Animal target detection system and method based on airborne radar and intelligent identification

Abstract

The invention provides an animal target detection system and method based on airborne radar and intelligent identification, and belongs to the technical field of radar detection. The system comprises an unmanned aerial vehicle, an antenna array fused with a fuselage of the unmanned aerial vehicle, and a radar host. The radar host comprises a signal processing unit; the antenna array is designed to be of a deformable structure and can be attached to the surface of a fuselage of the unmanned aerial vehicle. The antenna array is used for sending a radio frequency signal of a terahertz frequency band to a target area under the control of the system and receiving an echo signal reflected by an obstacle; and the signal processing unit is used for receiving the echo signals, extracting multi-dimensional features from the echo signals by using a spatio-temporal feature extraction network based on an attention mechanism, and classifying the multi-dimensional features through a deep reinforcement learning algorithm to obtain animal categories. The airborne radar detection method is light in weight and low in cost, the efficiency and precision of biodiversity observation can be improved, and the monitoring cost is greatly reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar detection, and particularly relates to an animal target detection system and method based on airborne radar and intelligent recognition. Background Art

[0002] Biodiversity observation is a key means to evaluate the health status of ecosystems, monitor changes in species distribution, and guide environmental protection. However, the current mainstream biodiversity observation methods mainly rely on ground-based manual surveys and satellite remote sensing technology, and these traditional methods are difficult to meet the needs of modern ecological monitoring. Among them, although ground-based manual surveys can obtain detailed ecological data, the survey efficiency is severely limited by terrain conditions and weather factors, and a large amount of human resources need to be invested. The observation cost is high, and it is difficult to achieve continuous monitoring of large-scale areas. There are also safety risks in surveys in remote or dangerous areas.

[0003] Although satellite remote sensing technology has expanded the observation range, its spatial resolution is difficult to meet the identification needs of individual organisms, and the time resolution is limited, making it difficult to capture the dynamic changes of biological activities. At the same time, the quality of remote sensing data is easily affected by atmospheric conditions and surface occlusion, and it lacks the ability to detect underground biological activities, making it difficult to meet the needs of refined research. Existing airborne radar systems also have limitations in biological target detection. The traditional antenna structure is difficult to adapt to complex terrain environments, the signal processing algorithm has insufficient recognition accuracy for dynamic targets, and the system's anti-interference ability is limited.

[0004] In view of the many deficiencies of the existing technology, it is urgent to develop a new type of biodiversity observation system that can overcome the limitations of traditional methods and achieve high-precision, low-cost real-time monitoring. Especially in complex forest environments, the system needs to have advanced functions such as high-resolution imaging, intelligent target recognition, and environmental adaptability to meet the urgent needs of modern ecological monitoring. This not only requires a breakthrough in hardware design, but also needs to innovate in signal processing and intelligent recognition algorithms, so as to establish a comprehensive biodiversity monitoring solution. Summary of the Invention

[0005] The present invention provides an animal target detection system and method based on airborne radar and intelligent recognition. By integrating a lightweight antenna array, an embedded radar host, and an edge intelligent processing module, it realizes real-time detection and classification recognition of animal targets in the wild environment, and has the advantages of scalability, low cost, high efficiency, etc., and is suitable for wide deployment in ecological environment monitoring tasks.

[0006] The present invention provides an animal target detection system based on airborne radar and intelligent recognition. The system includes a drone, an antenna array integrated with the drone fuselage, and a radar host; wherein, the radar host includes a signal processing unit. The antenna array is designed as a deformable structure and can conform to the surface of the fuselage of the drone.

[0007] The antenna array is used to send radio frequency signals to a target area under system control and receive echo signals reflected by obstacles. The signal processing unit is used to receive the echo signals, extract multi-dimensional features from the echo signals by using a spatio-temporal feature extraction network based on an attention mechanism, and classify the multi-dimensional features through a deep reinforcement learning algorithm to obtain the animal category.

[0008] According to an animal target detection system based on an airborne radar and intelligent recognition provided by the present invention, the radar host can adaptively adjust the radar beam direction and power distribution through an intelligent control module to achieve efficient detection in a complex environment.

[0009] According to an animal target detection system based on an airborne radar and intelligent recognition provided by the present invention, the antenna array adopts an ultra-thin conformable array structure constructed of flexible materials and can deform and attach according to the curvature of the drone fuselage, thereby reducing aerodynamic drag and improving the space utilization rate of radar deployment; the array is adapted to operate in the millimeter wave band and supports high-resolution far-field beamforming.

[0010] According to an animal target detection system based on an airborne radar and intelligent recognition provided by the present invention, the radar host includes a signal processing unit composed of an embedded processor and an AI acceleration chip.

[0011] According to an animal target detection system based on an airborne radar and intelligent recognition provided by the present invention, the system constructs a radar perception model based on a multi-level ecological environment, including: modeling the millimeter wave reflection characteristics of the canopy layer, adaptively modeling the scattering characteristics of the understory vegetation layer, and a ground clutter suppression module combined with a deep learning algorithm to improve the animal target resolution ability in a complex environment.

[0012] According to an animal target detection system based on an airborne radar and intelligent recognition provided by the present invention, the feedback control mechanism of the system includes a state estimator enhanced by deep learning, a path optimization algorithm based on reinforcement learning, and a terahertz beam adaptive regulation unit, and realizes parameter optimization through distributed edge computing to ensure the stability of the system in a complex electromagnetic environment.

[0013] According to an animal target detection system based on an airborne radar and intelligent recognition provided by the present invention, the system adopts an enhanced Kalman filtering algorithm based on deep learning, realizes super-resolution target tracking through distributed MIMO, and adjusts system parameters such as flight state optimization, terahertz beam regulation, and gain dynamic matching by using an adaptive parameter estimation technique.

[0014] The present invention also provides an animal target detection method based on airborne radar and intelligent recognition, which is applied to the signal processing unit in any of the above embodiments of the animal target detection system based on airborne radar and intelligent recognition. The method includes: Receiving an echo signal; Inputting the echo signal into a spatio-temporal feature extraction network based on an attention mechanism to obtain multi-dimensional features extracted by the spatio-temporal feature extraction network based on the attention mechanism from the echo signal; Using a deep reinforcement learning algorithm to classify the multi-dimensional features to obtain the animal category.

[0015] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and running on the processor. When the processor executes the computer program, the animal target detection method based on airborne radar and intelligent recognition is implemented.

[0016] The present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the animal target detection method based on airborne radar and intelligent recognition is implemented.

