An Active Optoelectronic Detection and Tracking Method for Unmanned Aerial Vehicles Based on Radio Reverse Protocol Parsing Technology

By analyzing UAV communication messages using radio reverse protocol parsing technology and combining this with unscented Kalman filtering to predict target displacement, the optoelectronic system is driven to actively track the UAV. This solves the problem of traditional technologies being unable to analyze the UAV's flight intentions and achieves efficient low-altitude security tracking.

CN123113990APending Publication Date: 2026-11-17SICHUAN CHUANGSHU SMART TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611524464.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-29
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies cannot effectively analyze the communication messages of drones in low-altitude security, which makes it difficult for photoelectric tracking systems to know the flight intentions in advance, making it difficult to deal with the sudden evasive maneuvers of drones, resulting in target loss and response delays.

Method used

By employing radio reverse protocol parsing technology, density peak clustering and multi-kernel learning support vector machine are used to classify UAV radio frequency signals, demodulate and decrypt communication messages, extract maneuver commands, and combine unscented Kalman filtering to predict target displacement, driving the photoelectric system to actively track.

Benefits of technology

It enables real-time acquisition of drone flight intentions, significantly reduces response latency, avoids target loss, and improves continuous and stable tracking capabilities in low-altitude security scenarios.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123113990A_ABST
    Figure CN123113990A_ABST
Patent Text Reader

Abstract

This invention discloses an active photoelectric detection and tracking method for unmanned aerial vehicles (UAVs) based on radio reverse protocol parsing technology, relating to the fields of low-altitude security and UAV technology. By receiving low-altitude radio frequency signals, frequency hopping bands are extracted using clustering and classification methods. Maneuver commands and flight intentions are obtained through blind demodulation, protocol parsing, and decryption. The UAV's position is estimated by combining multi-node time difference and angle difference, and the maneuver commands are introduced into filtering to predict the target displacement. Based on this, the attitude and focal length of the photoelectric turntable are pre-adjusted, and safety is ensured through mechanical limits and bandwidth constraints. Images are acquired in a preset state, the UAV area is detected and identified, and features are extracted and fed back to the servo system to achieve closed-loop tracking. This invention solves the problems of unresolved encrypted frequency hopping communication and easy loss of passive photoelectric responses, realizing intention-driven active preset tracking and improving the reliability and tracking continuity of UAV detection in complex environments.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of low-altitude security and unmanned aerial vehicle (UAV) technology, specifically a method for active photoelectric detection and tracking of UAVs based on radio reverse protocol parsing technology. Background Technology

[0002] Low-altitude security and drone countermeasures have become crucial safeguards for public safety and low-altitude airspace order. With the rapid proliferation of drone technology, accurate detection and stable tracking are core elements in preventing potential aerial threats. Current mainstream detection methods largely rely on active radar detection or traditional spectrum matching. These methods are highly susceptible to interference from building reflections in complex urban environments and often only provide a rough target location. This coarse location information forces subsequent optoelectronic detection equipment into a passive response state, performing large-scale blind searches within a limited field of view, significantly consuming target acquisition time and making it difficult to respond to sudden low-altitude intrusions.

[0003] The root cause of this passive response lies in the fact that modern drones generally employ frequency-hopping spread spectrum communication and proprietary encryption protocols. Traditional detection methods can only detect the presence of radio frequency signals but cannot deeply analyze the specific content of the communication messages. Because they cannot decipher the drone's equipment model and navigation control commands, the optoelectronic system completely lacks prior information when intervening. For example, when a small drone using frequency-hopping communication suddenly intrudes into the monitored area and performs an abrupt stop and turn to avoid an obstacle, the optoelectronic tracking system can only rely on visual images of the target after it has moved, as its flight intentions and maneuver commands were not known in advance from the radio frequency signals. This delayed tracking, relying solely on changes in visual images, means that the optoelectronic servo system cannot adjust its focus and turntable attitude in time when facing a target moving violently or crossing obstacles, resulting in frequent target loss. Summary of the Invention

[0004] The purpose of this invention is to provide an active photoelectric detection and tracking method for unmanned aerial vehicles (UAVs) based on radio reverse protocol parsing technology, which can realize integrated intelligent detection and stable tracking from radio frequency signal parsing, flight intention perception to photoelectric active pre-set tracking.

[0005] The objective of this invention can be achieved through the following technical solutions: This application provides an active photoelectric detection and tracking method for unmanned aerial vehicles (UAVs) based on radio reverse protocol parsing technology, including the following steps: S1. Obtain radio frequency signals in the low-altitude airspace, and process the radio frequency signals using a frequency hopping signal classification method that combines density peak clustering and multi-kernel learning support vector machine to obtain frequency hopping bands; S2. Demodulate and reverse-parse the protocol for the frequency hopping band, extract the communication message, determine the encryption identifier, if the encryption identifier exists and the system has legally obtained the decryption key, then decrypt to obtain the plaintext message, extract the maneuvering command from the plaintext message, and determine the flight intention of the UAV based on the maneuvering command; S3. Obtain the rough position of the UAV, and process the rough position using a filtering prediction algorithm that incorporates the flight intention to obtain the predicted target displacement; wherein, the filtering prediction algorithm is an unscented Kalman filter, and the maneuver command is used as a known external input to the state transition equation; S4. Calculate the turntable attitude and focal length state of the photoelectric system according to the target displacement, and use the cost function based on mechanical limit pre-detection to determine whether the turntable attitude exceeds the preset threshold. If it does not exceed the preset threshold, drive the photoelectric system according to the turntable attitude and focal length state to obtain the preset state. S5. Acquire a first visual image in the preset state, obtain a second visual image through target detection, extract target features through a twin network to determine the precise position of the UAV, and perform preset tracking.

[0006] The beneficial effects of this invention are as follows: Unsupervised aggregation of energy abrupt change points in the time-frequency energy spectrum is performed using the density peak clustering algorithm to separate candidate frequency hopping signal segments. A multi-kernel learning support vector machine with weighted summation of radial basis kernel, wavelet kernel, and polynomial kernel is constructed to finely classify the candidate segments and accurately extract the frequency hopping bands. Based on this, the original bit stream is recovered by cyclic spectrum correlation blind demodulation. Synchronization word, frame header, and payload fields are adaptively segmented using a protocol field boundary detection model. A hardware-level decryption engine based on a legitimate authorization key is called to decrypt in real time, extracting maneuvering commands such as velocity increment, angular velocity change, and altitude change commands from the plaintext message. The continuous flight intention of the UAV is determined by kinematic mapping. This solves the technical problem that traditional spectrum detection cannot parse encrypted frequency hopping communication messages, realizes the real-time acquisition of UAV flight intention, and provides a priori information basis for subsequent active pre-positioned tracking. By jointly estimating the time difference of arrival and angle difference of arrival of at least three distributed radio frequency receiving nodes, a rough azimuth of the UAV is obtained. The decrypted maneuver commands are quantized into control vectors and embedded in the Sigma point prediction step of unscented Kalman filtering, so that the state prediction explicitly reflects the UAV's active control behavior. When the maneuver commands change abruptly, the process noise covariance matrix is ​​amplified by an adaptive factor to increase the uncertainty weight of the prediction covariance to adapt to the sudden change in motion model. At the same time, the azimuth, pitch angle changes and focal length adjustments of the photoelectric turntable are calculated inversely based on the predicted displacement. A cost function including horizontal mechanical limit, pitch mechanical limit, maximum angular velocity and servo bandwidth is established for feasibility pre-check. The turntable is driven to rotate in advance and preset the focal length only within the threshold range. This solves the problem that traditional photoelectric tracking can only passively follow the target displacement and is difficult to deal with sudden evasive maneuvers. It realizes active preset tracking driven by intention from after-the-fact pursuit, which significantly shortens the response delay and effectively avoids tracking loss. After the optoelectronic system reaches the preset state, it acquires a first visual image, uses an attention mechanism to guide the model to focus on the potential target area, and fuses multi-scale features through a feature pyramid network to accurately segment the area where the UAV is located, outputting a second visual image with bounding boxes. Then, it uses a twin network to extract the deep appearance and motion features of the target, combines camera intrinsic and extrinsic parameters, and accurately maps the pixel coordinates to a three-dimensional spatial position in the world coordinate system. This is fed back to the optoelectronic servo control system through a real-time communication interface, forming a composite control closed loop that combines predictive feedforward and visual feedback. When visual tracking temporarily fails due to occlusion or rapid rotation, the system automatically reverts to the radio frequency prediction guidance mode to reacquire the target. This solves the problems of easy target loss and insufficient robustness of visual tracking in complex backgrounds, achieving high-precision and high-continuity closed-loop tracking. It also constructs a dual-modal complementary mechanism of radio frequency sensing and visual tracking, significantly improving the continuous and stable tracking capability in low-altitude security scenarios. Attached Figure Description

