A UAV Infrared Target Localization System Integrating Multi-Model KF Algorithm and Deep Learning

Abstract

The application provides an unmanned aerial vehicle infrared target positioning system fusing multi-model Kalman filtering and deep learning, and relates to the technical field of unmanned aerial vehicle navigation and positioning. The application discloses an unmanned aerial vehicle infrared target positioning system fusing multi-model Kalman filtering and deep learning, and relates to the technical field of unmanned aerial vehicle navigation and positioning. The application discloses an unmanned aerial vehicle infrared target positioning system fusing multi-model Kalman filtering and deep learning, and relates to the technical field of unmanned aerial vehicle navigation and positioning. The application discloses an unmanned aerial vehicle infrared target positioning system fusing multi-model Kalman filtering and deep learning, and relates to the technical field of unmanned aerial vehicle navigation and positioning. The application discloses an unmanned aerial vehicle infrared target positioning system fusing multi-model Kalman filtering and deep learning, and relates to the technical field of unmanned aerial vehicle navigation and positioning. The application discloses an unmanned aerial vehicle infrared target positioning system fusing multi-model Kalman filtering and deep learning, and relates to the technical field of unmanned aerial vehicle navigation and positioning. The application discloses an unmanned aerial vehicle infrared target positioning system fusing multi-model Kalman filtering and deep learning, and relates to the technical field of unmanned aerial vehicle navigation and positioning. The application discloses an unmanned aerial vehicle infrared target positioning system fusing multi-model Kalman filtering and deep learning, and relates to the technical field of unmanned aerial vehicle navigation and positioning. The application discloses an unmanned aerial vehicle infrared target positioning system fusing multi-model Kalman filtering and deep learning, and relates to the technical field of unmanned aerial vehicle navigation and positioning. The application discloses an unmanned aerial vehicle infrared target positioning system fusing multi-model Kalman filtering and deep learning, and relates to the technical field of unmanned aerial vehicle navigation and positioning. The application discloses an unmanned aerial vehicle infrared target positioning system fusing multi-model Kalman filtering and deep learning, and relates to the technical field of unmanned aerial vehicle navigation and positioning. The application discloses an unmanned aerial vehicle infrared target positioning system fusing multi-model Kalman filtering and deep learning, and relates to the technical field of unmanned aerial vehicle navigation and positioning. The application discloses an unmanned aerial vehicle infrared target positioning system fusing multi-model Kalman filtering and deep learning, and relates to the technical field of unmanned

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) navigation and positioning technology, specifically to an UAV infrared target positioning system that integrates multi-model KF algorithm (Kalman filter) and deep learning. Background Technology

[0002] According to Chinese Patent No. CN105467416A, a precise positioning system for unmanned aerial vehicles (UAVs) includes a flight control system, a GPS positioning system, and an ultrasonic positioning system. The flight control system is connected to an infrared receiver, the GPS positioning system, and the ultrasonic positioning system, respectively. The infrared receiver is connected to a remote controller. Before reaching the area requiring precise landing, the GPS positioning system guides the flight control system based on manual remote control. Upon reaching the area, the ultrasonic positioning system automatically activates, continuously transmitting signals between the ultrasonic receiver and the ultrasonic transmitter. The flight control system continuously calculates the distance between them, updates the target position, and adjusts the flight attitude as needed. Therefore, this invention achieves basic positioning with the GPS positioning system and precise positioning with the ultrasonic positioning system. By combining the GPS and ultrasonic positioning systems, precise positioning of the UAV can be achieved, allowing it to land accurately in a specific area.

[0003] Currently, various drone positioning technologies exist on the market, with GPS positioning systems based on satellite signals and ultrasonic positioning systems utilizing the propagation characteristics of sound waves being the most common. A typical drone precision positioning system includes a flight control system, a GPS positioning system, and an ultrasonic positioning system. This system switches between GPS and ultrasonic positioning systems via a switch on the remote controller. It utilizes the signal interaction between the ultrasonic transmitter and receiver, calculates distance by measuring the time difference, and combines this with a triangulation algorithm to achieve positioning. For example, when the distance detection module detects that the drone is less than ten meters from the precise landing area, the flight control system activates the ultrasonic positioning system. The ultrasonic transmitter is set at a right angle, and the controller uploads the calculated distance to the flight control system via the wireless communication module, thereby adjusting the flight attitude.

[0004] However, existing technologies have revealed the following problems in practical applications: 1. While GPS positioning systems can achieve relatively accurate positioning in open spaces, their accuracy drops significantly in environments where signals are easily blocked or interfered with, such as urban high-rises, canyons, dense forests, and indoors, due to the characteristics of satellite signal propagation. Signal loss can even occur, preventing drones from accurately determining their location and severely impacting mission execution. 2. Ultrasonic positioning systems have a relatively limited range, generally suitable for short-distance, small-area precise landing positioning. Furthermore, ultrasonic signals are easily affected by environmental factors during propagation, such as airflow, temperature changes, and humidity fluctuations. These factors alter the propagation speed and direction of sound waves, leading to deviations in distance measurement and making it difficult to guarantee positioning accuracy. 3. Existing positioning systems often rely on a single technology or a simple combination of technologies, lacking comprehensive consideration and adaptability to complex environments and diverse mission requirements. For example, in real-world scenarios involving complex multi-source data fusion, dynamic environmental changes, and high-precision target positioning, existing technologies struggle to meet the demands for efficient, accurate, and reliable positioning.

[0005] Therefore, a UAV infrared target localization system that integrates multi-model Kalman filtering and deep learning is needed to solve the above problems. Summary of the Invention

[0006] Technical problems to be solved

[0007] To address the shortcomings of existing technologies, this invention provides a UAV infrared target localization system that integrates multi-model Kalman filtering and deep learning, thus solving the problems mentioned in the background section.

[0008] Technical solution

[0009] To achieve the above objectives, this invention provides the following technical solution: a UAV infrared target localization system integrating multi-model Kalman filtering and deep learning, comprising a main system. The main system includes a data acquisition module, a multi-source data fusion module, a feature extraction and enhancement module, a dynamic model optimization module, a target state estimation module, a localization calculation module, and a flight decision-making and planning module. The data acquisition module deploys a high-resolution infrared sensor to accurately capture target infrared radiation information. Simultaneously, it integrates a high-definition visual camera and a high-precision lidar to collect environmental data from all directions. It employs an adaptive sampling strategy based on fuzzy logic, particle swarm optimization, and adaptive thresholding. The sampling frequency and resolution of the sensor are dynamically adjusted according to environmental complexity and target motion state. Fuzzy logic is used to evaluate the degree of fuzziness in environmental complexity and target motion state; particle swarm optimization is used to find the optimal solution among multiple possible combinations of sampling parameters; and the adaptive thresholding dynamically adjusts the sampling triggering conditions according to the rate of change of the data. This ensures data quality while reducing data redundancy and processing burden, guaranteeing data richness and integrity.

[0010] Preferably, the multi-source data fusion module employs an adaptive fusion algorithm that integrates biological neuron collaboration mechanisms, tensor decomposition, and deep learning attention mechanisms to simulate information interaction and integration between neurons. It performs feature encoding on infrared, visual, and lidar data. Tensor decomposition decomposes multi-source data from a high-dimensional space into a low-dimensional subspace to extract the core features of the data. The deep learning attention mechanism dynamically adjusts the weights of each sensor data according to the dynamic changes in the environment through competitive learning and collaborative feedback, achieving deep fusion of multi-source data and providing a high-quality data foundation for subsequent processing. At the same time, cross-modal consistency constraints are introduced by constructing a cross-modal similarity metric function based on Wasserstein distance, cosine similarity, and mutual information to ensure that the data from various modalities maintain information consistency and complementarity during the fusion process.