[0017] The present invention also provides a computer program product, including a computer program. When the computer program is executed by a processor, the animal target detection method based on airborne radar and intelligent recognition is implemented.

[0018] The animal target detection system and method based on airborne radar and intelligent recognition provided by the present invention includes a drone, an antenna array integrated with the drone fuselage, and a radar host. After the system is started, the transmitter of the radar host generates a high-frequency electromagnetic wave signal in the terahertz frequency band, which is transmitted by the antenna array attached to the fuselage surface to form a directional beam. During the propagation of these electromagnetic waves in space, after encountering a target, reflections occur, forming an echo signal containing target feature information. In addition, the signal processing unit in the radar host uses a spatio-temporal feature extraction network based on an attention mechanism to extract multi-dimensional features from the echo signal, and classifies the multi-dimensional features through a deep reinforcement learning algorithm to obtain the animal category. This method does not rely on satellite remote sensing technology, only relies on drone and radar technologies, can reduce costs, and uses electromagnetic wave signals in the terahertz frequency band, which can improve the resolution and the detection accuracy of biodiversity. This innovative technical solution breaks through the limitations of traditional airborne radar systems and realizes high-precision detection and recognition of wild animals in complex forest environments. By integrating a number of cutting-edge technologies, this system not only significantly improves the efficiency and accuracy of biodiversity observation, but also greatly reduces the monitoring cost, providing strong technical support for ecological environment protection and biodiversity research.

[0019] Furthermore, the receiving antenna is designed with a metamaterial structure, which can precisely manipulate electromagnetic waves at the sub-wavelength scale, significantly improving the spatial resolution of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0021] Figure 1 It is a schematic diagram of the composition structure of the animal target detection system based on airborne radar and intelligent recognition provided by the present invention.

[0022] Figure 2 It is a schematic diagram of the radar host integrated with the fuselage provided by the present invention.

[0023] Figure 3 It is a schematic diagram of the main modules of the radar host provided by the present invention.

[0024] Figure 4 It is a schematic diagram of the flow of the animal target detection method based on airborne radar and intelligent recognition provided by the present invention.

[0025] Figure 5 It is a schematic diagram of the structure of the animal target detection system based on airborne radar and intelligent recognition provided by the present invention.

[0026] Figure 6 It is a schematic diagram of the structure of the animal target detection device based on airborne radar and intelligent recognition provided by the present invention.

[0027] Figure 7 It is a schematic diagram of the structure of the electronic device provided by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0028] To make the objectives, technical solutions, and advantages of the present invention clearer, the following will clearly and completely describe the technical solutions of the present invention with reference to the accompanying drawings. The embodiments described herein are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0029] The following will describe the specific embodiments of the present invention in conjunction with Figures 1-5 Describe the specific embodiments of the present invention.

[0030] The present invention provides an animal target detection system based on airborne radar and intelligent recognition, as Figure 1 shownFigure 1 shows a schematic diagram of the composition structure of an animal target detection system based on airborne radar and intelligent recognition, including a drone 101, an antenna array 102 integrated with the drone fuselage, and a radar host, such as Figure 2 shown Figure 2 shows a schematic diagram of the radar host integrated with the fuselage. The radar host includes a host shell 201, which is used to provide signal I / O interfaces and a heat dissipation surface, connect the signal processing module and the antenna frame module, and ensure that the overall temperature of the whole machine remains within a reasonable range during real-time signal transmission and reception; a transmitter 202, which is used to provide a high-power radio frequency signal with a carrier modulated by a specific modulation for the radar, and make the radio frequency signal radiate out by the antenna through the feeder and the transceiver switch; a receiver 203, which is used to preselect, amplify, filter, and demodulate the target echo signal accompanied by noise, clutter, and interference received by the antenna; other components, which are used to filter signals, remove clutter, perform spectrum analysis, power calculation, etc. The radar host is integrated with the fuselage to make the overall meet the aerodynamic shape, eliminate the influence of external equipment on the aircraft, and improve flight stability and control accuracy.

[0031] Among them, the antenna array 102 integrated with the drone fuselage can be a lightweight conforming antenna array.

[0032] The radar host includes a signal processing unit, a transmitter, and a receiver; the above antenna array 102 is composed of a flexible dielectric substrate and a conforming microstrip antenna, and has characteristics such as bendability, low profile, and high integration, and can effectively conform to the contour of the drone fuselage, improve space utilization rate and reduce aerodynamic drag. The system adopts an integrated design of a special-shaped antenna frame and the aircraft, and the antenna array is integrated with the wing structure, and the power supply and data transmission systems of the aircraft are reused, effectively reducing the hardware cost.

[0033] such as Figure 1 shown, the antenna array 102 includes a transmitting antenna and a receiving antenna. The transmitting antenna is arranged on the surface of the drone wing or in the embedded structure, and the receiving antenna is arranged perpendicular to the wing direction and parallel to the central axis of the drone fuselage to enhance the directivity and angular resolution of the radar system. The antenna array operates in the terahertz band or the millimeter wave band, supports multi-channel transmission and reception, and can achieve high-precision beamforming and target perception.

[0034] Combined with Figure 3 shown Figure 3 shows a schematic diagram of the main modules of the radar host; the radar host includes a signal processing unit, a transmitter, and a receiver.

[0035] Specifically, the transmitter in the above radar host emits a modulated electromagnetic wave signal. This electromagnetic wave signal can be, for example, an electromagnetic wave signal in the terahertz frequency band. The terahertz frequency band is a special frequency band between microwaves and infrared rays in the electromagnetic spectrum, with a frequency range of 0.1 THz to 10 THz (corresponding wavelength range of 3 mm to 30 μm). It is the transition region between millimeter waves (frequency range of 30 - 300 GHz) and light waves. The terahertz band has a wider bandwidth and higher resolution. After the electromagnetic wave signal in the terahertz frequency band is transmitted to the transmitting antenna, it can send a directional beam to the target area, and the receiving antenna is used to receive the echo signal reflected by the obstacle.

[0036] The signal processing unit receives the echo signal reflected by the above obstacle, extracts multi-dimensional features from the echo signal using a spatio-temporal feature extraction network based on the attention mechanism, and classifies the multi-dimensional features through a deep reinforcement learning algorithm to obtain the animal category.