[0007] To better understand and implement this application, the technical solution is described in detail below with reference to the accompanying drawings.

[0008] Figure 1 A flowchart illustrating an active photoelectric detection and tracking method for unmanned aerial vehicles based on radio reverse protocol parsing technology, provided in Embodiment 1 of this application; Figure 2 A flowchart illustrating step S3 in an active photoelectric detection and tracking method for unmanned aerial vehicles based on radio reverse protocol parsing technology, provided in Embodiment 1 of this application; Figure 3 This is a flowchart illustrating step S4 of an active photoelectric detection and tracking method for unmanned aerial vehicles (UAVs) based on radio reverse protocol parsing technology, as provided in Embodiment 1 of this application. Detailed Implementation

[0009] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, exemplary embodiments will be described in detail below, examples of which are illustrated in the accompanying drawings. In the following description, when referring to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.

[0010] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used herein are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0011] The following detailed description of the specific implementation methods, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided in detail.

[0012] Example 1, please refer to Figures 1-3 This embodiment provides an active photoelectric detection and tracking method for unmanned aerial vehicles (UAVs) based on radio reverse protocol parsing technology, including the following steps: S1. Obtain radio frequency signals in the low-altitude airspace, and process the radio frequency signals using a frequency hopping signal classification method that combines density peak clustering and multi-kernel learning support vector machine to obtain frequency hopping bands; Further, step S1 specifically includes: S11. Deploy broadband radio frequency receiving arrays at the edge of key low-altitude airspace to collect radio frequency signals of the communication link between the UAV and the remote controller in real time; In this system, a distributed radio frequency receiving array consisting of three to five broadband software-defined radio receiving nodes is set up at the boundaries of airport airspace, military restricted areas, or security areas for large events. Each receiving node covers the commonly used ISM bands for UAVs, such as 2.4GHz, 5.8GHz, 915MHz, and 433MHz, and collects data in parallel with an instantaneous bandwidth of 20MHz. The sampling rate is set to 40Msps to ensure that the complete frequency hopping signal transition process can be captured. Each node achieves nanosecond-level time synchronization through a GPS timing module and transmits the collected IQ complex baseband data to the central processing unit in real time.

[0013] S12. Perform time-frequency transformation on the acquired broadband radio frequency signal to generate a time-frequency energy spectrum covering the frequency hopping communication band; The central processing unit processes the IQ data uploaded by each receiving node in frames, with a frame length of 1024 sampling points and a frame overlap rate of 87.5%. After adding a Hanning window to each frame of data, a fast Fourier transform is performed to calculate the power spectral density in each time slot. This constructs a two-dimensional time-frequency energy spectrum with time as the horizontal axis, frequency as the vertical axis, and energy values ​​in grayscale or color. The time-frequency resolution reaches the millisecond and kilohertz levels, clearly presenting the striped energy trajectory of the frequency-hopping signal rapidly jumping between different frequency points.

[0014] S13. The density peak clustering algorithm is used to perform unsupervised aggregation of energy abrupt change points in the time-frequency energy spectrum, and candidate frequency hopping signal segments with periodic jumping characteristics are separated from background noise and fixed frequency interference. Furthermore, density peak clustering algorithms specifically include: Each pixel in the time-frequency energy spectrum is considered as a data point, with a time coordinate t, a frequency coordinate f, and an energy amplitude value A(t,f). First, median filtering is applied to the energy spectrum to suppress isolated noise spikes. Then, an energy threshold is set (three times the mean of the background noise), and all energy abrupt changes above the energy threshold are extracted to form a candidate point set. For each point in the candidate point set, calculate its local density. The formula is: ;in, It is the Euclidean distance between point i and point j in the time-frequency plane. The cutoff distance parameter is set to the 2nd percentile of the distances between all point pairs; Calculate the minimum distance from each point to points with a higher density: For the point with the highest density, Take the maximum distance among all points; by x-axis A decision graph is plotted with the vertical axis. Points with both high ρ and high δ are selected as cluster centers. Each non-cluster center point is assigned to the cluster to which its nearest high-density neighbor belongs, eventually forming several energy point clusters. For each cluster, extract its time-frequency range. Calculate the time-domain span Δt and frequency-domain span Δf for all points within the cluster. If the cluster exhibits equally spaced periodicity in the time domain (detected by the autocorrelation function) and multiple discrete frequency points in the frequency domain (number of frequency points ≥ 2), it is identified as a candidate frequency-hopping signal segment; if Δf is very small and continuous in the time domain, it is identified as fixed-frequency interference and eliminated.

[0015] S14. Construct a multi-kernel learning support vector machine classifier. The kernel function is composed of a weighted sum of radial basis kernel, wavelet kernel and polynomial kernel. It performs fine classification on candidate frequency hopping signal segments, distinguishes UAV frequency hopping signals from other non-cooperative signals, and outputs a set of frequency hopping bands.