[0011] The feature extraction enhancement module adopts a convolutional neural network architecture that integrates quantum enhancement, generative adversarial networks, and residual networks. It utilizes the superposition and entanglement properties of quantum states to accelerate the feature extraction process. The generator in the generative adversarial network generates simulated feature data, and the discriminator distinguishes between real and simulated features. More representative features are mined through adversarial training. The residual network solves the gradient vanishing problem in deep networks through skip connections, enhancing the network's ability to extract complex features. Combined with attention mechanisms and improved dilated convolution techniques, it focuses on key target features, expands the receptive field, and enhances the ability to extract target features in complex scenes.

[0012] The dynamic model optimization module introduces a model optimization strategy based on Generative Adversarial Networks (GANs), transfer learning, and genetic algorithms. The generator generates simulated target and scene data, the discriminator distinguishes between real and simulated data, and the generalization ability of the model is improved through adversarial training. The transfer learning transfers model knowledge trained in similar scenarios, accelerating the convergence speed of the model in new scenarios. The genetic algorithm searches for the optimal model configuration in the model space by encoding, selecting, crossing over, and mutating the model's structure and parameters, achieving rapid optimization and adaptive adjustment of the model.

[0013] The target state estimation module employs an algorithm that integrates reinforcement learning, adaptive Kalman filtering, and particle filtering to model the target state estimation problem as a Markov decision process. By using a reward function for the UAV target localization scenario, it guides the agent to learn the optimal state estimation strategy. The adaptive Kalman filter dynamically adjusts the filtering parameters based on the rapid changes in target motion and complex environmental noise interference. The particle filter estimates the target state through sampling and weight updates of a large number of particles.

[0014] The positioning and calculation module is based on the principle of multi-view data fusion and improved triangulation. It combines the real-time attitude information of the UAV with the mechanism of dynamic coordinate transformation, error compensation and least squares fitting. By monitoring the flight attitude and environmental parameter changes of the UAV in real time, the coordinate system is dynamically adjusted in the positioning process and the measurement error is compensated online. The least squares fitting is used to find the optimal positioning solution in multiple sets of measurement data, adapting to the changes in the flight attitude of the UAV and the dynamic interference of the environment in real time, so as to achieve high-precision three-dimensional positioning of the target.

[0015] The flight decision-making and planning module employs a dual-deep Q-network architecture based on deep reinforcement learning, ant colony optimization, and A algorithm. Combined with experience replay and priority experience replay mechanisms, it effectively reduces overfitting in Q-value estimation and accelerates the learning process. The deep reinforcement learning learns the optimal flight strategy. The ant colony optimization algorithm simulates pheromone transmission during ant foraging to find potential optimal paths in the path search space. The A algorithm utilizes heuristic functions to quickly search for the optimal path from the current position to the target position. Simultaneously, it considers the UAV's energy consumption, flight safety constraints, and dynamic environmental factors to plan the optimal flight trajectory in real time. This ensures that the UAV can efficiently track targets while achieving long endurance and safe flight.

[0016] Preferably, in the multi-source data fusion module, an adaptive fusion algorithm based on the biological neuron collaboration mechanism, tensor decomposition, and deep learning attention mechanism constructs a neuron connection weight matrix to simulate the excitation and inhibition relationship between neurons. Combined with the low-dimensional feature representation after tensor decomposition and the weight allocation of the attention mechanism, it achieves efficient fusion of sensor data, improves the system's adaptability to complex and changing environments, and enhances the accuracy of target perception.

[0017] Preferably, the parallel computing capability of the qubits in the feature extraction enhancement module simultaneously explores the optimal solution direction during the parameter update process of the convolution kernel. The adversarial training mechanism of the generative adversarial network prompts the network to mine representative features. The skip connections of the residual network ensure the effective extraction of deep features and accelerate model convergence. Moreover, it can mine deep target features that are difficult to discover using traditional methods.

[0018] Preferably, in the dynamic model optimization module, the generative adversarial network (GAN) trains against adversarial forces to make the data distribution generated by the generator gradually approach the distribution of real data. The transfer learning reduces the training time and sample requirements of the model in new scenarios through feature mapping and knowledge transfer between the source domain and the target domain. The genetic algorithm further improves the performance and adaptability of the model by optimizing the model structure and parameters.

[0019] Preferably, under the conditions of target motion pattern and environmental noise, the target state estimation module uses reinforcement learning to guide the agent to select the optimal estimation strategy, and adaptive Kalman filtering to quickly adjust the filtering parameters. Particle filtering improves estimation accuracy through particle sampling and weight updates, keeping the estimation error within a small range. This effectively improves the stability and reliability of target state estimation.

[0020] Preferably, when considering the energy consumption of the UAV, the flight decision planning module establishes an energy model to correlate factors such as flight speed, altitude, and attitude with energy consumption. When planning the flight trajectory, it uses deep reinforcement learning, ant colony algorithm, and A algorithm to prioritize paths with low energy consumption. When considering flight safety constraints, it combines environmental perception data to avoid obstacles and dangerous areas, thereby ensuring the flight safety of the UAV.

[0021] Preferably, when collecting data, the data acquisition module dynamically and intelligently adjusts the sampling frequency and resolution of the sensor according to the real-time changes in environmental complexity and target motion state, so as to minimize data redundancy and processing burden while ensuring data quality.

[0022] Preferably, the multi-source data fusion module introduces a cross-modal similarity measurement function based on Wasserstein distance, cosine similarity, and mutual information during the data fusion process. This function measures the similarity and complementarity between different modalities from various perspectives, ensuring that different modalities maintain a high degree of consistency and complementarity of information during the fusion process, thereby further improving the quality and reliability of the fused data.

[0023] The different modal data involved here specifically include infrared radiation data collected by infrared sensors, which can reflect the thermal characteristic distribution of the target and play a key role in identifying the target's heat-generating parts and temperature differences; visible light image data acquired by high-definition vision cameras, which can present the target's shape, color, texture and other appearance features, helping to identify the target's category and details from a visual perspective; and distance data collected by high-precision lidar, which can construct three-dimensional spatial structure information of the target and its surrounding environment, accurately depicting the target's position and relative distance to surrounding objects.

[0024] This module encodes features from the aforementioned different modalities of data by simulating information interaction and integration between neurons. Tensor decomposition is used to decompose multi-source data from a high-dimensional space into a low-dimensional subspace, extracting the core features of the data. The deep learning attention mechanism dynamically adjusts the weights of each sensor's data based on dynamic environmental changes, such as weather changes, lighting changes, and targets entering different scenes, through competitive learning and collaborative feedback. This achieves deep fusion of multi-source data, providing a high-quality data foundation for subsequent processing.

[0025] Simultaneously, cross-modal consistency constraints are introduced. By constructing a cross-modal similarity metric function based on Wasserstein distance, cosine similarity, and mutual information, the consistency and complementarity of information between different modalities are ensured during the fusion process. Specifically, Wasserstein distance is used to measure the differences between the distributions of different modalities to ensure the consistency of the data distribution in the overall data; cosine similarity measures the similarity between feature vectors of different modalities from the perspective of the angle between vectors, capturing the similarity of features; and mutual information evaluates the degree of information sharing between different modalities, highlighting the complementarity of the data.