[0037] Among them, the spatio-temporal feature extraction network based on the attention mechanism is a network architecture that, when processing data with spatio-temporal characteristics (such as videos, time series sensor data, etc.), uses the attention mechanism to enhance the ability to extract important spatio-temporal features. Its network structure mainly includes an input layer, a spatio-temporal feature extraction layer (usually implemented by CNN, RNN, or LSTM), an attention layer, a fusion and decision layer. The attention layer introduces the attention mechanism, enabling the network to automatically learn the importance weights of different spatio-temporal positions in the data and perform weighted fusion on the features of these positions. A common implementation method of the attention layer is the self-attention mechanism.

[0038] The deep reinforcement learning algorithm is a branch of machine learning. Its core idea is to let an agent learn how to take actions through interactions with the environment to maximize the long-term cumulative reward. It mimics the process by which humans or animals learn to make decisions through trial and error. MDP (Markov Decision Process) is an important mathematical model in the deep reinforcement learning algorithm. It consists of a state set, an action set, a state transition probability, and a reward function. The state transition function describes the probability of transitioning to the next state given the current state and action. The reward function defines the immediate reward obtained by taking a certain action in a certain state.

[0039] Specifically, the signal processing unit first performs preprocessing on the echo signal, including the following steps: Step 1: Data preprocessing: Step 1.1: Perform analog-to-digital conversion on the echo signal to generate a time-domain discrete sequence.

[0040] Step 1.2: Use the wavelet threshold denoising algorithm to eliminate environmental noise. Specifically, perform 3-layer decomposition using the Daubechies4 wavelet basis, and set the soft threshold function to process the high-frequency coefficients.

[0041] Step 1.3: Segment the signal through a sliding window to generate time series signal segments.

[0042] Step 2: Construct a spatio-temporal feature extraction network based on the attention mechanism. This spatio-temporal feature extraction network based on the attention mechanism includes a dual-branch feature extraction layer, a multi-head self-attention layer, and a multi-dimensional feature fusion layer; the dual-branch feature extraction layer includes a spatial branch and a temporal branch; the spatio-temporal feature extraction network based on the attention mechanism is specifically used for: inputting the echo signal into the spatial branch and the temporal branch in the dual-branch feature extraction layer respectively to obtain the shallow texture features output by the spatial branch and the temporal features output by the temporal branch; inputting the shallow texture features and the temporal features into the multi-head self-attention layer to enable the multi-head self-attention layer to perform convolution operations on the shallow texture features and the temporal features to obtain an attention weight matrix, and performing Hadamard product operation on the attention weight matrix and the temporal features to obtain an enhanced feature vector output by the multi-head self-attention layer; inputting the enhanced feature vector and the shallow texture features into the multi-dimensional feature fusion layer to enable the multi-dimensional feature fusion layer to splice the enhanced feature vector and the shallow texture features to obtain the multi-dimensional features output by the multi-dimensional feature fusion layer. Specifically, it includes the following steps: Step 2.1: Construct a dual-branch feature extraction layer: The dual-branch feature extraction layer includes a spatial branch and a temporal branch, where Spatial branch: Use 5 layers of 1D-CNN (Convolutional Neural Network, the convolutional kernel size is [7, 5, 3, 3, 3], and the number of channels is [32, 64, 128, 256, 512]), and each layer is followed by batch normalization and ReLU activation. This spatial branch is used to extract shallow texture features.

[0043] Temporal branch: Use a bidirectional LSTM (Long Short-Term Memory, the number of hidden layer units is 256) to handle the time dimension dependence. This spatial branch is used to extract temporal features.

[0044] Step 2.2: Embed a multi-head self-attention layer (4 attention heads) to calculate the convolution results obtained from the above spatial branch and temporal branch to obtain an attention weight matrix.

[0045] Step 2.3: The feature fusion layer performs Hadamard product on the output of the above temporal branch and the attention weight matrix to generate an enhanced feature vector.

[0046] Step 3: Multi-dimensional feature fusion layer, including the following specific steps: Step 3.1: Hierarchical fusion of spatio-temporal features: Primary features: The original time-frequency features (referring to the spectral centroid and bandwidth extracted from the echo signal through STFT).

[0047] Intermediate features: The shallow texture features extracted by the above-mentioned spatial branch CNN.

[0048] Advanced features: The enhanced feature vectors extracted by the above-mentioned temporal branch LSTM-attention mechanism.

[0049] Step 3.2: Adopt the feature concatenation method to fuse the above multi-dimensional features (including primary features, intermediate features, and advanced features) to form a 4096-dimensional feature vector.

[0050] Step 3.3: Perform feature dimensionality reduction through a learnable parameter matrix (such as using the PCA-Whitening algorithm) to finally generate a 512-dimensional optimized feature vector.

[0051] Step 4: Design of a deep reinforcement learning classifier (capable of running a deep reinforcement learning algorithm), the specific steps are as follows: Step 4.1: Define the Markov decision process: State space S: Obtain the normalized values of the 512-dimensional optimized feature vector; Action space A: Perform threshold judgment on the above normalized values to obtain the animal category, for example: {felidae, canidae, ungulates, birds, others} five categories of animals.

[0052] Step 5: Phased training strategy: Use the animal radar echo database (containing 100,000 labeled samples) to train the classifier.

[0053] The above embodiments provide an animal target detection system based on airborne radar and intelligent recognition. The system includes a drone, an antenna array integrated with the drone fuselage, and a radar host. The system adopts an integrated design of a special-shaped antenna frame and the aircraft, where the antenna array is integrated with the wing structure, and the power supply and data transmission systems of the aircraft are reused, effectively reducing the hardware cost. When the system is started, the transmitter of the radar host generates high-frequency electromagnetic wave signals in the terahertz frequency band, which are transmitted through the antenna array integrated with the fuselage to form a directional beam. During the propagation of these electromagnetic waves in space, when they encounter a target, reflections occur, forming echo signals containing target feature information. In addition, the signal processing unit in the radar host uses a spatio-temporal feature extraction network based on the attention mechanism to extract multi-dimensional features from the echo signals, and classifies the multi-dimensional features through a deep reinforcement learning algorithm (such as MDP, Markov Decision Process, Markov decision process) to obtain the animal category. This method does not rely on satellite remote sensing technology, only relies on drone and radar technologies, can reduce costs, and uses electromagnetic wave signals in the terahertz frequency band, which can improve the resolution and the detection accuracy of biodiversity.

[0054] In one embodiment, the above antenna array adopts a metasurface structure design to achieve electromagnetic wave manipulation at the sub-wavelength scale.