[0016] Furthermore, for each candidate frequency hopping signal segment output by S13, the central processing unit calculates the following six characteristic parameters: First, the frequency hopping period estimate is obtained by detecting the time interval between energy stripes in the time-frequency energy spectrum and using the peak-to-average method of the autocorrelation function to obtain the time period of the frequency hopping signal jumping from one frequency point to the next. Second, frequency hopping bandwidth, which is the difference between the highest and lowest frequencies covered by the signal segment throughout the entire frequency hopping process, reflects the frequency occupancy range of the UAV communication channel. Third, dwell time, which is the length of time a signal stays on each single frequency point. The dwell time varies for different types of drones or different communication modes. Fourth, the randomness entropy value of the channel sequence is calculated by statistically analyzing the occurrence order of frequency hopping points to distinguish between pseudo-random frequency hopping sequences and fixed sequences. Fifth, the signal-to-noise ratio (SNR), which is the ratio of signal energy to background noise energy, is used to evaluate signal quality. Sixth, the texture features of the time-frequency energy spectrum are extracted using the gray-level co-occurrence matrix method to extract indicators such as the uniformity, contrast, and correlation of the energy distribution; The above six features constitute a six-dimensional feature vector, which is used to describe the essential properties of each candidate frequency hopping signal segment; Constructing a multi-kernel combination kernel function: The classifier uses a weighted combination of three different kernel functions. The first is the radial basis function (RBF) kernel, which is based on the Euclidean distance between two feature vectors and can effectively handle nonlinear classification problems. The second is the wavelet kernel, which is based on the Mexican cap wavelet form and has a good ability to capture local abrupt changes in signals, making it suitable for identifying frequency switching edges in frequency-hopping signals. The third is the second-order polynomial kernel function, which maps to a high-dimensional space through the inner product of feature vectors, and can express the interaction relationships between features. The three kernel functions are each multiplied by a weighting coefficient and then summed to form the final multi-kernel function. All three weighting coefficients are positive and their sum equals one; these coefficients are not fixed but optimized based on the actual electromagnetic environment. During the training phase before system deployment, a large number of known frequency-hopping signal samples (including frequency-hopping signals from common drone brands, WiFi frequency-hopping signals, Bluetooth frequency-hopping signals, and fixed-frequency interference signals) are collected and labeled as positive or negative samples. Then, a five-fold cross-validation method is used, dividing the samples into five parts. Four parts are used for training and one for validation in turn. A grid search is used to traverse all possible combinations of weighting coefficients, and the set of coefficients that achieves the highest classification accuracy is selected as the final configuration. Training the Support Vector Machine (SVM) classifier: The aforementioned multi-kernel function is embedded into the kernel function framework of the SVM. The training objective of the SVM is to find an optimal classification hyperplane that maximizes the margin between positive and negative samples. Specifically, the labeled training sample feature vectors are input into the system to solve a convex quadratic optimization problem. The constraint of this problem is that the product of the classification result of each sample and its true label is not less than one minus a slack variable, while minimizing the sum of the inverse of the classification margin and the penalty of the slack variable. The penalty parameter controls the tolerance for misclassified samples and is set to one hundred based on the balance between positive and negative samples in the actual scenario. By solving this optimization problem, a set of support vectors (i.e., the training samples closest to the classification boundary) and their corresponding Lagrange multipliers are obtained. These support vectors and multipliers together determine the final classification decision function. The system classifies and outputs a set of frequency hopping bands: For each candidate frequency hopping signal segment extracted in real time by S13, the system first calculates its six-dimensional feature vector according to the method in step one. Then, it substitutes this feature vector into the pre-trained multi-core support vector machine classifier to calculate the decision function value. If the decision function value is greater than zero, the segment is determined to be a UAV frequency hopping signal; if it is less than or equal to zero, it is determined to be other non-cooperative signals (such as civilian WiFi frequency hopping, Bluetooth frequency hopping, illegal image transmission interference, etc.) and is removed. For candidate segments identified as UAV frequency hopping signals, extract all frequency values ​​appearing in the segment, merge identical frequency points within adjacent time windows, and remove false frequency points caused by noise; finally, output a set of frequency hopping bands, which contains all frequency values ​​of the frequency hopping channels currently used by the UAV for communication; this set of frequency hopping bands will be used as input to step S2 to guide subsequent blind demodulation processing.

[0017] Specifically, step S1 effectively solves the technical problems of UAV frequency hopping signals being susceptible to fixed-frequency interference and the difficulty in accurately extracting non-cooperative signals such as WiFi / Bluetooth aliasing under complex electromagnetic backgrounds. It achieves highly reliable identification and output of frequency hopping bands, providing accurate frequency guidance for subsequent blind demodulation and protocol reverse analysis.

[0018] S2. Demodulate and reverse-parse the protocol for the frequency hopping band, extract the communication message, determine the encryption identifier, if the encryption identifier exists and the system has legally obtained the decryption key, then decrypt to obtain the plaintext message, extract the maneuvering command from the plaintext message, and determine the flight intention of the UAV based on the maneuvering command; Furthermore, step S2 specifically includes: S21. Perform blind demodulation processing on the signal in each frequency band of S1 output, and recover the original bit stream of the UAV flight control link or remote control link from the frequency hopping sequence. After obtaining the set of frequency hopping bands output by S1, the central processing unit performs the following blind demodulation operation for each frequency hopping band: Based on the dwell time and transition time of the frequency-hopping signal, the entire frequency-hopping sequence is divided into multiple single-frequency dwell segments according to time. For each dwell segment, the modulation parameters of the signal are estimated using the cyclic spectrum correlation method. Specifically, the fourth-order cyclic cumulant of the received signal is calculated, and the modulation type is determined by detecting the peak value at the cyclic frequency. If a single peak is detected and the cyclic spectrum has significant energy at zero frequency, it is determined to be frequency-shift keying (FPS) modulation; if positive and negative symmetrical double peaks are detected, it is determined to be phase-shift keying (PSK) modulation. For signals determined to be FPS, the digital symbols corresponding to each frequency point are determined by statistically analyzing the distribution center of the instantaneous frequency. For signals determined to be PSK, the phase change sequence is recovered through a carrier synchronization loop and mapped to binary bits. After completing the symbol decision for each dwell segment, the bit streams of each segment are concatenated in chronological order according to the frequency hopping sequence to form a complete original bit stream. Simultaneously, frame synchronization is checked using the preamble sequence in the frequency hopping sequence. If the check passes, blind demodulation is confirmed to be successful. If it fails, the modulation parameter estimates are adjusted and demodulation is repeated, with a maximum of three attempts. Finally, the original bit stream corresponding to the frequency hopping band is output, which contains complete information about the UAV flight control downlink data or remote control uplink data.

[0019] S22. Input the original bitstream into the protocol field boundary detection model, and adaptively cut out the synchronization word, frame header and payload fields from the bitstream without prior knowledge of the UAV's private protocol format; The protocol field boundary detection model employs a method based on statistical features and machine learning, specifically including: A sliding window analysis is performed on the bitstream, setting an 8-bit wide window and moving the window one bit at a time, with the starting position of the bitstream as the origin. At each window position, the frequency of the 8-bit value within that window is calculated. The frequency distribution of the 8-bit values ​​across all windows in the entire bitstream is statistically analyzed, forming a probability vector of length 256. Synchronization words typically have a very high repetition probability; therefore, the 8-bit value corresponding to the peak value in the probability vector is the candidate synchronization word. After determining the synchronization word, the system uses the synchronization word as an anchor point and extracts a fixed-length bit sequence as a candidate area for the frame header. The frame header typically contains several fixed-pattern fields, such as the protocol version number, message type, encryption identifier bit field, and payload length field. The boundaries of these fields are automatically identified using a mutual information-based method: for each candidate boundary point, the correlation between the bit values ​​on both sides of the boundary point is calculated, and when the mutual information value changes from high to low, that position is the field boundary. After identifying the frame header structure, based on the payload length field in the frame header, a bit sequence of the corresponding length is truncated from the end of the frame header as the payload field. If the payload length field is not encrypted, it is read directly; if it is encrypted, only the start and end positions of the payload are marked, and it is extracted after decryption. Through the above adaptive segmentation, the system can accurately separate the synchronization word, frame header, payload field, and any possible parity bits from the original bit stream without pre-storing any private protocol templates.

[0020] S23. Identify the encryption flag field in the frame header. If it is determined to be encrypted, call the hardware-level decryption engine to decrypt the encrypted payload in real time to obtain the plaintext message. The decryption engine only performs the decryption operation when the corresponding key is obtained through a legitimate authorization method. The encryption identifier field in the frame header typically occupies one or two bits. The system extracts the value of the encryption identifier field based on the frame header field structure determined in S22; the mapping rule is: binary 00 represents plaintext transmission, 01 represents simple XOR encryption, 10 represents Advanced Encryption Standard (AES), and 11 represents a proprietary encryption algorithm.

[0021] If the encryption identifier is determined to be 00, the payload field is a plaintext message and can be read directly. If it is determined to be 01, 10, or 11, a hardware-level decryption engine is invoked to decrypt the encrypted payload in real time. This hardware-level decryption engine is based on a legally authorized key. It uses a pre-configured key that matches the target UAV's communication protocol (obtained through legal means, such as authorization from the flight control chip manufacturer or provision by relevant law enforcement agencies) to decrypt the encrypted payload in real time and output a plaintext message. The entire decryption process is only implemented when legally authorized or for national or public safety purposes, and does not involve illegal active attacks or cracking activities.