[0026] Beneficial effects

[0027] This invention provides a UAV infrared target localization system that integrates multi-model Kalman filtering and deep learning. It has the following advantages:

[0028] 1. The data acquisition module of this invention employs an adaptive sampling strategy based on fuzzy logic, particle swarm optimization, and adaptive thresholding. This strategy dynamically and intelligently adjusts the sensor's sampling frequency and resolution according to real-time changes in environmental complexity and target motion state. This effectively avoids data redundancy and significantly reduces the burden of subsequent processing while ensuring data quality. The multi-source data fusion module utilizes an adaptive fusion algorithm that integrates biological neuron collaboration mechanisms, tensor decomposition, and deep learning attention mechanisms. This not only enables efficient feature encoding of different modalities such as infrared, vision, and LiDAR, but also extracts core features through tensor decomposition and dynamically adjusts data weights using deep learning attention mechanisms to achieve deep fusion. Simultaneously, a cross-modal similarity metric function based on Wasserstein distance, cosine similarity, and mutual information is introduced to ensure high consistency and complementarity of different modalities during the fusion process. This provides a high-quality data foundation for subsequent modules and significantly improves the accuracy and efficiency of the entire system in acquiring and processing target and environmental information.

[0029] 2. The feature extraction enhancement module in this invention adopts a convolutional neural network architecture that integrates quantum enhancement, generative adversarial networks (GANs), and residual networks, exhibiting unique advantages. Quantum enhancement utilizes the superposition and entanglement properties of quantum states to accelerate the feature extraction process during convolutional kernel parameter updates. GANs, through adversarial training between the generator and discriminator, uncover more representative features, while the skip connections of the residual network solve the gradient vanishing problem in deep networks, enhancing the ability to extract complex features. Combining attention mechanisms with improved dilated convolution techniques further focuses on key target features and expands the receptive field. The dynamic model optimization module introduces optimization strategies based on GANs, transfer learning, and genetic algorithms. GANs improve the model's generalization ability, transfer learning accelerates model convergence in new scenarios, and genetic algorithms search for the optimal model configuration, achieving rapid optimization and adaptive adjustment of the model. These modules work synergistically, enabling the system to accurately extract deep-level target features, quickly adapt to different scenarios, and significantly improve the overall performance and adaptability of the model.

[0030] 3. The target state estimation module in this invention employs an algorithm that integrates reinforcement learning, adaptive Kalman filtering, and particle filtering. It models the target state estimation problem as a Markov decision process, guiding the agent to learn the optimal estimation strategy through a specific reward function. Adaptive Kalman filtering dynamically adjusts filtering parameters based on target motion changes and environmental noise interference. Particle filtering improves estimation accuracy through particle sampling and weight updates, effectively enhancing the stability and reliability of target state estimation and keeping the estimation error within a minimal range. The positioning and calculation module, based on multi-view data fusion and improved triangulation principles, combines real-time UAV attitude information with dynamic coordinate transformation, error compensation, and least squares fitting mechanisms to adapt to changes in UAV flight attitude and dynamic environmental interference in real time, achieving high-precision three-dimensional positioning of the target. This high-precision target state estimation and positioning capability enables the system to accurately lock the target position even in complex environments, providing a reliable basis for subsequent flight decisions.

[0031] 4. The flight decision-making and planning module in this invention adopts a dual-deep Q-network architecture based on deep reinforcement learning, ant colony optimization, and A algorithm. Combined with experience replay and priority experience replay mechanisms, it effectively reduces the overfitting problem in Q-value estimation and accelerates the learning process. Deep reinforcement learning learns the optimal flight strategy, while the ant colony optimization and A algorithm respectively simulate pheromone transmission in ants foraging and utilize heuristic functions to find potential optimal paths in the path search space. Simultaneously, considering the UAV's energy consumption, flight safety constraints, and dynamic environmental factors, an energy model is established to correlate flight parameters with energy consumption. When planning the flight trajectory, low-energy paths are prioritized, and environmental perception data is used to avoid obstacles and dangerous areas. This enables the UAV to intelligently plan the optimal trajectory during flight, achieving long endurance and safe flight while efficiently tracking targets, greatly improving the system's practicality and reliability. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the main system framework of the present invention;

[0033] Figure 2 This is a flowchart illustrating the system usage of the present invention;

[0034] Figure 3 This is a simulation diagram of data acquisition for the present invention;

[0035] Figure 4 This is a simulation diagram of the normalized data processing of the present invention. Detailed Implementation

[0036] 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. Specific Implementation Example 1:

[0038] like Figure 1-4 As shown, an infrared target localization system for unmanned aerial vehicles (UAVs) integrating multi-model Kalman filtering and deep learning includes a main system. The main system comprises a data acquisition module, a multi-source data fusion module, a feature extraction and enhancement module, a dynamic model optimization module, a target state estimation module, a localization solution module, and a flight decision-making and planning module. The data acquisition module deploys a high-resolution infrared sensor to accurately capture the infrared radiation information of the target. It also integrates a high-definition visual camera and a high-precision lidar to collect environmental data from all directions. An adaptive sampling strategy based on fuzzy logic, particle swarm optimization, and adaptive thresholding is employed. This strategy dynamically adjusts the sensor's sampling frequency and resolution according to environmental complexity and target motion state. Fuzzy logic is used to evaluate the degree of fuzziness in environmental complexity and target motion state; particle swarm optimization is used to find the optimal solution among multiple possible combinations of sampling parameters; and the adaptive thresholding dynamically adjusts the sampling triggering conditions according to the rate of change of the data. This approach ensures data quality while reducing data redundancy and processing burden, guaranteeing data richness and completeness.

[0039] The multi-source data fusion module employs an adaptive fusion algorithm that integrates biological neuron collaboration mechanisms, tensor decomposition, and deep learning attention mechanisms. This algorithm simulates information interaction and integration between neurons and performs feature encoding on infrared, visual, and lidar data. Tensor decomposition breaks down multi-source data from a high-dimensional space into a low-dimensional subspace to extract the core features of the data. The deep learning attention mechanism dynamically adjusts the weights of each sensor's data based on dynamic changes in the environment through competitive learning and collaborative feedback, achieving deep fusion of multi-source data and providing a high-quality data foundation for subsequent processing. At the same time, cross-modal consistency constraints are introduced. By constructing a cross-modal similarity metric function based on Wasserstein distance, cosine similarity, and mutual information, the module ensures that the data from various modalities maintain consistency and complementarity during the fusion process.

[0040] The feature extraction enhancement module adopts a convolutional neural network architecture that integrates quantum enhancement, generative adversarial networks, and residual networks. It utilizes the superposition and entanglement properties of quantum states to accelerate the feature extraction process. The generator in the generative adversarial network generates simulated feature data, and the discriminator distinguishes between real and simulated features. Through adversarial training, more representative features are mined. The residual network solves the gradient vanishing problem in deep networks by skip connections, enhancing the network's ability to extract complex features. Combined with attention mechanisms and improved dilated convolution techniques, it focuses on key target features, expands the receptive field, and enhances the ability to extract target features in complex scenes.

[0041] The dynamic model optimization module introduces a model optimization strategy based on Generative Adversarial Networks (GANs), transfer learning, and genetic algorithms. The generator generates simulated target and scene data, the discriminator distinguishes between real and simulated data, and the generalization ability of the model is improved through adversarial training. Transfer learning transfers the model knowledge trained in similar scenarios to accelerate the convergence speed of the model in new scenarios. The genetic algorithm searches for the optimal model configuration in the model space by encoding, selecting, crossing over, and mutating the model's structure and parameters, thereby achieving rapid optimization and adaptive adjustment of the model.

[0042] The target state estimation module uses an algorithm that integrates reinforcement learning, adaptive Kalman filtering, and particle filtering to model the target state estimation problem as a Markov decision process. By using a reward function for the UAV target localization scenario, the agent is guided to learn the optimal state estimation strategy. The adaptive Kalman filter dynamically adjusts the filtering parameters according to the rapid changes in the target motion and the interference of complex environmental noise. The particle filter estimates the target state by sampling a large number of particles and updating the weights.