[0055] Among them, the metasurface structure refers to a material with a two-dimensional planar structure, and the basic unit of the constituent material is at the sub-wavelength scale, so it can achieve electromagnetic wave manipulation at the sub-wavelength scale.

[0056] Specifically, the above receiving antenna is made of a metasurface structure material. The metasurface structure material is an artificial electromagnetic material composed of periodically or non-periodically arranged or implanted basic units with special-designed sub-wavelength sizes into the basic material body, and has physical properties that do not exist in natural materials, and can regulate electromagnetic waves. For example, it can be made of graphene material.

[0057] The above embodiments adopt a receiving antenna with a metasurface structure, so it can accurately manipulate electromagnetic waves at the sub-wavelength scale, significantly improving the spatial resolution ability of the system.

[0058] In one embodiment, the above signal processing unit is composed of a neuromorphic chip and a three-dimensional photonic integrated circuit processor.

[0059] Specifically, to match the signal processing capabilities of the above antenna array, the signal processing unit uses a three-dimensional photonic integrated circuit processor to implement signal processing. The three-dimensional photonic integrated circuit processor is an advanced chip technology that achieves optoelectronic co-integration by vertically stacking multiple layers of photonic and electronic devices. Its core role is to break through the performance bottleneck of traditional electronic chips by leveraging the high-bandwidth and low-latency characteristics of optical signals, supporting THz-level bandwidth.

[0060] In addition, the above spatio-temporal feature extraction network based on the attention mechanism and the deep reinforcement learning classifier can run on a neuromorphic chip. A neuromorphic chip is an integrated circuit that mimics the structure and information processing mechanism of biological neurons, aiming to imitate the human brain's neural network at the hardware level to achieve efficient and low-power intelligent computing. The neuromorphic chip consists of an artificial neuron array, synaptic weight storage (such as memristors), and a spiking neural network.

[0061] In one embodiment, the feedback control mechanism of the system includes a deep learning-enhanced state estimator, a reinforcement learning-based path optimization algorithm, and a terahertz beam adaptive regulation unit, which cooperate with distributed edge computing nodes to perform real-time optimization management of various system parameters, ensuring the stability and response ability of the system in a complex electromagnetic environment. The various system parameters include antenna system parameters (such as antenna gain, beam width, beam pointing, etc.), transmitted signal power, signal bandwidth, pulse repetition frequency, radar system performance parameters (such as detection probability, false alarm rate, operating range, etc.), data transmission and communication parameters (such as data transmission rate, communication link quality, etc.), as well as the fractal speed and altitude of the UAV, and the attitude stability of the UAV.

[0062] Among them, the deep learning-enhanced state estimator is a system that combines deep learning technology with traditional state estimation methods, used to improve the accuracy, robustness, and adaptability of state estimation. State estimation is the process of inferring the internal state of a system based on the observable data of the system using mathematical models and algorithms. The deep learning-enhanced state estimator incorporates deep learning algorithms on the basis of traditional state estimation, leveraging the powerful feature extraction and learning capabilities of deep learning to estimate the state of the system more accurately. In this embodiment, it is mainly used to accurately estimate the state of the UAV, such as attitude, position, speed, etc., to ensure flight safety.

[0063] The path optimization algorithm based on reinforcement learning is a type of algorithm that uses the principles of reinforcement learning to solve path planning and optimization problems. It mainly realizes path optimization by the interaction between the agent and the environment to learn the optimal strategy. In this embodiment, it is applied to the path planning of unmanned aerial vehicles (UAVs), helping the UAVs to plan the optimal path in a complex forest environment, avoid obstacles, and reach the target position efficiently. The path optimization algorithm based on reinforcement learning adaptively adjusts the flight path and radar scanning strategy of the UAV according to the environmental feedback, realizing continuous and stable detection of animal targets in complex dynamic scenarios.

[0064] The terahertz beam adaptive control unit is a system based on adaptive beamforming technology. By performing real-time processing on the output signals of the antenna array or sensor array, it dynamically adjusts the direction and shape of the beam to enhance the desired signal and suppress the interference signal. Its working principle is as follows: The antenna array receives signals from different directions, including the desired signal and interference signals, and preprocesses the collected signals, such as filtering, amplification, etc., to remove noise and interference; calculates the covariance matrix of the received signals, and calculates the weight vector of the adaptive array according to the covariance matrix, which is used to adjust the direction and shape of the beam, and applies the calculated weight vector to the output signals of the antenna array to form a beam in the desired direction while suppressing the signals in the interference direction.

[0065] In this embodiment, by adaptively adjusting the direction and shape of the terahertz beam, the accuracy and resolution of target detection can be improved, the anti-interference ability of the radar can be enhanced, and the target can be effectively detected and tracked.

[0066] In one embodiment, the above-mentioned deep learning enhanced state estimator specifically includes: adopting the enhanced Kalman filtering algorithm and fusing a deep neural network for state estimation, and realizing high-precision tracking of the target through a distributed MIMO radar architecture. At the same time, an adaptive parameter estimation module is also introduced (see the above description of the feedback control mechanism) to dynamically adjust the flight attitude of the flight platform, the direction of the terahertz beam, and the system gain to optimize the radar imaging performance and target perception ability.

[0067] Specifically, in the field of radar detection, the movement of animal targets is usually continuous and constantly changing. By using the enhanced Kalman filtering algorithm to estimate and update the motion state of the target in real time, the future position and speed of the animal target can be predicted more accurately. The deep learning enhanced state estimator in this embodiment fuses the enhanced Kalman filtering algorithm and a deep learning model, jointly models the historical trajectory and real-time observation data of the animal target, predicts the future position and speed of the animal target through feature learning, and improves the spatial resolution and robustness of target tracking under the collaborative perception of the combined distributed MIMO antenna array.

[0068] Furthermore, the signal processing unit of the system also includes a deep learning module, which incorporates a variety of data augmentation strategies to simulate the radar echo characteristics under different weather and terrain conditions, and constructs a database of radar scattering characteristics of wild animals. This module uses a multi-scale feature fusion network, integrating spatial attention and temporal attention mechanisms, to extract key feature information from the scattering characteristics and motion patterns of animal targets. Through transfer learning strategies, the recognition ability under small sample conditions is effectively improved, and the recognition accuracy of wild animals can reach over 92%.

[0069] In one embodiment, to enhance the adaptability of the system in different ecological environments, a hierarchical environmental modeling mechanism is designed in the signal processing flow.