[0022] S24. Extract maneuver commands from the plaintext payload using the command semantic template library. This template library stores the command mapping relationship of various UAV flight control protocols. Convert the command fields into the UAV's velocity vector change, attitude angle setpoint, or heading lock flag. After kinematic mapping, determine the UAV's continuous flight intention.

[0023] After obtaining the plaintext message, the flight control commands within are further parsed. The command semantic template library is a pre-built database that stores the flight control protocol command sets of mainstream drone brands on the market. The construction process of this template library is as follows: By collecting remote control signals from different drone models (such as DJI Matrice series, Autel series, Parnot series, etc.) under known operations, the meaning of the payload fields in each protocol is analyzed through reverse engineering, and a mapping table of command operation codes and physical actions is established. For private protocols that have not yet been included, the system adopts an online learning method. When an unknown command pattern is detected, the pattern is recorded and marked as to be analyzed. The template library is subsequently expanded manually or automatically. During actual unpacking, the first byte or the first two bytes of the plaintext message are read as the instruction opcode, and the corresponding instruction type is retrieved from the template library. Maneuver instructions include: First, the velocity vector change command contains the velocity increment values ​​or target velocity values ​​of the UAV in the three directions of front-back, left-right, and up-down in the body coordinate system. After extracting these values, they are converted into velocity vector changes in the geocentric coordinate system. Second, the attitude angle setting command contains the expected values ​​of the UAV's roll angle, pitch angle and yaw angle. These angle values ​​are recorded as the UAV's target attitude, and the heading change rate and turning radius are further calculated. Third, the heading lock command, which keeps the drone on its current heading, indicating that the drone will enter a straight-line cruise state. Fourth, altitude hold or altitude change command, which includes the target altitude value or rate of ascent and descent, based on which the system determines whether the drone will ascend, descend or fly horizontally; Fifth, the flight mode switching command, which indicates that the drone switches between modes such as hovering, return to home, automatic tracking, and attitude enhancement, and the system predicts the drone's subsequent movement patterns based on this command; After extracting the specific values ​​of the aforementioned command fields, the system executes a kinematic mapping step. This mapping converts the digitized command fields into a continuous description of flight intent with physical meaning. For example, if the command opcode corresponds to "velocity vector change," and the velocity increments in the three directions are 5 meters per second forward, 2 meters per second lateral, and 0 meters per second vertically, the system maps it to "perform a level flight turn maneuver to the northwest at a speed of approximately 5.3 meters per second"; if the command is "roll angle set to 25 degrees, pitch angle 10 degrees, yaw angle unchanged," the system maps it to "a high-G turn with a right roll and nose-down approach"; if the command is "altitude change of -3 meters per second," it maps it to "a dive maneuver with a rate of 3 meters per second to descend."

[0024] The continuous flight intentions of the UAV are output, represented in the form of semantic tags, including but not limited to: constant speed straight cruise, sudden stop hovering, high G-force hovering, rapid climb, dive descent, and serpentine evasive maneuvers. This flight intention will serve as the enhanced input for the unscented Kalman filter in step S3 to predict the future displacement of the UAV and enable the electro-optical system to perform active pre-tracking.

[0025] Specifically, step S2 extracts maneuvering commands from encrypted frequency hopping communication in real time and determines the UAV's flight intention through blind demodulation, protocol reverse parsing, and decryption with a legitimate key. This solves the technical problem that traditional detection cannot parse flight control messages and achieves deep perception of the UAV's active control behavior.

[0026] S3. Obtain the approximate location of the UAV, and process the approximate location using a filtering prediction algorithm that incorporates the flight intention to obtain the predicted target displacement. The filtering prediction algorithm is an unscented Kalman filter, and the maneuver command is used as a known external input to the state transition equation. Furthermore, step S3 specifically includes: S31. Using the joint estimation method of time difference of arrival and angle difference of arrival, the target UAV's spatial azimuth, pitch angle and radial distance relative to the receiving array are calculated by using synchronously collected data from at least three distributed radio frequency receiving nodes on the same frequency hopping band, as a rough azimuth. While S1 and S2 continuously analyze the frequency hopping signal, at least three nodes in the distributed radio frequency receiving array (usually arranged as a triangular or quadrilateral array with side lengths of tens to hundreds of meters) synchronously capture signals within the same frequency hopping band; each node is equipped with a high-precision global positioning system or atomic clock to achieve nanosecond-level time synchronization.

[0027] For each frequency-hopping dwell segment, the system performs the following operations: Calculating the time difference of arrival: Using one node as the reference node, the signals received by the other nodes are subjected to a generalized cross-correlation operation with the signal from the reference node. Specifically, the two signals are multiplied in the frequency domain and then subjected to an inverse Fourier transform to obtain the cross-correlation function. The time delay corresponding to the peak position is the time difference between the arrival of the signals at the two nodes. Since the dwell time of the frequency-hopping signal at each frequency point is very short (usually hundreds of microseconds), the system only selects a few dwell segments with the highest signal-to-noise ratio for time difference estimation to improve accuracy. Calculating the angle of arrival difference: Each receiving node is equipped with a small antenna array (such as a uniform circular array or a cross array), and the direction of arrival of the signal is measured by phase interferometry. Specifically, the phase difference between the signals received by different antenna elements is calculated, and combined with the antenna spacing and signal wavelength, the azimuth and elevation angles of the signal relative to the node array are calculated. The azimuth information measured by each of the three nodes, combined with the known positional relationships between the nodes, can form multiple directional rays, and their intersection points represent the approximate location of the target. The time difference of arrival (TDOA) data and angle difference of arrival (ADO) data are jointly calculated. Specifically, a point in space is taken as the hypothetical target location. The theoretical values ​​of the time difference and angle difference of arrival from that point to each node are calculated, and then compared with the actual measured values ​​to construct an error function. Gradient descent or particle swarm optimization algorithms are used to search for the spatial point that minimizes the error function; this point is the three-dimensional spatial coordinate of the target UAV. From this coordinate, the azimuth angle (with true north as the reference), pitch angle (elevation angle above the horizontal plane), and radial distance (straight-line distance) relative to the center of the receiving array are extracted. Due to the discontinuity of frequency-hopping signals and multipath interference, the accuracy of this positioning result is typically on the order of meters to tens of meters; therefore, it is called a coarse azimuth.

[0028] This rough orientation information will be used as the initial observation for the unscented Kalman filter, and will be used to initialize or correct the state estimation of the filter.

[0029] S32. A fusion prediction algorithm based on unscented Kalman filtering and maneuver intent enhancement is used to quantize the flight intent determined in S2 into a maneuver command vector. The maneuver command vector is extracted from the decrypted UAV communication message and includes velocity increment, angular velocity change, and altitude change commands. As a known external input to the state transition equation, the target displacement is predicted using the following formula:

[0030] in, For time k, the state estimate includes the UAV's position and velocity. Let f(⋅, ⋅) be the covariance matrix, and f(⋅, ⋅) be the nonlinear kinematic state transition function embedded with maneuver commands, used to map the current state of the UAV and external control commands to the next state; and For unscented transformation weights, The process noise covariance matrix describes random disturbances that the system model cannot explain. The prior prediction of the UAV state at time k+1 based on the information at time k, where the position component is used to calculate the preset displacement of the photoelectric turntable; n is the dimension of the state vector x. The nonlinear kinematic state transition function with embedded maneuver commands is specifically expressed as follows:

[0031] in, , and These are position, velocity, and acceleration, respectively. For attitude quaternions, Angular velocity, This is the quaternion multiplication operator. The quaternion representing the attitude change caused by the angular velocity increment over time Δt. The filtering period (prediction time step) is typically 20-50 milliseconds. and From the maneuver command vector The acceleration increment and angular velocity increment decomposed from the equation, when When the set threshold is exceeded, i.e. when the maneuver command vector undergoes a sudden change, the predicted covariance matrix is... The uncertainty weights of the corresponding dimensions are increased, specifically through adaptive adjustment factors. The process noise covariance matrix is ​​scaled to accommodate abrupt changes in the motion model when the UAV performs evasive maneuvers.