[0043] The positioning and calculation module is based on the principle of multi-view data fusion and improved triangulation. It combines the real-time attitude information of the UAV with the mechanism of dynamic coordinate transformation, error compensation and least squares fitting. By monitoring the flight attitude and environmental parameter changes of the UAV in real time, the coordinate system is dynamically adjusted in the positioning process and the measurement error is compensated online. Least squares fitting is used to find the optimal positioning solution in multiple sets of measurement data. It adapts to the flight attitude changes of the UAV and the dynamic interference of the environment in real time, and achieves high-precision three-dimensional positioning of the target.

[0044] The flight decision-making and planning module employs a dual-deep Q-network architecture based on deep reinforcement learning, ant colony optimization, and A algorithm. Combined with experience replay and priority experience replay mechanisms, it effectively reduces overfitting in Q-value estimation and accelerates the learning process. Deep reinforcement learning learns the optimal flight strategy. The ant colony optimization algorithm simulates pheromone transmission during ant foraging to find potential optimal paths in the path search space. The A algorithm utilizes heuristic functions to quickly search for the optimal path from the current position to the target position. Simultaneously, it considers the UAV's energy consumption, flight safety constraints, and dynamic environmental factors to plan the optimal flight trajectory in real time. This ensures that the UAV can efficiently track targets while achieving long endurance and safe flight.

[0045] In the multi-source data fusion module, an adaptive fusion algorithm based on the biological neuron collaboration mechanism, tensor decomposition, and deep learning attention mechanism constructs a neuron connection weight matrix to simulate the excitation and inhibition relationship between neurons. It combines the low-dimensional feature representation after tensor decomposition and the weight allocation of the attention mechanism to achieve efficient fusion of sensor data, thereby improving the system's adaptability to complex and changing environments and the accuracy of target perception.

[0046] In the feature extraction enhancement module, the parallel computing capability of qubits simultaneously explores the optimal solution direction during the parameter update of the convolution kernel. The adversarial training mechanism of the generative adversarial network prompts the network to mine representative features. Among them, the skip connections of the residual network ensure the effective extraction of deep features and accelerate model convergence. Moreover, it can mine deep target features that are difficult to discover using traditional methods.

[0047] In the dynamic model optimization module, the generative adversarial network (GAN) trains against adversarial forces to make the data distribution generated by the generator gradually approach the distribution of real data. Transfer learning reduces the training time and sample requirements of the model in new scenarios through feature mapping and knowledge transfer between the source and target domains. The genetic algorithm further improves the performance and adaptability of the model by optimizing the model structure and parameters.

[0048] Under the conditions of target motion patterns and environmental noise, the target state estimation module uses reinforcement learning to guide the agent to select the optimal estimation strategy. Adaptive Kalman filtering quickly adjusts the filtering parameters, and particle filtering improves estimation accuracy through particle sampling and weight updates, keeping the estimation error within a small range. This effectively improves the stability and reliability of target state estimation.

[0049] When considering the energy consumption of the UAV, the flight decision-making and planning module establishes an energy model that correlates factors such as flight speed, altitude, and attitude with energy consumption. It uses deep reinforcement learning, ant colony algorithm, and Algorithm A to prioritize low-energy-consumption paths when planning flight trajectories. When considering flight safety constraints, it combines environmental perception data to avoid obstacles and dangerous areas, ensuring the safety of UAV flight.

[0050] When collecting data, the data acquisition module dynamically and intelligently adjusts the sampling frequency and resolution of the sensor according to the real-time changes in environmental complexity and target motion state, so as to minimize data redundancy and processing burden while ensuring data quality.

[0051] In the process of data fusion, the multi-source data fusion module introduces a cross-modal similarity measurement function based on Wasserstein distance, cosine similarity, and mutual information. This function measures the similarity and complementarity between different modalities from various perspectives, ensuring that different modalities maintain a high degree of consistency and complementarity of information during the fusion process, and further improving the quality and reliability of the fused data.

[0052] The different modal data involved here specifically include infrared radiation data collected by infrared sensors, which can reflect the thermal characteristic distribution of the target and play a key role in identifying the target's heat-generating parts and temperature differences; visible light image data acquired by high-definition vision cameras, which can present the target's shape, color, texture and other appearance features, helping to identify the target's category and details from a visual perspective; and distance data collected by high-precision lidar, which can construct three-dimensional spatial structure information of the target and its surrounding environment, accurately depicting the target's position and relative distance to surrounding objects.

[0053] This module encodes features from the aforementioned different modalities of data by simulating information interaction and integration between neurons. Tensor decomposition is used to decompose multi-source data from a high-dimensional space into a low-dimensional subspace, extracting the core features of the data. The deep learning attention mechanism dynamically adjusts the weights of each sensor's data based on dynamic environmental changes, such as weather changes, lighting changes, and targets entering different scenes, through competitive learning and collaborative feedback. This achieves deep fusion of multi-source data, providing a high-quality data foundation for subsequent processing.

[0054] Simultaneously, cross-modal consistency constraints are introduced. By constructing a cross-modal similarity metric function based on Wasserstein distance, cosine similarity, and mutual information, the consistency and complementarity of information between different modalities are ensured during the fusion process. Specifically, Wasserstein distance is used to measure the differences between the distributions of different modalities to ensure the consistency of the data distribution in the overall data; cosine similarity measures the similarity between feature vectors of different modalities from the perspective of the angle between vectors, capturing the similarity of features; and mutual information evaluates the degree of information sharing between different modalities, highlighting the complementarity of the data. Specific Implementation Example 2:

[0056] like Figure 1-4 As shown, the following is the specific usage process of the content mentioned in Example 1 in a real-world application scenario:

[0057] 1. Data Acquisition Phase:

[0058] After takeoff, the data acquisition module begins operation. A high-resolution infrared sensor continuously scans the surrounding environment, accurately capturing infrared radiation information emitted by targets. For example, in nighttime or smoky environments, it can clearly detect the location and approximate outline of heat-generating targets. Simultaneously, a high-definition vision camera captures visible light images, obtaining details of the target's appearance. If the target is a vehicle, it can identify its color, shape, and other features. A high-precision lidar emits a laser beam to measure the distance to the target and surrounding objects, constructing a three-dimensional spatial model.

[0059] An adaptive sampling strategy based on fuzzy logic, particle swarm optimization, and adaptive thresholding operates in real time. The system utilizes fuzzy logic to assess the complexity of the environment and the fuzziness of the target's motion state based on factors such as the density of objects, changes in lighting, and changes in the target's speed and direction. For example, when a drone enters a high-rise urban area, the environmental complexity increases, and the fuzzy logic outputs a corresponding result. The particle swarm optimization algorithm, based on the fuzzy logic's evaluation results, quickly searches for the optimal sampling parameters from numerous possible combinations of sampling frequency and resolution. If a sudden acceleration of the target's motion is detected, the particle swarm optimization adjusts the sampling frequency to capture the target's dynamics at a higher frequency. The adaptive thresholding dynamically determines whether to trigger sampling based on the rate of change of data such as infrared radiation intensity, changes in target features in the visual image, and fluctuations in LiDAR distance data, ensuring the acquisition of key data while avoiding data redundancy caused by over-collection.

[0060] 2. Multi-source data fusion stage:

[0061] The multi-source data fusion module receives infrared, visual, and lidar data from the data acquisition module. An adaptive fusion algorithm based on biological neuron collaboration mechanisms, tensor decomposition, and deep learning attention mechanisms is initiated. First, it simulates information interaction between neurons, encodes features from different modalities, and transforms the data into a unified, processable format.

[0062] Tensor decomposition algorithms break down high-dimensional, multi-source data into low-dimensional subspaces, extracting core features and reducing data processing volume while preserving key information. For example, they can integrate and refine features from 3D LiDAR data and 2D visual image data.