[0070] It should be noted that in the field of radar detection, tracking targets hidden in vegetation is extremely challenging, and vegetation occlusion easily interferes with radar signals. First, by modeling the electromagnetic scattering characteristics of vegetation, the target echo can be more accurately distinguished from the vegetation echo, thereby improving the detection accuracy of targets hidden in or interacting with vegetation (such as ground vehicles, low-altitude aircraft, wild animals, etc.); second, the vegetation structure is complex and diverse, and its scattered echo may generate false targets or clutter in the radar received signal. Precise vegetation modeling helps to identify these false echoes, reduce the false alarm rate, and enable the radar to more reliably detect real targets; third, different vegetation structures have different scattering and attenuation effects on radar signals. Based on the vegetation structure model, the distribution, intensity, and variation law of vegetation clutter can be more accurately estimated and predicted, so as to design more effective clutter suppression algorithms and improve the performance of radar signal processing.

[0071] This application adopts multi-level environmental modeling, including a canopy layer electromagnetic response model designed based on optimized electromagnetic characteristics, an adaptive model describing the scattering characteristics of the understory vegetation layer, and a model enhanced by deep learning to suppress clutter in the surface layer.

[0072] Specifically, the received echo signal is input into the data acquisition and preprocessing module for calibration, fusion, and cleaning; then the preprocessed data is respectively input into the vegetation modeling module and the animal target modeling module to construct the vegetation model of the target area for subsequent differentiation from the animal target signal; finally, it is input into the model tracking and optimization module, which estimates the target state in real time according to the target motion model and radar observation data, and predicts the target position and trajectory. Finally, the optimization module dynamically adjusts the parameters of the vegetation model and the target motion model according to the evaluation of the tracking results.

[0073] This mechanism uses radar data obtained by an antenna array to model the electromagnetic responses of different vegetation structures (canopy layer, understory vegetation layer, surface layer), and cooperates with a deep neural network to model and suppress clutter distribution. The system introduces lightweight wave-absorbing materials and array optimization technologies to reduce the false alarm rate caused by environmental clutter. By integrating a polarization feature pyramid network and a context-aware spatio-temporal attention mechanism, the system can effectively handle multi-source interference and achieve accurate identification and classification of animal targets of different sizes and species, with an overall classification accuracy exceeding 92%.

[0074] By comparing with a preset feature library, the signal processing unit can also determine the type and category of the target (such as large mammals, medium mammals, small mammals, etc.), and provide further analysis based on the attribute information of the target. The realization of all these functions not only improves the accuracy of target positioning and tracking, but also enables the system to maintain high stability and accuracy under complex environmental conditions, thus more effectively completing the animal target detection task.

[0075] In one embodiment, as Figure 4 shown Figure 4 is a schematic flow chart of an animal target detection method based on airborne radar and intelligent recognition. This method is applied to the signal processing unit in the above-mentioned animal target detection system based on airborne radar and intelligent recognition, and this method includes the following steps.

[0076] Step 401, receive the echo signal.

[0077] Among them, the echo signal is the signal after the target radio frequency signal sent by the radar is reflected by the obstacle.

[0078] Specifically, use the receiving antenna to receive the echo signal reflected by the obstacle.

[0079] Step 402, input the echo signal into a spatio-temporal feature extraction network based on an attention mechanism, and obtain multi-dimensional features extracted by the spatio-temporal feature extraction network based on the attention mechanism from the echo signal.

[0080] Specifically, the signal processing unit receives the echo signal reflected by the above-mentioned obstacle, extracts multi-dimensional features from the echo signal by using a spatio-temporal feature extraction network based on an attention mechanism, and classifies the multi-dimensional features through a deep reinforcement learning algorithm to obtain the animal category.

[0081] Specifically, the signal processing unit first performs preprocessing on the echo signal, including the following steps: Step 1: Data preprocessing: Step 1.1: Perform analog-to-digital conversion on the echo signal to generate a time-domain discrete sequence.

[0082] Step 1.2: Use the wavelet threshold denoising algorithm to eliminate environmental noise. Specifically, perform 3-layer decomposition using the Daubechies4 wavelet basis and set the soft threshold function to process the high-frequency coefficients.

[0083] Step 1.3: Segment the signal through a sliding window to generate time series signal segments.

[0084] Step 2: Construct a spatio-temporal feature extraction network based on the attention mechanism, including the following steps: Step 2.1: Construct a dual-branch feature extraction structure: Spatial branch: Use a 5-layer 1D-CNN (Convolutional Neural Network, with convolutional kernel sizes [7, 5, 3, 3, 3] and channel numbers [32, 64, 128, 256, 512]). After each layer, batch normalization and ReLU activation are connected. This spatial branch is used to extract shallow texture features.

[0085] Temporal branch: Use a bidirectional LSTM (Long Short-Term Memory, with 256 hidden layer units) to handle the time dimension dependence. This spatial branch is used to extract time features.

[0086] Step 2.2: Embed the multi-head self-attention mechanism (4 attention heads) to calculate the convolutional results obtained from the above spatial branch and temporal branch to obtain the attention weight matrix.

[0087] Step 2.3: The feature fusion layer performs the Hadamard product of the outputs of the above spatial branch and temporal branch with the attention weight matrix to generate an enhanced feature vector.

[0088] Step 3: Multi-dimensional feature fusion, including the following specific steps: Step 3.1: Perform hierarchical fusion of spatio-temporal features: Primary features: The original time-frequency features (referring to the spectral centroid and bandwidth extracted from the echo signal through STFT).

[0089] Intermediate features: Namely, the shallow texture features extracted by the CNN in the above spatial branch.

[0090] Advanced features: The enhanced feature vector extracted by the LSTM-attention mechanism in the above temporal branch.

[0091] Step 3.2: Use the feature concatenation method to fuse the above multi-dimensional features (including primary features, intermediate features, and advanced features) to form a 4096-dimensional feature vector.

[0092] Step 3.3: Perform feature dimensionality reduction through a learnable parameter matrix (e.g., using the PCA-Whitening algorithm) to finally generate an optimized feature vector of 512 dimensions.

[0093] Step 403: Use a deep reinforcement learning algorithm to classify the multi-dimensional features to obtain the animal category.

[0094] Step 4: The specific steps of the deep reinforcement learning algorithm are as follows: Step 4.1: Define the Markov decision process: State space S: Obtain the normalized values of the 512-dimensional optimized feature vector; Action space A: Perform threshold judgment on the above normalized values to obtain the animal category, for example: five categories of animals {felidae, canidae, ungulates, birds, others}.

[0095] Step 5: Train the policy in stages: Use an animal radar echo database (containing 100,000 labeled samples) to train the classifier.

