV targets and providing reliable trajectory data support for low-altitude safety management.

[0057] In target identification and tracking unit 3, identifying non-cooperative drones includes the following steps:

[0058] Cooperative drones typically fly along pre-set paths, while non-cooperative drones often follow unknown or unauthorized paths. Path analysis is needed to identify anomalies, analyze whether the target's flight path is unknown or unauthorized, and calculate the curvature of the target's trajectory. Rate of change of velocity and distance from the no-fly zone The formula is ; ; ;in, and The first and second derivatives of the trajectory coordinates, Design a path validity function for the current speed. ,when At the same time, it determines abnormal paths, improving the detection accuracy of abnormal paths such as sudden turns and incursions into no-fly zones. It detects whether signal characteristics are abnormal, and combines classification results and tracking information to comprehensively determine non-cooperative status. Non-cooperative drones may be using abnormal communication protocols or have no signal transmission. It is necessary to identify suspicious targets through signal analysis, collect the target's radio frequency signal, and extract the carrier frequency. Modulation method Signal strength and spectral characteristics Calculate the cosine similarity between the signal features and the known cooperative drone roster. ,when or signal strength When determining signal anomalies, relying on a single feature can easily lead to false alarms. It is necessary to integrate path, signal, and classification information to improve decision reliability and incorporate a comprehensive path score. Signal scoring and classification confidence The weighted fusion is performed to obtain the final weighted fusion value. ,but An alarm is triggered only when a drone is identified as non-cooperative after three or more consecutive frames, thus suppressing momentary misjudgments, improving the ability to identify non-cooperative drones, overcoming the limitations of single-feature judgment, and providing a reliable intelligent monitoring method for low-altitude security.

[0059] The decision response unit 4 determines non-cooperative drones based on the classification results and tracking information, and triggers an alarm signal when a non-cooperative drone is identified. At the same time, it outputs the target position, speed and category information to the ground station display.

[0060] In decision response unit 4, triggering an alarm signal includes the following steps:

[0061] Single alarm methods are susceptible to environmental interference, while multimodal alarms can improve alarm reach and ensure timely emergency response. When a non-cooperative drone is identified, at least one of the following alarms is activated: sound alarm, light alarm, or wireless signal alarm. Simultaneously, the alarm information is transmitted to the ground station via a wireless channel. The ground station outputs target location, speed, and category information, including the following steps:

[0062] The system generates target status data packets in real time and transmits them to the ground station display via a communication interface. The ground station interface dynamically displays the position coordinates, velocity vector, and target category identifier, solving the problem of insufficient response of traditional alarm systems in complex environments and providing timely and comprehensive decision support for low-altitude safety management.

[0063] This invention uses sensor unit 1 to simultaneously acquire images, 3D point clouds, and acoustic signals. Data processing and fusion unit 2 performs multi-source data alignment and weighted fusion to extract texture, shape, and color features from the images. The point cloud is used to generate a depth map, and the acoustic signals are mapped to a 2D feature map. Based on sensor reliability, weights are dynamically allocated to generate a unified feature map. Target recognition and tracking unit 3 uses an optimized deep learning model to identify the type of UAV and combines Kalman filtering to achieve accurate prediction and updating of position and velocity status. Decision response unit 4 triggers real-time audible and visual alarms and wireless alarms for non-cooperative UAVs and transmits target dynamic information to the ground station, improving the target detection accuracy of UAVs in complex environments.

[0064] The second objective of this invention is to provide a method for implementing a drone detection system based on vision, lidar, and sound waves, comprising the following steps:

[0065] S1. The device collects image data, 3D point cloud data, and sound signals around the drone using visual sensors, lidar, and acoustic sensors.

[0066] S2. Perform Gaussian filtering on the image data to remove noise, normalize the pixel values ​​to the range of [0,1], extract edge, texture and color features, perform statistical filtering on the point cloud data to remove noise, and project the three-dimensional point cloud onto the two-dimensional image plane to generate a depth map. Perform frequency domain filtering on the acoustic signal to remove noise, extract the time domain amplitude, frequency domain spectrum and cepstral coefficient features, and map them into a two-dimensional matrix feature map.

[0067] S3. Based on the dynamic allocation of weights according to the reliability of each sensor, the visual feature map, depth map feature map and acoustic two-dimensional feature map are weighted and summed to generate a unified multi-source fusion feature map.

[0068] S4. Input the fused feature map into the optimized deep learning model, extract high-level features through convolutional layers to identify the target location, and classify it into cooperative and non-cooperative drones. Based on the target dynamic model, predict the position, velocity and acceleration. Input the deep learning detection results into the Kalman filter module to update the target state estimate.

[0069] S5 analyzes whether the target's flight path is abnormal and whether the signal characteristics are unauthorized. Combining the classification results and tracking information, it determines the non-cooperative drone, triggers an alarm, and generates a target status data packet in real time. Finally, it transmits the data to the ground station display via the communication interface to dynamically display the target's real-time status.