[0032] It should be explained that unscented Kalman filtering is a recursive filtering method for state estimation of nonlinear systems. In this embodiment, the maneuvering command extracted from the decrypted communication message in S2 is used as a known external control input and introduced into the prediction step of the filter, so that the state prediction can reflect the active intention of the UAV.

[0033] S33. Extract the position component from the predicted state and subtract it from the position component at the current time to obtain the predicted displacement of the target UAV within a future time window.

[0034] In this process, after S32 completes the state prediction, the predicted state vector of the UAV at the next moment is obtained. The position coordinate part (denoted as the predicted position) is extracted from this vector, and the position coordinate part (denoted as the current position) is extracted from the optimal estimated state at the current moment. Each coordinate component of the predicted position is subtracted from the corresponding coordinate component of the current position to obtain the eastward displacement, northward displacement and skyward displacement. These three values ​​together constitute the predicted target displacement vector. This predicted displacement has a clear physical meaning: it represents the amount of spatial position change that the UAV will undergo within one future filtering cycle under ideal conditions, driven by its current flight intention and without external random interference. Because this prediction incorporates real-time maneuvering commands (such as turning, acceleration / deceleration, and ascent / descent) obtained from communication messages, it has the ability to anticipate sudden evasive maneuvers performed by the UAV, whereas traditional filtering methods based on kinematic extrapolation can only respond passively after the UAV has actually moved.

[0035] The predicted displacement will be transmitted to step S4 to calculate the preset attitude and focal length of the photoelectric turntable, so that the photoelectric system can turn to the position where the target will be reached in advance, instead of following its historical trajectory, thereby achieving true active preset tracking.

[0036] Specifically, a rough orientation is obtained through joint estimation of TDOA / ADOA, and the decrypted maneuver commands are used as the control vectors to embed the prediction steps of the unscented Kalman filter. The nonlinear state transition function of the embedded maneuver commands is used for state prediction, which solves the technical problem that traditional filtering prediction can only be based on historical trajectory extrapolation and cannot predict the UAV's active avoidance maneuvers. It realizes the target displacement prediction of the fused flight intention and provides accurate displacement guidance for the active pre-positioning of the optoelectronic system.

[0037] S4. Calculate the turntable attitude and focal length state of the photoelectric system according to the target displacement, and use the cost function based on mechanical limit pre-detection to determine whether the turntable attitude exceeds the preset threshold. If it does not exceed the preset threshold, drive the photoelectric system according to the turntable attitude and focal length state to obtain the preset state. Furthermore, step S4 specifically includes: S41. Based on the predicted displacement, combined with the current actual azimuth and pitch angles of the photoelectric turntable, the azimuth and pitch angle changes that the turntable needs to rotate are calculated through inverse kinematics. At the same time, the focal length adjustment of the zoom motor is calculated based on the radial distance change in the predicted displacement. This step is executed after the predicted target displacement is obtained in step S3. The predicted displacement is a three-dimensional vector represented by a geographic coordinate system (east, north, and sky), reflecting the distance and direction that the UAV is expected to move in the next filtering cycle.

[0038] Read the current actual azimuth and pitch angles of the photoelectric turntable. These two angles are fed back in real time by an absolute encoder inside the turntable, in degrees. The azimuth angle ranges from 0° to 360° (0° is due north, increasing clockwise), and the pitch angle ranges from -30° to +60° (0° is horizontal, positive upwards). Perform inverse kinematics calculation: Assume that the line of sight of the UAV relative to the photoelectric turntable base at the current moment is known (which can be calculated from the state estimation position and the coordinates of the turntable base in step S3). Superimpose the coordinates of the endpoint of the predicted displacement onto the current target position to obtain the expected target position at the future moment. Then calculate the spatial direction of the expected position relative to the turntable base to obtain the expected azimuth and expected pitch angles. Calculate the azimuth change: Subtract the current actual azimuth from the desired azimuth. If the difference is greater than 180°, subtract 360° (choosing the reverse rotation path); if the difference is less than -180°, add 360°. The absolute value of the resulting azimuth change should not exceed 180°. The pitch change is directly calculated by subtracting the current actual pitch from the desired pitch. The focal length adjustment is calculated based on the change in radial distance: Radial distance refers to the straight-line distance from the photoelectric turntable base to the target UAV. The current radial distance and the expected radial distance (calculated using the expected position coordinates) are calculated separately. If the expected radial distance is less than the current radial distance (the target is closer), the focal length needs to be increased to maintain the appropriate size of the target in the imaging image; otherwise, the focal length is decreased. The focal length adjustment is calculated according to the principle of "keeping the target imaging size unchanged" based on the distance ratio: Expected focal length = current focal length × (current radial distance / expected radial distance). The zoom motor uses a stepper motor, and each pulse corresponds to a fixed focal length change step size. The system converts the focal length adjustment into the number and direction of the stepper motor pulses.

[0039] S42. Establish a cost function that includes the horizontal mechanical limit angle, pitch mechanical limit angle, maximum angular acceleration, and servo bandwidth of the photoelectric turntable. Perform a feasibility pre-check on the calculated turntable attitude to determine whether the required rotation angle touches the mechanical limit and whether the angular velocity exceeds the servo rated upper limit. Furthermore, the cost function is expressed as:

[0040] in, These are the control parameters for azimuth, elevation, and focal length. and The horizontal and pitch mechanical limit thresholds for the photoelectric turntable. Let ω be the angular velocity of the turntable. α, β, and γ are the upper limit of the servo bandwidth and the penalty coefficients. When J(θ) > 0, the turntable posture is determined to exceed the preset threshold, the system prohibits the execution of the drive and maintains the current state; when J(θ) = 0, the posture is determined to be feasible and the preset drive is allowed to be executed.

[0041] The cost function performs a feasibility pre-check on the calculated turntable attitude using three independent constraints: First, it checks whether the desired azimuth angle exceeds the horizontal mechanical limit threshold of the photoelectric turntable (e.g., -170° to +170°). If it does, the first penalty is triggered. Second, it checks whether the desired pitch angle exceeds the pitch mechanical limit threshold (e.g., -30° to +60°). If it does, the second penalty is triggered. Finally, it calculates the required turntable angular velocity by dividing the azimuth and pitch angle changes by the prediction time window. If the absolute value of this angular velocity exceeds the servo bandwidth limit (typically 30° / s to 60° / s), the third penalty is triggered. If any one of these three constraints is triggered, the cost function value is greater than 0, indicating that the turntable attitude exceeds the preset threshold. The system then prohibits the execution of the drive and maintains the current state. Only when none of the three constraints are triggered and the cost function value is 0 is the attitude deemed feasible, allowing the execution of the preset drive. In addition, it is possible to check whether the required angular acceleration exceeds the maximum angular acceleration allowed by the turntable (for example, by roughly estimating by dividing the change in angular velocity by the time window) as an auxiliary safety judgment to ensure that the drive command does not exceed the physical response capability of the motor.

[0042] S43. If all the turntable attitude parameters are within the preset threshold range, the servo driver outputs the drive current to control the photoelectric turntable to rotate to the predicted orientation in advance, and at the same time drives the zoom motor to preset the focal length, so that the photoelectric system is in the preset state before the target UAV enters the field of view.