[0096] The above embodiment provides an animal target detection system based on an airborne radar and intelligent recognition. The system includes a drone platform equipped with radar equipment, an integrated antenna array, and a radar host. After the system is started, the transmitting module of the radar host generates a high-frequency electromagnetic wave signal in the terahertz frequency band, which is transmitted through the directional antenna array integrated on the surface of the drone body to form a high-resolution beam. During the propagation of the electromagnetic wave, after encountering an animal target, it is reflected to form an echo signal containing the target feature information. The receiving antenna adopts array optimization design and spatial filtering technology, which can improve the receiving ability of weak target signals while ensuring the lightweight of the structure, effectively enhancing the spatial resolution of the system. The signal processing unit inside the radar host integrates a spatio-temporal feature extraction network based on the attention mechanism, which can extract multi-dimensional target information including distance, speed, direction, and scattering features from the radar echo. This module performs classification learning and optimization judgment on the target features through a deep reinforcement learning algorithm, and finally realizes the accurate recognition and classification of animal targets. This method does not rely on external satellite remote sensing data, and only completes the target detection task based on the drone platform and the airborne radar system, with good flexibility and deployment efficiency. The use of terahertz band signals significantly improves the resolution of the radar system and the recognition ability of small animals in complex habitats, thereby improving the biodiversity monitoring accuracy of the system in the wild environment.

[0097] The following uses a specific embodiment to illustrate the specific implementation process of the animal target detection method based on an airborne radar and intelligent recognition proposed in this application. The method includes the following steps: (1) The terahertz wave signal transmitted by the radar host generates an echo signal after encountering an obstacle, and the echo signal is received by the antenna array integrated in the drone wing. The receiving antenna is a deformable structure that can fit the surface of the drone fuselage; in addition, the material structure of the receiving antenna is a metasurface structure that can precisely manipulate electromagnetic waves at the sub-wavelength scale, effectively improving the quality and accuracy of the received signal.

[0098] (2) The received echo signal is preprocessed (including noise and interference removal, signal amplification, filtering, etc.) to improve the signal quality; Then clutter suppression is performed. The system combines multi-layer clutter suppression with micro-Doppler feature extraction. It separates ground clutter and vegetation clutter through the height information h and the polarization scattering matrix, and uses three-pulse MTI processing to remove static clutter and extract the micro-Doppler features of the target to distinguish moving and static targets, improving the clutter suppression ability of the system by 15 - 20 dB; In clutter suppression, the influence of vegetation on the signal needs to be considered. Since the leaf structures of different vegetation layers are different, they have different electromagnetic response characteristics. This application conducts multi-level environmental modeling of the vegetation environment in the forest, including an electromagnetic response model of the canopy layer designed based on optimized electromagnetic characteristics and an adaptive model describing the scattering characteristics of the understory vegetation layer. For example, the amplitude and phase distribution of vegetation clutter are estimated according to the scattering loss model of radar signals in each layer of the above vegetation, and then the clutter suppression model enhanced by deep learning suppresses the clutter in the surface layer.

[0099] Specifically, the steps for surface clutter suppression are as follows: (2.1) Vegetation parameter collection and analysis: Collect detailed parameters of the canopy layer in the target area, including tree species type, average tree height, crown size, leaf shape and size, leaf density, branch structure, etc. At the same time, obtain the electromagnetic characteristic parameters of the trees, such as the complex permittivity of leaves and branches, which can be obtained through on-site measurement or reference to literature data.

[0100] Analyze the influence of different parameters on the electromagnetic scattering characteristics of the canopy layer, determine the key parameters, and provide a basis for model construction.

[0101] (2.2) Electromagnetic response model construction and optimization: Initial model selection: Based on the physical optics model and the equivalent medium model, construct an initial electromagnetic response model of the canopy layer. The physical optics model is used to describe the specular reflection and shadow effects of large-scale targets (such as tree trunks, main branches, etc.) in the canopy layer; the equivalent medium model regards the canopy layer as a whole homogeneous isotropic medium and is used to describe the overall propagation characteristics of radar signals in the canopy layer.

[0102] Model Optimization: Introduce optimization algorithms such as genetic algorithms and particle swarm optimization algorithms to optimize and adjust the model parameters. Taking the actually measured radar echo data as the objective function, through iterative optimization, minimize the error between the model output and the actual data, so as to obtain the above-mentioned electromagnetic response model of the canopy layer designed based on optimized electromagnetic characteristics.

[0103] (2.3) Understory Vegetation Feature Extraction: Classify and extract the features of the understory vegetation in the target area, including parameters such as the types, heights, densities, and leaf area indices of herbaceous plants and shrubs. Use field surveys, remote sensing image analysis and other means to obtain this information.

[0104] (2.4) Divide the understory vegetation into different sub-regions according to its distribution and characteristics for subsequent adaptive modeling.

[0105] (2.5) Adaptive Scattering Model Establishment: Initial Model Setting: Adopt a model based on geometric structure, such as a ray tracing model or a geometric optics model, as the initial model of the scattering characteristics of the understory vegetation layer. According to the geometric structure characteristics of the understory vegetation, set the basic parameters of the model, such as the positions, shapes, and sizes of the scatterers.

[0106] Adaptive Adjustment Mechanism: Use adaptive algorithms in machine learning, such as adaptive neural networks and adaptive support vector machines, to adaptively adjust the above initial model according to the actual radar echo data samples and vegetation feature parameters. By continuously learning and updating the model parameters, the model can adapt to the scattering characteristic changes of the understory vegetation in different regions and different growth states.

[0107] (2.6) Clutter Suppression Model Based on Deep Learning: Data Preparation and Preprocessing: Collect a large amount of radar echo data containing vegetation clutter and target signals as the training data set. Perform preprocessing on the data, including data cleaning, normalization, feature extraction and other operations to improve the quality and usability of the data.

[0108] Network Structure Design: Design a deep learning network, such as a convolutional neural network (CNN) or a recurrent neural network (RNN) and its variants, for suppressing clutter in the surface layer. The network input is the radar echo signal, and the output is the target signal after clutter suppression.

[0109] Training Process: Use the training set to train the deep learning network, and adjust the network parameters through optimization algorithms (such as stochastic gradient descent, Adam optimization algorithm, etc.) to minimize the error between the network output and the actual target signal. During the training process, use the validation set to verify the model to prevent overfitting.