[0070] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A drone detection system based on vision, lidar, and acoustic waves, characterized in that, include: The sensor unit (1) includes a visual sensor, a lidar and an acoustic sensor, which are used to collect image data, three-dimensional point cloud data and acoustic signals around the UAV, respectively. The data processing fusion unit (2) extracts the texture, shape and color features of the image data through the data preprocessing module, projects the point cloud data onto the image plane to generate a depth map, extracts the time domain, frequency domain and cepstral features of the acoustic signal, and maps them to a two-dimensional feature map for multi-source data alignment. Then, the feature fusion module performs weighted fusion of the preprocessed visual, lidar and acoustic features to generate a unified feature map. The data processing and fusion unit (2) extracts time-domain, frequency-domain, and cepstral features of the acoustic signal and maps them to a two-dimensional feature map, including the following steps: Noise is removed by frequency domain filtering or wavelet transform, and the time-domain amplitude features, frequency-domain spectrum features, and cepstral coefficient features of the acoustic signal are extracted. The features are then mapped into a two-dimensional matrix form to align with visual and lidar data. In the data processing fusion unit (2), the feature fusion module performs weighted fusion including the following steps: Based on the reliability of different sensors, weights are dynamically allocated, and the visual feature map, lidar depth map feature map, and acoustic two-dimensional feature map are weighted and summed to generate a unified multi-source fusion feature map. The target recognition and tracking unit (3) performs target detection and classification on the unified feature map through a deep learning model, identifies the UAV and distinguishes between cooperative and non-cooperative targets, and predicts and updates the position, speed and acceleration of the UAV by combining the target dynamic model and the deep learning detection results through the Kalman filter module. The decision response unit (4) judges non-cooperative drones based on the classification results and tracking information, and triggers an alarm signal when a non-cooperative drone is identified, while outputting the target position, speed and category information to the ground station display.

2. The UAV detection system based on vision, lidar, and sound waves according to claim 1, characterized in that, The data processing and fusion unit (2) extracts texture, shape, and color features from image data, including the following steps: By removing noise through Gaussian filtering, the image pixel values ​​are normalized to the range of [0,1], and edge features, texture features, and color features are extracted.

3. The UAV detection system based on vision, lidar, and sound waves according to claim 1, characterized in that, The data processing and fusion unit (2) projects point cloud data onto an image plane to generate a depth map, which includes the following steps: By removing noise from the point cloud through statistical filtering, the 3D point cloud is projected onto the 2D image plane to generate a 2D depth map containing depth information.

4. The UAV detection system based on vision, lidar, and sound waves according to claim 1, characterized in that, In the target recognition and tracking unit (3), the deep learning model performs target detection and classification on the unified feature map, including the following steps: An optimized deep learning model is used as the basic architecture. High-level features of the fused data are extracted through convolutional layers to identify target locations and classify them as cooperative or non-cooperative drones.

5. The UAV detection system based on vision, lidar, and sound waves according to claim 1, characterized in that, In the target recognition and tracking unit (3), the Kalman filter module combines the target dynamic model and deep learning detection results to perform prediction and state update, including the following steps: The target dynamic model predicts the position, velocity, and acceleration at the next moment, and the deep learning detection results are used as measurements to update the Kalman filter state estimate.

6. The UAV detection system based on vision, lidar, and sound waves according to claim 1, characterized in that, The target identification and tracking unit (3) identifies non-cooperative drones by including the following steps: Analyze whether the target's flight path is unknown or unauthorized, detect whether the signal characteristics are abnormal, and combine the classification results and tracking information to make a comprehensive judgment on the non-cooperative status.

7. The UAV detection system based on vision, lidar, and sound waves according to claim 1, characterized in that, In the decision response unit (4), triggering the alarm signal includes the following steps: When a non-cooperative drone is identified, at least one of the following is activated: an audible alarm, a visual alarm, or a wireless signal alarm. Simultaneously, the alarm information is transmitted wirelessly to the ground station. The ground station outputs the target's location, speed, and category information, including the following steps: Target status data packets are generated in real time and transmitted to the ground station display via the communication interface. The position coordinates, velocity vector, and target category identifier are dynamically displayed on the ground station interface.

8. A method for implementing a UAV detection system based on vision, lidar, and sound waves as described in any one of claims 1-7, characterized in that: Includes the following steps: S1. The device collects image data, 3D point cloud data, and sound signals around the drone using visual sensors, lidar, and acoustic sensors. S2. Perform Gaussian filtering on the image data to remove noise, normalize the pixel values ​​to the range of [0,1], extract edge, texture and color features, perform statistical filtering on the point cloud data to remove noise, and project the three-dimensional point cloud onto the two-dimensional image plane to generate a depth map. Perform frequency domain filtering on the acoustic signal to remove noise, extract the time domain amplitude, frequency domain spectrum and cepstral coefficient features, and map them into a two-dimensional matrix feature map. S3. Based on the dynamic allocation of weights according to the reliability of each sensor, the visual feature map, depth map feature map and acoustic two-dimensional feature map are weighted and summed to generate a unified multi-source fusion feature map. S4. Input the fused feature map into the optimized deep learning model, extract high-level features through convolutional layers to identify the target location, and classify it into cooperative and non-cooperative drones. Based on the target dynamic model, predict the position, velocity and acceleration. Input the deep learning detection results into the Kalman filter module to update the target state estimate. S5 analyzes whether the target's flight path is abnormal and whether the signal characteristics are unauthorized. Combining the classification results and tracking information, it determines the non-cooperative drone, triggers an alarm, and generates a target status data packet in real time. Finally, it transmits the data to the ground station display via the communication interface to dynamically display the target's real-time status.

Images