[0063] Deep learning attention mechanisms dynamically adjust the weights of data from various sensors based on environmental changes such as weather variations, lighting conditions, and the target entering different scenes. For example, in strong light environments, the weight of visual camera data is appropriately reduced, while the weights of infrared sensor and LiDAR data are increased. Simultaneously, by constructing a cross-modal similarity metric function based on Wasserstein distance, cosine similarity, and mutual information, it ensures that different modalities maintain high consistency and complementarity in terms of overall distribution (guaranteed by Wasserstein distance), feature vector similarity (measured by cosine similarity), and information sharing (evaluated by mutual information) during the fusion process, generating high-quality fused data.

[0064] 3. Feature extraction enhancement stage:

[0065] The feature extraction and enhancement module acquires the fused data and processes it using a convolutional neural network architecture that integrates quantum enhancement, generative adversarial networks, and residual networks. The quantum enhancement part leverages the superposition and entanglement properties of quantum states, utilizing the parallel computing power of qubits during convolution operations to simultaneously explore the optimal solution directions for updating the parameters of multiple convolution kernels, greatly accelerating the feature extraction process.

[0066] In Generative Adversarial Networks (GANs), the generator produces simulated feature data, while the discriminator strives to distinguish between real and simulated features. Through continuous adversarial training, the network is encouraged to uncover more representative target features. For example, it can accurately extract unique texture, shape, and other features for different types of targets. Residual networks, through skip connections, solve the vanishing gradient problem in deep network training, ensuring the network's effective extraction of deep features. Combined with attention mechanisms and improved dilated convolution techniques, they focus on key target features, expand the receptive field, and thus enhance the ability to extract target features in complex scenes, such as accurately identifying key parts of a target in a cluttered background.

[0067] 4. Dynamic model optimization stage:

[0068] The dynamic model optimization module employs a model optimization strategy based on Generative Adversarial Networks (GANs), transfer learning, and genetic algorithms. The generator of the GAN generates simulated target and scene data, while the discriminator distinguishes between real and simulated data. Through continuous adversarial training, the data distribution generated by the generator gradually approximates the real data distribution, improving the model's generalization ability for various targets and scenes.

[0069] Transfer learning transfers knowledge from models trained in similar scenarios (such as similar geographical environments, lighting conditions, and target motion patterns) to the current model, accelerating the model's convergence speed in new scenarios. For example, a model previously trained in a mountainous environment can transfer some of its knowledge to scenarios with similar terrain. Genetic algorithms encode, select, crossover, and mutate the model's structure (such as the number of network layers and neurons) and parameters (such as weights and biases), searching for the optimal model configuration in the model space to achieve rapid optimization and adaptive adjustment of the model to adapt to different target localization requirements.

[0070] 5. Target State Estimation Stage:

[0071] The target state estimation module employs an algorithm that integrates reinforcement learning, adaptive Kalman filtering, and particle filtering. The target state estimation problem is modeled as a Markov decision process. Through a carefully designed reward function tailored to the UAV target localization scenario, the agent is guided to learn the optimal state estimation strategy under different states (such as different stages of target motion or different relative positions between the UAV and the target).

[0072] When the target's motion state changes rapidly (e.g., sudden acceleration or turning) or is affected by complex environmental noise (e.g., electromagnetic interference or weather noise), the adaptive Kalman filter quickly and dynamically adjusts the filter parameters to maintain an accurate estimate of the target's state. Particle filtering, through sampling and weight updates of a large number of particles, further refines the target's state estimation. For example, even when the target is briefly obscured, the distribution and weight changes of the particles can still reasonably infer the target's position and state.

[0073] 6. Location and Solution Stage:

[0074] The positioning and calculation module is based on multi-view data fusion and improved triangulation principles, combined with real-time UAV attitude information (including pitch, yaw, and roll angles). By monitoring the UAV's flight attitude and changes in environmental parameters (such as wind speed and air pressure) in real time, a dynamic coordinate transformation mechanism dynamically adjusts the coordinate system during the positioning process to ensure positioning accuracy.

[0075] The error compensation mechanism compensates for sensor measurement and calculation errors online. Least squares fitting searches for the optimal positioning solution among multiple sets of measurement data, comprehensively considering various factors to achieve high-precision 3D positioning of the target. For example, even when a UAV continuously adjusts its attitude and position during flight, it can always accurately calculate the target's 3D coordinates.

[0076] 7. Flight Decision-Making and Planning Phase:

[0077] The flight decision-making and planning module employs a dual-deep Q-network architecture based on deep reinforcement learning, ant colony optimization, and A algorithm. By combining experience replay and priority experience replay mechanisms, overfitting in Q-value estimation is reduced, accelerating the learning process. Deep reinforcement learning learns the optimal flight strategy based on different environments (such as urban environments, mountainous environments, and severe weather environments).

[0078] Ant colony optimization (ACO) searches for potential optimal paths in the path search space by simulating pheromone transfer during ant foraging. ACO utilizes a heuristic function to quickly search for the optimal path from the current position to the target position. Simultaneously, it considers the drone's energy consumption (related to flight speed, altitude, attitude, and air resistance), flight safety constraints (such as avoiding no-fly zones, obstacles, and severe weather areas), and dynamic environmental factors (such as wind direction changes and sudden changes in the target's direction of movement), planning the optimal flight trajectory in real time to ensure the drone achieves long endurance and safe flight while efficiently tracking the target. For example, when the drone's battery is low, it prioritizes planning a low-energy-consumption return path; when encountering obstacles ahead, it promptly adjusts the flight path to avoid them. Specific Implementation Example 3:

[0080] like Figure 1-4 As shown, the following is a supplement to the content of Example 1:

[0081] Fuzzy logic membership functions: When assessing environmental complexity, taking an urban environment as an example, factors such as building density and road density can be considered. Assuming building density is represented by the number of buildings per square kilometer, a number less than 50 is defined as "low complexity" with a membership degree of 0-0.3; a number between 50 and 150 is defined as "medium complexity" with a membership degree of 0.3-0.7; and a number greater than 150 is defined as "high complexity" with a membership degree of 0.7-1. For target motion, taking target speed as an example, a speed less than 10 m / s is defined as "low speed" with a membership degree of 0-0.3; a speed between 10-30 m / s is defined as "medium speed" with a membership degree of 0.3-0.7; and a speed greater than 30 m / s is defined as "high speed" with a membership degree of 0.7-1. These membership function divisions can be adjusted based on extensive experimental data and experience in real-world scenarios.

[0082] Particle swarm optimization (PSO) algorithm parameters: The number of particles can be between 20 and 100. For example, when optimizing data acquisition parameters for a small area, the number of particles can be set to 30, and the inertia weight can be adjusted between 0.4 and 0.9. When fast convergence is required, the inertia weight can be set to 0.4; if a wider search space is desired, it can be set to 0.9. Learning factors are typically divided into cognitive learning factors and social learning factors, both generally ranging from 1 to 4. For example, a cognitive learning factor of 1.5 and a social learning factor of 1.8 determine the degree to which particles move towards their historical best position and the global best position.

[0083] Adaptive Kalman filter noise covariance matrix: During initialization, for the process noise covariance matrix, if the target motion is relatively stable, its diagonal elements can be set to small values, such as [0.01, 0.01, 0.01] (assuming the state vector is three-dimensional, representing the target's position, velocity, and acceleration respectively). For the measurement noise covariance matrix, it is determined according to the sensor accuracy. If the infrared sensor has high accuracy, its corresponding matrix elements can be set to 0.001, while the elements corresponding to sensors with slightly lower accuracy can be set to 0.01. During the update process, adjustments can be made according to changes in the uncertainty of the target motion. When the target motion suddenly accelerates or is subject to significant disturbance, the matrix value should be appropriately increased to better track changes in the target state.