[0110] (2.7)Model integration and system implementation: Integrate the optimized electromagnetic response model of the canopy layer, the adaptive scattering model of the understory vegetation layer, and the clutter suppression model based on deep learning to form a complete multi-layer vegetation scattering characteristic analysis and clutter suppression system.

[0111] Design data interfaces and signal transmission mechanisms to ensure data flow and collaborative work among the models.

[0112] Implement this technical solution in an actual radar system, integrate the system into the radar signal processing platform, and conduct field tests to collect radar echo data, including scenarios of different vegetation types, different target types and motion states, etc.

[0113] Evaluate the performance of the system. The main indicators include clutter suppression effect (such as clutter suppression ratio, target detection probability, etc.), target signal fidelity (such as signal distortion, signal-to-noise ratio, etc.), real-time performance and stability of the system, etc. According to the test results, further optimize and improve the system to enhance its performance and reliability.

[0114] Through the above technical solutions, it is possible to effectively model the scattering loss of radar signals in vegetation, and use the deep learning-enhanced model to suppress surface clutter, improving the radar's detection and recognition capabilities for targets in vegetation-covered areas.

[0115] In addition, the signal also passes through the combination of graph neural network and spatio-temporal attention mechanism to achieve more accurate time-domain and frequency-domain feature analysis.

[0116] (3)Extract target feature information. The feature extraction process is to extract the features of the target from the processed signal above, such as speed, size, shape, etc.

[0117] See the description part of the spatio-temporal feature extraction network based on the attention mechanism in the above text for the specific feature extraction process.

[0118] (4)Target detection and classification and recognition: According to the extracted target features, use deep reinforcement learning algorithms to distinguish different types of animal targets. It is also possible to use the improved feature pyramid network and the context-aware spatio-temporal attention mechanism to analyze and classify the target features, achieving accurate recognition of different categories of wild animals. The classification system adopts a deep residual network structure and improves the classification accuracy through a hierarchical pre-training strategy.

[0119] (5)Target tracking: Adopt an enhanced Kalman filtering algorithm based on deep learning and combine it with distributed MIMO technology to track the target.

[0120] (6) While performing target tracking, the above enhanced Kalman filtering algorithm based on deep learning can also be used to co-control the system parameters of the airborne radar system: specifically including: First, fuse the target and UAV states: Construct a unified state vector, including the position, velocity, acceleration of the target, and the flight state parameters of the UAV (such as position, velocity, attitude, etc.). Through this fusion method, the states of the target and the UAV can be estimated and predicted simultaneously. Feature extraction based on deep learning: Use a deep learning model to extract the target features in the radar echo signal and input them as observation information into the Kalman filter. At the same time, obtain the current state information of the UAV using the sensor data of the UAV (such as IMU data, GPS data, etc.).

[0121] Then, Kalman filter prediction and update: Use the prediction equation of the Kalman filter to predict the states of the target and the UAV at the current moment based on the state estimation value at the previous moment. Compare the target features extracted by deep learning and the UAV sensor data with the predicted values, and calculate the residuals. Then use the Kalman gain to update the state estimation value to obtain a more accurate state estimation of the target and the UAV.

[0122] Finally, perform feedback control: Use the updated UAV state estimation to optimize the system working parameters in real time, including flight altitude, speed, and beam pointing angle. At the same time, use the reinforcement learning optimization module to adaptively adjust the flight path and beam pointing.

[0123] The above process can also be designed modularly respectively to form a schematic structural diagram of an animal target detection system based on an airborne radar and intelligent recognition as shown in Figure 5 which includes a transmitter, a receiver, an antenna array, other components, and an echo signal preprocessing unit, a clutter suppression unit, a target detection and recognition unit, a target tracking unit, and a feedback control unit in the signal processing unit.

[0124] Next, the animal target detection device based on an airborne radar and intelligent recognition provided by the present invention will be described. The animal target detection device based on an airborne radar and intelligent recognition described below can be correspondingly referred to the animal target detection method based on an airborne radar and intelligent recognition described above.

[0125] As shown in Figure 6 shown, Figure 6 a schematic structural diagram of an animal target detection device based on an airborne radar and intelligent recognition is provided. The device includes the following modules: A signal reception module 601, used to receive echo signals; A feature extraction module 602, configured to input the echo signal into a spatio-temporal feature extraction network based on an attention mechanism, and obtain multi-dimensional features extracted by the spatio-temporal feature extraction network based on the attention mechanism from the echo signal; An animal classification module 603, configured to classify the multi-dimensional features by using a deep reinforcement learning algorithm to obtain an animal category.

[0126] In an embodiment, the spatio-temporal feature extraction network based on the attention mechanism includes a dual-branch feature extraction layer, a multi-head self-attention layer, and a multi-dimensional feature fusion layer; the dual-branch feature extraction layer includes a spatial branch and a temporal branch; the above-mentioned feature extraction module 602 is further configured to: Input the echo signal into the spatial branch and the temporal branch in the dual-branch feature extraction layer respectively, to obtain shallow texture features output by the spatial branch and temporal features output by the temporal branch; input the shallow texture features and the temporal features into the multi-head self-attention layer, so that the multi-head self-attention layer performs a convolution operation on the shallow texture features and the temporal features to obtain an attention weight matrix, and perform a Hadamard product operation on the attention weight matrix and the temporal features to obtain an enhanced feature vector output by the multi-head self-attention layer; input the enhanced feature vector and the shallow texture features into the multi-dimensional feature fusion layer, so that the multi-dimensional feature fusion layer performs splicing on the enhanced feature vector and the shallow texture features to obtain multi-dimensional features output by the multi-dimensional feature fusion layer.

[0127] Figure 7 Illustrates a schematic structural diagram of an electronic device, such as Figure 7As shown in the figure, the electronic device adopts a high-performance heterogeneous computing architecture, including: a processor 710, a communications interface 720, a memory 730, and a communication bus 740. Among them, the processor 710 adopts a multi-core design, integrates a neural network acceleration unit, has a main frequency of 3.5 GHz, and supports a single-precision floating-point operation ability of up to 12 TFLOPS; the communications interface 720 supports a variety of high-speed data transmission protocols, including 10 Gb Ethernet, optical fiber communication, and wireless data links, to achieve high-speed interconnection of internal components and external devices of the system; the memory 730 adopts a hierarchical storage architecture, including a cache, a DDR4 main memory, and an NVMe solid-state drive, with a total capacity of up to 1 TB, and supports ECC error correction function to ensure data integrity; the communication bus 740 adopts the high-speed PCIe 4.0 specification, with a bandwidth of 64 GB / s, to support efficient data exchange between devices. The processor 710, the communications interface 720, and the memory 730 complete data transmission and control signal interaction with each other through the communication bus 740. The processor 710 can call the logical instructions in the memory 730 to execute the animal target detection method based on the airborne radar and intelligent recognition. The method includes: receiving an echo signal; inputting the echo signal into a spatio-temporal feature extraction network based on the attention mechanism to obtain multi-dimensional features extracted by the spatio-temporal feature extraction network based on the attention mechanism from the echo signal; using a deep reinforcement learning algorithm to classify the multi-dimensional features to obtain the animal category.