[0043] Specifically, when the cost function determination result in S42 is feasible (cost function value is zero), the system executes the preset driving operation: The central processing unit sends motion commands to the turntable servo drive via a real-time fieldbus (such as EtherCAT, CANopen, or RS-422). The commands include: target azimuth angle, target pitch angle, maximum rotational speed limit (usually set to 80% of the servo bandwidth limit to allow for margin), and maximum acceleration limit (set to 50% of the rated value to avoid mechanical shock). After receiving the command, the servo driver's built-in position closed-loop controller reads the current actual angle fed back by the encoder, calculates the position deviation, and generates a torque current command through a proportional-integral-derivative (PID) control algorithm. This command drives the permanent magnet synchronous motor or brushless DC motor to rotate. The driver controls the motor to accelerate, maintain a constant speed, and decelerate according to a preset S-shaped speed curve, allowing the turntable to move smoothly to the target position. During the movement, the encoder provides real-time position feedback, and the driver continuously corrects the deviation, ultimately achieving a positioning accuracy within 0.02°. The central processing unit simultaneously sends a focal length preset command to the zoom motor controller, including the target focal length value and the direction of movement. The zoom motor controller calculates the required number of step pulses based on the difference between the current focal length (feedback from a potentiometer or Hall sensor) and the target focal length, and then outputs pulses of the corresponding frequency and number to drive the stepper motor to rotate, thereby moving the lens assembly to the expected position; After the turntable and zoom system reach the target position, each controller returns a "preset complete" status flag to the central processing unit. At this point, the optoelectronic system is in the preset state: the center of the field of view is aligned with the future target position predicted in step S3, and the focal length is adapted to the imaging requirements at that distance. Subsequently, the system waits for the target UAV to fly into the field of view and enters the visual tracking process in step S5; If a new predicted displacement update is received during the preset process (before the turntable reaches the target) (because step S3 runs continuously), the system will immediately interrupt the current motion, re-execute S41~S43, calculate the new target attitude based on the latest predicted value, and issue a new command. This dynamic adjustment mechanism ensures that the photoelectric system always faces the latest predicted orientation; If S42 determines that the turntable attitude exceeds a threshold (e.g., the desired azimuth angle exceeds the mechanical limit), the system will prohibit any drive operations, maintain the current position of the turntable and zoom system, and generate a warning message (containing the reason for exceeding the threshold) to report to the operator or the superior control system. This safety mechanism effectively avoids equipment damage caused by prediction errors.

[0044] Specifically, by solving the turntable attitude and focal length through inverse kinematics and using the cost function of mechanical limit and servo bandwidth constraints for feasibility pre-check, the turntable is driven to be preset only within the safety threshold. This solves the problem that traditional optoelectronic systems lack an active preset mechanism and are prone to mechanical collisions or response delays due to prediction deviations. It achieves a safe and reliable preset state and ensures that the optoelectronic system is in the best observation attitude before the target enters the field of view.

[0045] S5. Acquire a first visual image in the preset state, obtain a second visual image through target detection, extract target features through a twin network to determine the precise position of the UAV, and perform preset tracking.

[0046] Furthermore, step S5 specifically includes: S51. After the optoelectronic system reaches the preset state, the visible light or infrared sensor on the optoelectronic turntable deployed in the airport airspace or the security area of ​​a large event is activated to continuously collect video streams in the low-altitude airspace and obtain first visual images containing intruding drone targets. In this process, once step S4 completes the preset drive, the photoelectric turntable has rotated to the predicted orientation, and the zoom motor has been adjusted to the expected focal length, the system confirms that the preset status flag is valid, and the central processing unit immediately starts the photoelectric sensor to acquire the video stream. The photoelectric turntable typically integrates dual light sensors: a visible light camera (for color image acquisition during the day or under sufficient light conditions) and an infrared thermal imager (for thermal target detection at night, in low light, or in smoky environments). Based on the preset time period, light sensor data, and scene requirements, one of the channels or both channels can be automatically selected or activated simultaneously. The video stream is fed into the central processing unit in the form of a continuous frame sequence of 30 to 60 frames per second. The resolution of each frame is typically 1920×1080 pixels or higher. Since step S4 has already aligned the field of view center with the future target position predicted in step S3, the drone target is likely to appear in the area near the center of the first visual image without the need for a global search of the entire image. This "pre-guided" approach significantly narrows the potential range of the target, reducing the computational burden on subsequent target detection.

[0047] S52. Using a target detection model based on attention mechanism and feature pyramid network, end-to-end detection of drone targets in the first visual image is performed. The drone's location is accurately segmented from complex ground and aerial backgrounds such as buildings, trees, and birds, and a second visual image with target bounding boxes is output. For each frame of the first visual image captured, the system runs a pre-trained deep learning object detection model. This model is based on a convolutional neural network architecture, and its core components include an attention mechanism and a feature pyramid network. The role of the attention mechanism is to focus the model on the region in the image most likely to contain the drone target. Specifically, the model calculates the self-attention weight for each spatial location in the shallow feature map, highlighting feature responses related to the drone's shape and motion pattern, while suppressing background noise (such as swaying leaves or moving clouds). The Feature Pyramid Network is used to address the issue of large scale variations of drone targets in images. When the drone is far from the optoelectronic system, it occupies only tens of pixels in the image; when it is close, it may occupy hundreds of pixels. The Feature Pyramid Network, through top-down paths and lateral connections, fuses deep, strong semantic features with shallow, high-resolution location information to generate multi-scale feature representations, enabling the model to detect both large and small targets simultaneously. This target detection model has been trained on a large number of low-altitude security scene images. The training samples include various drone images under different lighting, weather, and background conditions, as well as common interference objects such as birds, balloons, and plastic bags as negative samples. The model outputs one or more bounding boxes (each bounding box consists of center coordinates, width, height, and confidence score). Only bounding boxes with a confidence score exceeding a set threshold (e.g., 0.7) are considered valid drone targets. For dual-light systems, the model can process visible light and infrared images separately and fuse the results: if both channels detect a bounding box at the same location, the confidence score is further increased. After target detection, the system crops or marks the area where the drone is located from the original first-view image, generating a second-view image with a clear target bounding box. This second-view image will serve as input for subsequent feature extraction.

[0048] S53. Extract the appearance and motion features of the UAV from the second visual image using deep learning tracking methods. Combine the world coordinate extrinsic parameters of the camera's internal parameters and installation position pre-calibrated by the optoelectronic system to calculate the precise pixel coordinates of the UAV in the image coordinate system. Then, through coordinate transformation, map the UAV to the three-dimensional spatial position in the world coordinate system with the optoelectronic turntable as the origin. After obtaining the target bounding box in the second visual image, the system uses a Siamese network as the core deep learning tracking method. The Siamese network consists of two sub-networks with the same structure and shared weights: one for extracting features from the template image (i.e., the appearance of the target in the initial frame) and the other for extracting features from the search image (the candidate region of the current frame). By calculating the similarity between the two feature vectors, the system can locate the precise position of the target in the current frame.