[0084] Convolutional Neural Network Architecture Parameters: In the feature extraction enhancement module, assume a network is built to identify small targets. The network can be set to 10 layers, including 5 convolutional layers and 5 fully connected layers. The kernel size of the convolutional layers can gradually increase from 3x3 to 5x5 to obtain features at different scales; for example, the first three convolutional layers have a kernel size of 3x3, and the last two have a kernel size of 5x5. The number of neurons in each layer can be calculated based on the kernel size and the number of input channels to ensure the effectiveness of feature extraction. The number of neurons in the fully connected layers can be set to 512, 256, 128, 64, and 10, respectively, for classifying and outputting the extracted features.

[0085] Genetic algorithms encode model structures as follows: For convolutional neural network structures, binary encoding can be used. For example, the number of network layers can be represented by several bits; assuming 4 bits, it can represent networks with 1-16 layers. For the type of each layer (convolutional layer or fully connected layer), 1 bit is used, with 0 representing a convolutional layer and 1 representing a fully connected layer. For the kernel size, if only 3x3 and 5x5 are considered, 1 bit is used, with 0 representing 3x3 and 1 representing 5x5. The number of neurons can be encoded according to its value range; for example, the range of the number of neurons can be divided into several intervals, and the corresponding binary number can be used to represent the interval.

[0086] Reward Function Design: In the UAV target localization scenario of the target state estimation module, the reward function can be designed as follows: A reward of +10 is given when the UAV's target localization error is less than 0.5 meters; a reward of +5 is given when the error is between 0.5 and 1 meter; and a reward of -5 is given when the error is greater than 1 meter. If the UAV approaches a dangerous area (such as near a high-voltage line) during flight, a reward of -20 is given. When the UAV successfully and continuously tracks the target for a certain period of time (such as 30 seconds), a reward of +15 is given. Simultaneously, considering the UAV's energy consumption, a reward of -3 is given for every certain percentage (such as 5%) of battery power consumed. This reward function comprehensively considers factors such as localization accuracy, safety, and energy consumption, guiding the agent to learn the optimal state estimation strategy. Specific Implementation Example 4:

[0088] like Figure 1-4 As shown, the key algorithm mentioned in Example 1 will be analyzed in detail below, including its core mathematical formulas and explanations:

[0089] 1. Data Acquisition Module

[0090] Fuzzy logic: Fuzzy logic is used to evaluate the complexity of an environment and the degree of fuzziness in a self-defined motion state. Taking the evaluation of environmental complexity as an example, assuming there are multiple influencing factors, such as object density D, light change rate L, and scene change frequency S, these factors are mapped to the [0,1] interval using fuzzy membership functions. For example, for object density D, the membership function μ... D (x) can be represented as:

[0091]

[0092] Among them, D min and D max These represent the minimum and maximum values ​​of the object's density, respectively. Then, using fuzzy inference rules, such as the weighted average method, the fuzzy value F of the environmental complexity is calculated.

[0093] F=w1μ D (D)+w2μ L (L)+w3μ S (S)

[0094] Here, w1, w2, and w3 represent the weights of each factor. In the system, this fuzzy value is used as a basis for adjusting the sampling parameters in the subsequent particle swarm optimization algorithm, helping the system to reasonably adjust the data acquisition frequency and resolution according to the complexity of the environment.

[0095] Particle Swarm Optimization (PSO) Algorithm: The PSO algorithm is used to find the optimal solution among a variety of possible combinations of sampled parameters. Each particle i has a position x in the n-dimensional search space. i =(x i1 ,x i2,…,x in ) represents a set of sampled parameters, velocity v i =(v i1 ,v i2 ,…,v in Particles update their position and velocity by tracking two extreme values: the individual extreme value (pbest). i And global extremum gbest.

[0096] The speed update formula is:

[0097] v ij (t+1)=ωv ij (t)+c1r1(p ij -x ij (t))+c2r2(g j -x ij (t))

[0098] Where ω is the inertia weight, c1 and c2 are learning factors, and r1 and r2 are random numbers between [0,1]. The position update formula is:

[0099] x ij (t+1)=x ij (t)+v ij (t+1)

[0100] In the system, the particle swarm optimization algorithm searches for the optimal combination of parameters such as sampling frequency and resolution based on the environmental complexity and target motion state evaluated by fuzzy logic, in order to balance data quality and processing burden.

[0101] 2. Multi-source data fusion module

[0102] Tensor decomposition: Tensor decomposition is used to decompose multi-source data from a high-dimensional space into low-dimensional subspaces to extract the core features of the data. Using a third-order tensor... For example, a commonly used form of CP decomposition is:

[0103]

[0104] Among them, a r ∈R I ,b r ∈R J ,c r ∈R K For factor vectors, Let R denote the vector outer product, and R be the rank of the decomposition. Through tensor decomposition, high-dimensional multi-source data (such as tensors composed of infrared, vision, and lidar data) are decomposed into low-dimensional vector combinations, extracting the core features of each modality data to provide a more concise and effective data representation for subsequent fusion.

[0105] Deep learning attention mechanism: The deep learning attention mechanism dynamically adjusts the weights of data from each sensor based on dynamic changes in the environment. Taking the self-attention mechanism as an example, for the input feature sequence Q, K, V, the attention output is:

[0106]

[0107] Where, d k Let K be the dimension of the key vector. In the system, features from different modalities are used as Q, K, and V inputs to the attention mechanism, respectively, to calculate the weights of each modality, thus achieving dynamic fusion of multi-source data. For example, when changes in ambient lighting cause a decrease in the quality of visual data, the attention mechanism will reduce the weight of visual data and relatively increase the weight of infrared and lidar data.

[0108] Cross-modal similarity measurement function:

[0109] Wasserstein distance: Used to measure the difference between different modal data distributions to ensure the consistency of the data distribution in the overall distribution. For two probability distributions P and Q, the one-dimensional Wasserstein distance (also known as Earth-Mover's distance) is defined as:

[0110]

[0111] in, It is the set of all joint distributions of P and Q. In the system, the similarity between the feature distributions of different modalities is determined by calculating the Wasserstein distance, ensuring that the fused data maintains consistency in distribution.

[0112] Cosine similarity: Measured from the perspective of the angle between vectors, it reflects the similarity between feature vectors of different modalities, capturing the similarity of features. For two feature vectors a and b, the cosine similarity is defined as:

[0113]

[0114] In the system, feature vectors extracted from different modalities are input into cosine similarity calculation to evaluate their similarity, which helps to preserve similar features during the fusion process and enhance the stability of the fused data.

[0115] Mutual information: assesses the degree of information sharing between data from different modalities, highlighting the complementarity of the data. For two random variables X and Y, mutual information is defined as:

[0116]

[0117] Here, p(x,y) is the joint probability distribution of X and Y, and p(x) and p(y) are their marginal probability distributions. In the system, by calculating the mutual information of data from different modalities, complementary information between them is discovered, thereby improving the quality of the fused data.

[0118] 3. Feature Extraction Enhancement Module

[0119] Quantum enhancement: Utilizing the superposition and entanglement properties of quantum states to accelerate the feature extraction process. In quantum computing, a qubit can exist in a superposition of 0 and 1, such as |ψ>=α|0>+β|1>, where α and β are complex numbers, and |α| 2 +|β| 2 =1. During the convolution kernel parameter update process, the parallel computing capability of qubits allows for the simultaneous exploration of multiple parameter update directions. Rapid parameter optimization is achieved through quantum gate operations, thereby accelerating the feature extraction process. For example, by transforming qubits through quantum gate operations such as the Hadamard gate, parallel updates of the convolution kernel parameters can be achieved. Compared to traditional methods, this approach can find the optimal parameters faster and improve feature extraction efficiency.