[0128] In addition, when the logical instructions in the above-mentioned memory 730 can be implemented in the form of software functional units and sold or used as an independent product, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The foregoing storage medium includes: various media such as a USB flash drive, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk, or an optical disc that can store program codes.

[0129] On the other hand, the present invention also provides a computer program product, which includes a computer program. The computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the animal target detection method based on airborne radar and intelligent recognition provided by the above-mentioned various methods. The method includes: receiving an echo signal; inputting the echo signal into a spatio-temporal feature extraction network based on an attention mechanism to obtain multi-dimensional features extracted by the spatio-temporal feature extraction network based on the attention mechanism from the echo signal; and using a deep reinforcement learning algorithm to classify the multi-dimensional features to obtain the animal category.

[0130] In another aspect, the present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it is implemented to execute the animal target detection method based on airborne radar and intelligent recognition provided by the above-mentioned various methods. The method includes: receiving an echo signal; inputting the echo signal into a spatio-temporal feature extraction network based on an attention mechanism to obtain multi-dimensional features extracted by the spatio-temporal feature extraction network based on the attention mechanism from the echo signal; and using a deep reinforcement learning algorithm to classify the multi-dimensional features to obtain the animal category.

[0131] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art can understand and implement it without creative labor.

[0132] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on such an understanding, the essence of the above technical solution or the part that contributes to the prior art can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disc, etc., and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0133] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An animal target detection system based on airborne radar and intelligent recognition, characterized in that: The system includes a drone, an antenna array integrated with the drone body, and a radar host; wherein, The radar host includes a signal processing unit; The antenna array is designed as a deformable structure and can be attached to the fuselage surface of the drone; The antenna array is used to send radio frequency signals in the terahertz frequency band to the target area under the control of the system, and receive echo signals reflected by obstacles; The signal processing unit is used to receive the echo signal, extract multi-dimensional features from the echo signal using a spatiotemporal feature extraction network based on an attention mechanism, and classify the multi-dimensional features through a deep reinforcement learning algorithm to obtain an animal category.

2. The animal target detection system based on airborne radar and intelligent recognition according to claim 1 is characterized in that: The antenna array adopts a metasurface structure to achieve sub-wavelength scale electromagnetic wave manipulation.

3. The animal target detection system based on airborne radar and intelligent recognition according to claim 1 is characterized in that: The signal processing unit is composed of a neuromorphic chip and a three-dimensional photonic integrated circuit processor.

4. The animal target detection system based on airborne radar and intelligent recognition according to claim 1 is characterized in that: The system adopts multi-level environmental modeling, including a canopy electromagnetic response model designed based on optimized electromagnetic characteristics, and an adaptive scattering model that describes the scattering characteristics of the understory vegetation layer; The signal processing unit is used to input the echo signal into a clutter suppression model based on deep learning to obtain an echo signal after clutter suppression; wherein the clutter suppression model based on deep learning is constructed based on the canopy electromagnetic response model designed based on optimized electromagnetic characteristics and an adaptive scattering model describing the scattering characteristics of the understory vegetation layer; The signal processing unit is used to input the echo signal after clutter suppression into the spatiotemporal feature extraction network based on the attention mechanism to obtain multi-dimensional features output by the spatiotemporal feature extraction network based on the attention mechanism; The signal processing unit is used to classify the multi-dimensional features through a deep reinforcement learning algorithm to obtain an animal category.

5. The animal target detection system based on airborne radar and intelligent recognition according to claim 1 is characterized in that: The feedback control mechanism of the system includes a deep learning enhanced state estimator, a reinforcement learning based path optimization algorithm and a terahertz beam adaptive control unit. Parameter optimization is achieved through distributed edge computing to ensure the stability of the system in complex electromagnetic environments.

6. An animal target detection method based on airborne radar and intelligent recognition, applied to the signal processing unit in the animal target detection system based on airborne radar and intelligent recognition as claimed in any one of claims 1 to 5, characterized in that: The method comprises: receiving an echo signal; Inputting the echo signal into a spatiotemporal feature extraction network based on an attention mechanism to obtain multi-dimensional features extracted from the echo signal by the spatiotemporal feature extraction network based on the attention mechanism; The multi-dimensional features are classified using a deep reinforcement learning algorithm to obtain animal categories.

7. The animal target detection method based on airborne radar and intelligent recognition according to claim 6 is characterized in that: The spatiotemporal feature extraction network based on the attention mechanism includes a dual-branch feature extraction layer, a multi-head self-attention layer and a multi-dimensional feature fusion layer; the dual-branch feature extraction layer includes a spatial branch and a temporal branch; the echo signal is input into the spatiotemporal feature extraction network based on the attention mechanism to obtain the multi-dimensional features extracted from the echo signal by the spatiotemporal feature extraction network based on the attention mechanism, including: Inputting the echo signal into the space branch and the time branch in the dual-branch feature extraction layer respectively, obtaining the shallow texture features output by the space branch and the time features output by the time branch; Inputting the shallow texture features and the time features into the multi-head self-attention layer, so that the multi-head self-attention layer performs a convolution operation on the shallow texture features and the time features to obtain an attention weight matrix, and performing a Hadamard product operation on the attention weight matrix and the time features to obtain an enhanced feature vector output by the multi-head self-attention layer; The enhanced feature vector and the shallow texture feature are input into the multi-dimensional feature fusion layer, so that the multi-dimensional feature fusion layer performs splicing on the enhanced feature vector and the shallow texture feature to obtain the multi-dimensional feature output by the multi-dimensional feature fusion layer.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the computer program, the animal target detection method based on airborne radar and intelligent identification as described in any one of claims 6 to 7 is implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the animal target detection method based on airborne radar and intelligent recognition as described in any one of claims 6 to 7 is implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the animal target detection method based on airborne radar and intelligent recognition as described in any one of claims 6 to 7 is implemented.

Images