[0049] The specific implementation steps are as follows: Initialize template: When the UAV target is detected for the first time in step S52, the system crops out the image region within the bounding box, normalizes it to a fixed size (e.g., 127×127 pixels), and extracts a high-dimensional feature vector through the template branch of the Siamese network. This vector represents the appearance features of the UAV (such as rotor configuration, body color, texture, etc.) and is denoted as the target template. For each subsequent frame, the system uses the target position of the previous frame as the center and crops a search region in the current frame that is slightly larger than the bounding box (e.g., 255×255 pixels). After normalization, it is fed into the search branch of the Siamese network to obtain the feature map of the search region. The cross-correlation between template features and search region features (i.e., convolution operation) is calculated to generate a response score map. The position with the highest score is the estimated position of the target in the current frame. To adapt to changes in UAV attitude and lighting, the system updates the template with the detection results of the current frame every certain period of time (e.g., every 10 frames) or when the response score is lower than the threshold. After obtaining the precise pixel coordinates (center point coordinates u, v) of the target in the image coordinate system, the system combines the pre-calibrated intrinsic parameter matrix of the photoelectric camera (including focal length, principal point coordinates, distortion coefficients, etc.) and the extrinsic parameter matrix of the installation position (i.e., the world coordinates of the photoelectric turntable base) to transform the pixel coordinates into a three-dimensional spatial position with the turntable center as the origin through coordinate transformation. The coordinate transformation steps include: first, performing distortion correction on the pixel coordinates; then, back-projecting the unit direction vector from the pinhole imaging model to the camera coordinate system; then, rotating the direction vector to the turntable coordinate system using extrinsic parameters; and finally, calculating the precise three-dimensional coordinates of the target by combining the radial distance of the target (which can be obtained from the predicted distance in step S3 or based on the estimation of the target imaging size).

[0050] S54. The calculated precise spatial position is fed back to the photoelectric servo control system through the real-time communication interface, driving the turntable to continuously align with the UAV, so that the UAV is always in the center of the photoelectric field, achieving continuous and stable preset tracking in low-altitude airspace.

[0051] After obtaining the precise three-dimensional spatial position of the UAV target in each frame, the system sends the position data to the servo control system of the photoelectric turntable via a real-time communication interface (such as EtherCAT, CANopen, or Gigabit Ethernet). The servo control system internally runs a position closed-loop control algorithm: it converts the received target spatial position into the desired turntable azimuth and pitch angles (inverse kinematics, similar to step S41), compares them with the current actual angles, calculates the deviation, and then outputs a drive current through a proportional-integral-derivative controller to continuously rotate the turntable and keep the target at the center of the photoelectric field.

[0052] To adapt to the continuous movement of the UAV, the system adopts a predictive-corrective composite control strategy: on the one hand, the predicted displacement in step S3 is used as a feedforward quantity to adjust the turntable attitude in advance; on the other hand, the visual tracking result in step S53 is used as a feedback quantity for fine correction. The fusion of the two achieves rapid response and high-precision tracking.

[0053] In addition, the system is equipped with a tracking re-detection mechanism: if the drone target is not detected for several consecutive frames (e.g., 5 frames) (i.e., the output of step S52 is empty or the confidence level is too low), the system determines that tracking is lost. At this time, it immediately reverts to the radio frequency prediction guidance mode in step S3 and searches for the possible location of the target again. Once the target is detected again, visual tracking is resumed.

[0054] Through steps S51 to S54, the optoelectronic system, guided by radio frequency analysis and intent prediction, achieves active pre-positioned tracking of UAV targets in low-altitude airspace, characterized by fast response, high accuracy, and strong anti-interference capability.

[0055] Specifically, step S5 detects the UAV region through an attention mechanism and a feature pyramid network, extracts features through a twin network, calculates the 3D position, and feeds it back to the servo system. This solves the problems of easy loss and insufficient robustness of traditional visual tracking in complex environments, and achieves high-precision and high-continuity closed-loop tracking. At the same time, a composite control mechanism combining predictive feedforward and visual feedback is constructed.

[0056] The decryption operation described in this method is only performed when legally authorized or for national or public safety purposes, and does not involve illegal active attacks or cracking activities. Any unauthorized interception or decryption of other people's communication signals is not within the scope of protection of this patent.

[0057] Example 2 provides an active photoelectric detection and tracking method for unmanned aerial vehicles (UAVs) based on radio reverse protocol parsing technology. The difference between this example and Example 1 lies in the visual tracking method. Example 1 employs a deep learning-based target feature extraction and tracking method using Siamese networks, while this example, targeting low-altitude security scenarios with limited computing resources or extremely high real-time requirements, such as mobile vehicle-mounted anti-UAV platforms and border patrol points, uses a fast visual tracking method based on discriminative correlation filters. The remaining steps (including radio frequency signal processing, protocol reverse parsing, maneuver command extraction, unscented Kalman filter prediction, and photoelectric preset driving) are the same as in Example 1 and will not be repeated here. Specific differences are described below: The specific implementation process of the visual tracking method in this embodiment (replacing step S53 of embodiment 1): After outputting the second visual image with the target bounding box in step S52, this embodiment no longer runs the Siamese network, but instead uses a Discriminative Correlation Filter (DCF) for high-speed target tracking. The specific implementation of this method includes the following sub-steps: Target appearance model initialization: When the UAV target is detected for the first time (i.e., the first valid bounding box is output in step S52), the system extracts various manually designed features of the target area with the center of the bounding box as the origin. This embodiment addresses the complex backgrounds commonly encountered by UAVs in low-altitude security scenarios (such as airport runways, city skylines, mountains, and trees), fusing three complementary features: Oriented gradient histogram features, used to capture the edge contours and rotor structure of the UAV; color naming features, used to distinguish UAVs from birds (UAVs typically have single-tone colors such as gray, white, and black, while bird feathers have varied colors); and local binary pattern texture features, used to enhance robustness to changes in lighting. These features are superimposed in a multi-channel format to form the target template. Training of the correlation filter: Based on the initialized target template, the system trains a minimum output squared error filter in the frequency domain. Specifically, with the target center as the origin, a search region slightly larger than the target boundary is extracted, generating a Gaussian-shaped expected response (the response value is highest at the center and gradually decreases towards the edge). The feature map is transformed to the frequency domain using a fast Fourier transform, and the closed-form least squares solution is solved to obtain the frequency domain coefficients of the correlation filter. This training process takes only tens of microseconds per frame, much faster than the forward inference time of the Siamese network. Rapid target location localization: For each subsequent frame, the system first extracts a search region of the same size centered on the target location of the previous frame and calculates a multi-channel feature map. After converting to the frequency domain, it multiplies the feature map element-wise with pre-trained filter coefficients, and then obtains the spatial response map through inverse Fourier transform. The maximum value of the response map is the precise pixel coordinate of the UAV target in the current frame. Since the entire localization process is performed entirely in the frequency domain (accelerated using fast Fourier transform), the computational complexity is far lower than that of deep learning models, enabling a processing speed of hundreds of frames per second on ordinary embedded devices. Adaptive Model Update: Considering the potential scale changes (from far to near or near to far), lighting changes (passing through cloud shadows), and partial occlusion (flying over trees or behind buildings) that drones may encounter in low-altitude security scenarios, this embodiment employs an adaptive update strategy. The system calculates the peak-to-side-lobe ratio of the response map for each frame. If this ratio is higher than a set threshold (indicating high tracking confidence), the filter coefficients and target template are updated using the detection results of the current frame (including target position and appearance features) at a fixed learning rate (usually 0.02). If the peak-to-side-lobe ratio is lower than the threshold (indicating possible occlusion or rapid movement), the update is paused to avoid filter contamination. Simultaneously, to address scale changes, this embodiment additionally maintains a one-dimensional scale-related filter. Each frame evaluates the responses of multiple scale candidate windows, selects the highest response as the optimal scale, and adjusts the target bounding box size accordingly. Fusion and verification with RF prediction results: A key improvement in this embodiment is the real-time fusion of the visual tracking results of the correlation filter with the RF predicted displacement from step S3. Specifically, the pixel positions obtained from visual tracking are transformed to obtain visual positioning space coordinates, and RF prediction yields predicted space coordinates. These two coordinates are combined in the central processing unit using an adaptive weighted fusion algorithm. Positioning results with high confidence are given greater weight, with confidence determined by the peak-to-sidelobe ratio of the response and the signal-to-noise ratio of the RF signal. When visual tracking is lost due to rapid target rotation or severe occlusion (the peak-to-sidelobe ratio of the response remains below a threshold), the system relies entirely on the RF prediction results to continue driving the photoelectric turntable. Simultaneously, the search area is expanded and the correlation filter is reinitialized. Visual tracking resumes once the target is detected again. This RF-enhanced correlation filter tracking mechanism fully leverages the advantages of radio frequency analysis, overcoming the weakness of pure visual methods in recovering lost targets.