[0120] Generative Adversarial Networks (GANs): In GANs, a generator produces simulated feature data, and a discriminator distinguishes between real and simulated features. Through adversarial training, more representative features are extracted. The objective functions of the generator G and the discriminator D are as follows:

[0121]

[0122] Where, p data (x) represents the real data distribution. The generator attempts to generate simulated data similar to the real data, causing the discriminator to misjudge. The discriminator, on the other hand, strives to correctly distinguish between real and simulated data. In the system, the generative adversarial network continuously challenges the training, prompting the generator to generate data that is closer to the real target features. In this process, the discriminator learns to better identify real features, thereby uncovering more representative target features.

[0123] Residual networks: Residual networks address the vanishing gradient problem in deep networks through skip connections, enhancing the network's ability to extract complex features. In a residual network, assuming the input is x, and the output after a series of convolutional layers is F(x), then the output of the residual block is:

[0124] y = F(x) + x

[0125] This skip connection method allows gradients to propagate more smoothly in the backpropagation process, even as the number of network layers increases, avoiding the vanishing gradient problem. In the system, residual networks are used for deep feature extraction from the fused data, effectively extracting key features of targets in complex scenes, such as accurately identifying subtle textures and structural features of targets in complex backgrounds.

[0126] 4. Target State Estimation Module

[0127] Adaptive Kalman Filtering: Adaptive Kalman filtering dynamically adjusts its filtering parameters based on rapid changes in target motion and complex environmental noise interference. The prediction equation for standard Kalman filtering is:

[0128]

[0129] P k|k-1 =AP k-1|k-1 A T +Q

[0130] The update equation is:

[0131] K k =P k|k-1 H T HP k|k-1 H T +R) -1

[0132]

[0133] P k|k =(IK k H)P k|k-1

[0134] in, It is the predicted state at time k. Let A be the updated state at time k, A be the state transition matrix, B be the control input matrix, and u be the input state. k It is the control input, P k|k-1 It is the prediction coerror, P k |k is the update covariance, Q is the process noise covariance, K k It is the Kalman gain, H is the observation matrix, and z is the Kalman gain. k Q is the observed value, and R is the observation noise covariance. In adaptive Kalman filtering, the values ​​of Q and R are dynamically adjusted based on the uncertainty of target motion and changes in environmental noise. For example, when the target motion suddenly accelerates or is subject to significant disturbance, the value of Q is increased to better track changes in the target state; when sensor measurement noise increases, the value of R is increased, and the weight of the observed value is reduced.

[0135] Particle filtering: Particle filtering estimates the target state by sampling a large number of particles and updating the weights. Let the target state be x. k The observed value is z k First, based on the prior distribution p(x0) for particles... Initialize the data, i = 1, ..., N, where N is the number of particles. At time k, according to the state transition probability p(x...k |x k-1 Sampling of particles yields... Then, based on the observed value z k Calculate the weights of the particles:

[0136]

[0137] Finally, the particles are resampled, retaining those with higher weights and discarding those with lower weights, resulting in a new particle set. In this system, particle filtering, through sampling and weight updates of a large number of particles, can handle non-Gaussian and nonlinear target state estimation problems. It accurately estimates the target state in complex environments, such as when the target is briefly occluded or has a complex trajectory, and can still reasonably infer the target's position and state by analyzing particle distribution and weight changes.

[0138] 5. Location Solving Module

[0139] Dynamic coordinate transformation: By monitoring the UAV's flight attitude and environmental parameter changes in real time, the coordinate system is dynamically adjusted during the positioning process. Assume the UAV's attitude is represented by a rotation matrix R, and the position vector is t. Under different flight attitudes, the target's coordinates x in the UAV's local coordinate system are transformed. local Transform to global coordinate system x global The formula is:

[0140] x global =Rx local +t

[0141] Wherein, the rotation matrix R is based on the pitch angle θ and yaw angle of the UAV. The roll angle ω is calculated. In the system, dynamic coordinate transformation ensures that the target's position information can be accurately converted to a unified global coordinate system even as the UAV's flight attitude constantly changes, providing an accurate data foundation for subsequent positioning calculations.

[0142] Error compensation: Online compensation for measurement errors. Assuming the measurement error is ∈ Z, the measured value is z, and the true value is x, then the estimated value after error compensation is... for:

[0143]

[0144] Error ∈ can be determined through sensor calibration data, statistical analysis of historical measurement data, or model-based error prediction. In the system, error compensation mechanisms can improve the accuracy of measurement data and reduce the impact of sensor errors on target positioning.

[0145] Least squares fitting: used to find the optimal localization solution among multiple sets of measurement data. Assume we have n sets of measurement data (x...i ,y i ,z i The least squares method aims to find a set of parameters, along with a model function f(x,y,z; θ) to be fitted. To minimize the sum of squared errors between the measured values ​​and the model predictions, i.e.:

[0146]

[0147] Where d i This refers to the actual distance or position information corresponding to the measurement data. In triangulation-based positioning, f(x,y,z; θ) may be a distance function related to the target position (x,y,z) and the UAV's position and attitude. By minimizing the sum of squared errors mentioned above, the optimal position estimate of the target is determined. In the system, the least squares fitting method comprehensively considers the measurement information after multi-view data fusion, combined with the UAV's real-time attitude, and adapts in real-time to changes in the UAV's flight attitude and dynamic environmental interference, achieving high-precision three-dimensional positioning of the target.

[0148] 6. Flight Decision Planning Module

[0149] Deep reinforcement learning: In a dual-deep Q-network architecture, the agent learns the optimal flight strategy through interaction with the environment. The core of Q-learning is learning a Q-function Q(s,a), which represents the long-term cumulative reward expectation of taking action a in state s. The update formula for the Q-function is:

[0150]

[0151] Where α is the learning rate, γ is the discount factor, and r t In state s t Take action a t The reward obtained later, s t +1 represents the next state. In the dual-deep Q-network, two neural networks are introduced: one to estimate the current Q-value (online network) and the other to estimate the target Q-value (target network) to reduce the overfitting problem in Q-value estimation. In the system, deep reinforcement learning learns to select the optimal flight actions, such as changing flight speed, direction, and altitude, based on the current position of the UAV, the target position, energy state, and environmental information, in order to efficiently track the target and meet energy and safety constraints.

[0152] Ant colony optimization (ACO) algorithm: The ACO algorithm searches for potential optimal paths in a path search space by simulating pheromone transfer during ant foraging. Ants release pheromones along the path, and the concentration of these pheromones influences the path choices of subsequent ants. Let C be the set of cities (or location nodes), and let p be the probability that ant k at position i will move to position j.i j k for:

[0153]

[0154] Where τ ij It is the pheromone concentration along the path from position i to position j, η ij =1 / d ij It is heuristic information (d ij J is the distance from position i to position j, α and β represent the relative importance of pheromones and heuristic information, respectively. k This is the set of possible positions for ant k to move to next. As the ant continues to move, the pheromone is updated according to the following formula:

[0155] τ ij =(1-ρ)τ ij +Δτ ij

[0156]

[0157] Where ρ is the pheromone evaporation rate. This represents the amount of pheromone left by ant k on path (i,j). In the system, the ant colony algorithm gradually explores potential optimal flight paths in the UAV's path search space through the accumulation and volatilization of pheromones. Considering path length, safety, and collaboration with other algorithms, it plans a reasonable flight trajectory for the UAV.

[0158] Algorithm A: Algorithm A uses a heuristic function to quickly find the optimal path from the current position to the target position.

[0159] The search process of Algorithm A is based on an evaluation function f(n) = g(n) + h(n), where g(n) is the actual cost from the starting node to node n, and h(n) is the estimated cost (heuristic function) from node n to the target node.

[0160] Algorithm A selects the node with the smallest f(n) value for expansion. For example, in a two-dimensional plane, if Euclidean distance is used as the heuristic function... Where (x) n ,y n ) is the coordinate of node n, (x goal ,y goal() represents the coordinates of the target node. In the system, Algorithm A combines the current position of the UAV and the target position to quickly search for a theoretically optimal path. At the same time, it considers flight safety constraints (such as avoiding obstacles), providing an important reference for the UAV's flight decisions. In collaboration with deep reinforcement learning and ant colony algorithms, it plans the optimal flight trajectory in real time to meet various conditions, ensuring that the UAV can achieve long endurance and safe flight while efficiently tracking the target.

[0161] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising a reference structure" does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0162] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An unmanned aerial vehicle infrared target positioning system fusing a multi-model KF algorithm and deep learning, comprising a main system, characterized in that: The main system comprises a data acquisition module, a multi-source data fusion module, a feature extraction enhancement module, a dynamic model optimization module, a target state estimation module, a positioning calculation module and a flight decision planning module, the data acquisition module deploys a high-resolution infrared sensor to accurately capture target infrared radiation information, and simultaneously integrates a high-definition visual camera and a high-precision laser radar to collect environmental data in all directions, and adopts an adaptive sampling strategy based on fuzzy logic, particle swarm optimization and adaptive threshold to dynamically adjust the sampling frequency and resolution of the sensor according to the environmental complexity and target motion state; The multi-source data fusion module uses an adaptive fusion algorithm that combines the biological neuron collaborative mechanism, tensor decomposition and deep learning attention mechanism to simulate the information interaction and integration between neurons, and encodes the features of infrared, visual and laser radar data, wherein the tensor decomposition decomposes the multi-source data from a high-dimensional space into a low-dimensional subspace to extract the core features of the data, and the deep learning attention mechanism dynamically adjusts the data weights of each sensor through competitive learning and collaborative feedback according to the dynamic changes of the environment, realizes the deep fusion of multi-source data, and provides a high-quality data basis for subsequent processing, while introducing a cross-modal consistency constraint to ensure the consistency and complementarity of information in the fusion process by constructing a cross-modal similarity measurement function based on the Wasserstein distance, cosine similarity and mutual information; The feature extraction enhancement module adopts a convolutional neural network architecture that combines quantum enhancement, generative adversarial networks and residual networks, uses the superposition and entanglement characteristics of quantum states to accelerate the feature extraction process, the generator in the generative adversarial network generates simulated feature data, and the discriminator distinguishes between real and simulated features, and through adversarial training, more representative features are extracted, the residual network solves the gradient vanishing problem in deep networks through skip connections to enhance the network's ability to extract complex features, and combines attention mechanisms and improved dilated convolution technology to focus on key target features, expand the receptive field and enhance the ability to extract target features in complex scenes; The dynamic model optimization module introduces a model optimization strategy based on generative adversarial networks, transfer learning and genetic algorithms, wherein the generator generates simulated target and scene data, and the discriminator distinguishes between real and simulated data to improve the model's generalization ability through adversarial training, the transfer learning migrates model knowledge trained in similar scenes to accelerate the convergence speed of the model in new scenes, and the genetic algorithm encodes, selects, crosses and mutates the structure and parameters of the model to search for the optimal model configuration in the model space, realizing fast optimization and adaptive adjustment of the model. The target state estimation module uses the algorithm of fusion reinforcement learning, adaptive Kalman filtering and particle filtering to model the target state estimation problem as a Markov decision process, guide the agent to learn the optimal state estimation strategy through the reward function for the unmanned aerial vehicle target positioning scene, the adaptive Kalman filtering dynamically adjusts the filtering parameters according to the rapid change of target motion and complex environmental noise interference, and the particle filtering estimates the target state through a large number of particle sampling and weight updating; The positioning calculation module is based on multi-view data fusion and improved triangulation principle, combines real-time attitude information of unmanned aerial vehicle, and fuses dynamic coordinate transformation, error compensation and least square fitting mechanism, dynamically adjusts the coordinate system in the positioning process through real-time monitoring of the flight attitude and environmental parameter changes of the unmanned aerial vehicle, and compensates the measurement error online, and the least square fitting is used to find the optimal positioning solution in a plurality of measurement data, and the flight attitude change and environmental dynamic interference of the unmanned aerial vehicle are adapted in real time, and high-precision three-dimensional positioning of the target is realized. The flight decision planning module adopts a double deep Q network architecture based on deep reinforcement learning, ant colony algorithm and A algorithm, combines experience replay and priority experience replay mechanism, effectively reduces the overfitting problem of Q value estimation, speeds up the learning process, the deep reinforcement learning learns the optimal flight strategy, the ant colony algorithm finds the potential optimal path in the path search space by simulating the pheromone transmission in the ant foraging process, and the A algorithm uses heuristic function to quickly search the optimal path from the current position to the target position, while considering the energy consumption, flight safety constraint and environmental dynamic factors of the unmanned aerial vehicle, and plans the optimal flight trajectory in real time.

2. The unmanned aerial vehicle infrared target positioning system of claim 1, wherein: In the multi-source data fusion module, the adaptive fusion algorithm based on biological neuron cooperation mechanism, tensor decomposition and deep learning attention mechanism simulates the excitation and inhibition relationship between neurons by constructing a neuron connection weight matrix, and realizes efficient fusion of sensor data by combining low-dimensional feature representation after tensor decomposition and weight distribution of attention mechanism, and improves the adaptability of the system to complex and variable environment and the accuracy of target perception. 3.The UAV infrared target positioning system of claim 1, wherein: In the feature extraction reinforcement module, the parallel computing capability of the quantum bits explores the optimal solution direction in the parameter updating process of the convolution kernel simultaneously, and the adversarial training mechanism of the generative adversarial network promotes the network to mine representative features, and the skip connection of the residual network ensures the effective extraction of deep features of the network and accelerates the convergence of the model.

4. The unmanned aerial vehicle infrared target positioning system of claim 1, wherein: In the dynamic model optimization module, the generative adversarial network gradually approaches the real data distribution through adversarial training, the transfer learning reduces the training time and sample demand of the model in the new scene through feature mapping and knowledge transfer between the source domain and the target domain, and the genetic algorithm further improves the performance and adaptability of the model through optimization search of the model structure and parameters.

5. The unmanned aerial vehicle infrared target positioning system of claim 1, wherein: The target state estimation module strengthens the learning of the intelligent agent to select the optimal estimation strategy under the target motion mode and environmental noise conditions, and the adaptive Kalman filter quickly adjusts the filter parameters, wherein the particle filter improves the estimation accuracy through particle sampling and weight updating, so that the estimation error is kept within a small range. 6.The UAV infrared target positioning system of claim 1, wherein: The flight decision planning module considers the energy consumption of the unmanned aerial vehicle, associates factors such as flight speed, height and attitude with energy consumption by establishing an energy model, and uses deep reinforcement learning, ant colony algorithm and A algorithm to plan the flight trajectory, preferentially selecting a path with low energy consumption, and in consideration of flight safety constraints, combining environmental perception data to avoid obstacles and dangerous areas, thereby ensuring the safety of the unmanned aerial vehicle flight.

7. The unmanned aerial vehicle infrared target positioning system of claim 1, wherein: The data acquisition module dynamically and intelligently adjusts the sampling frequency and resolution of the sensor according to the real-time changes in the complexity of the environment and the target motion state when collecting data, thereby ensuring data quality while minimizing data redundancy and processing burden. 8.The UAV infrared target positioning system of claim 1, wherein: The multi-source data fusion module introduces a cross-modal similarity measurement function based on Wasserstein distance, cosine similarity and mutual information during data fusion, which measures the similarity and complementarity between different modal data from various angles, ensures the high consistency and complementarity of information during the fusion of different modal data, and further improves the quality and reliability of the fused data.

Images