[0058] The advantages of this embodiment compared to Embodiment 1 are: the computational cost of the correlation filter is only one-hundredth to one-thousandth that of the twin network, and it can run in real time on low-cost embedded platforms (such as ARM Cortex-A series processors), significantly reducing system hardware costs; at the same time, frequency domain operations have good predictability and stability, making them suitable for real-time tracking in vibration environments such as vehicle-mounted and airborne systems. This embodiment is particularly suitable for UAV countermeasure scenarios that are sensitive to equipment power consumption and deployment costs, such as border patrols and temporary field deployments; although its robustness in extremely complex backgrounds is slightly lower than that of deep learning models, by fusing radio frequency prediction information, its overall tracking success rate can still reach a level similar to that of Embodiment 1.

[0059] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any brief modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for active photoelectric detection and tracking of unmanned aerial vehicles (UAVs) based on radio reverse protocol parsing technology, characterized in that: Includes the following steps: S1. Obtain radio frequency signals in the low-altitude airspace, and process the radio frequency signals using a frequency hopping signal classification method that combines density peak clustering and multi-kernel learning support vector machine to obtain frequency hopping bands; S2. Demodulate and reverse-parse the protocol for the frequency hopping band, extract the communication message, determine the encryption identifier, if the encryption identifier exists and the system has legally obtained the decryption key, then decrypt to obtain the plaintext message, extract the maneuvering command from the plaintext message, and determine the flight intention of the UAV based on the maneuvering command; S3. Obtain the rough position of the UAV, and process the rough position using a filtering prediction algorithm that incorporates the flight intention to obtain the predicted target displacement; wherein, the filtering prediction algorithm is an unscented Kalman filter, and the maneuver command is used as a known external input to the state transition equation; S4. Calculate the turntable attitude and focal length state of the photoelectric system according to the target displacement, and use the cost function based on mechanical limit pre-detection to determine whether the turntable attitude exceeds the preset threshold. If it does not exceed the preset threshold, drive the photoelectric system according to the turntable attitude and focal length state to obtain the preset state. S5. Acquire a first visual image in the preset state, obtain a second visual image through target detection, extract target features through a twin network to determine the precise position of the UAV, and perform preset tracking.

2. The active photoelectric detection and tracking method for unmanned aerial vehicles based on radio reverse protocol parsing technology according to claim 1, characterized in that: Step S1 specifically includes: S11. Deploy broadband radio frequency receiving arrays at the edge of key low-altitude airspace to collect radio frequency signals of the communication link between the UAV and the remote controller in real time; S12. Perform time-frequency transformation on the acquired broadband radio frequency signal to generate a time-frequency energy spectrum covering the frequency hopping communication band; S13. The density peak clustering algorithm is used to perform unsupervised aggregation of energy abrupt change points in the time-frequency energy spectrum, and candidate frequency hopping signal segments with periodic jumping characteristics are separated from background noise and fixed frequency interference. S14. Construct a multi-kernel learning support vector machine classifier. The kernel function is composed of a weighted sum of radial basis kernel, wavelet kernel and polynomial kernel. It performs fine classification on candidate frequency hopping signal segments, distinguishes UAV frequency hopping signals from other non-cooperative signals, and outputs a set of frequency hopping bands.

3. The active photoelectric detection and tracking method for unmanned aerial vehicles based on radio reverse protocol parsing technology according to claim 2, characterized in that: The weighting coefficients of the radial basis kernel, wavelet kernel, and polynomial kernel in the multi-kernel learning support vector machine are dynamically determined using a Bayesian optimization method based on cross-validation, so as to adaptively match the frequency band distribution characteristics of different frequency hopping modes. The density peak clustering algorithm automatically identifies cluster centers and separates candidate frequency hopping signal segments by calculating the local density and relative distance of each point in the time-frequency energy spectrum.

4. The active photoelectric detection and tracking method for unmanned aerial vehicles based on radio reverse protocol parsing technology according to claim 1, characterized in that: Step S2 specifically includes: S21. Perform blind demodulation processing on the signal in each frequency band of S1 output, and recover the original bit stream of the UAV flight control link or remote control link from the frequency hopping sequence. S22. Input the original bitstream into the protocol field boundary detection model, and adaptively cut out the synchronization word, frame header and payload fields from the bitstream without prior knowledge of the UAV's private protocol format; S23. Identify the encryption flag field in the frame header. If it is determined to be encrypted, call the hardware-level decryption engine to decrypt the encrypted payload in real time to obtain the plaintext message. The decryption engine only performs the decryption operation when the corresponding key is obtained through a legitimate authorization method. S24. Extract maneuver commands from the plaintext payload using the command semantic template library. This template library stores the command mapping relationship of various UAV flight control protocols. Convert the command fields into the UAV's velocity vector change, attitude angle setpoint, or heading lock flag. After kinematic mapping, determine the UAV's continuous flight intention.

5. The active photoelectric detection and tracking method for unmanned aerial vehicles based on radio reverse protocol parsing technology according to claim 4, characterized in that: The hardware-level decryption engine is a decryption engine based on a legitimate authorized key, which decrypts the encrypted payload in real time using a pre-configured key that matches the target UAV's communication protocol. The protocol field boundary detection model is based on sliding window statistics and mutual information analysis to perform adaptive segmentation of protocol fields without prior knowledge.

6. The active photoelectric detection and tracking method for unmanned aerial vehicles based on radio reverse protocol parsing technology according to claim 1, characterized in that: Step S3 specifically includes: S31. Using the joint estimation method of time difference of arrival and angle difference of arrival, the target UAV's spatial azimuth, pitch angle and radial distance relative to the receiving array are calculated by using synchronously collected data from at least three distributed radio frequency receiving nodes on the same frequency hopping band, as a rough azimuth. S32. A fusion prediction algorithm based on unscented Kalman filtering and maneuver intent enhancement is adopted to quantize the flight intent determined in S2 into a maneuver command vector, which is used as the known external input to the state transition equation to calculate the predicted target displacement: ; in, For time k, the state estimate includes the UAV's position and velocity. Let f(⋅, ⋅) be the covariance matrix, and f(⋅, ⋅) be the nonlinear kinematic state transition function embedded with maneuver commands, used to map the current state of the UAV and external control commands to the next state; and For unscented transformation weights, The process noise covariance matrix describes random disturbances that the system model cannot explain. The prior prediction of the UAV state at time k+1 based on the information at time k, where the position component is used to calculate the preset displacement of the photoelectric turntable; n is the dimension of the state vector x. S33. Extract the position component from the predicted state and subtract it from the position component at the current time to obtain the predicted displacement of the target UAV within the future time window.

7. The active photoelectric detection and tracking method for unmanned aerial vehicles based on radio reverse protocol parsing technology according to claim 3, characterized in that: The maneuver command vector includes at least velocity increment, angular velocity change, and altitude change command. When the maneuver command vector undergoes a sudden change, the uncertainty weight of the corresponding dimension in the prediction covariance matrix of the unscented Kalman filter increases. The process noise covariance matrix is ​​amplified by an adaptive factor to adapt to the sudden change in the motion model when the UAV performs evasive maneuvers.

8. The active photoelectric detection and tracking method for unmanned aerial vehicles based on radio reverse protocol parsing technology according to claim 1, characterized in that: Step S4 specifically includes: S41. Based on the predicted displacement, combined with the current actual azimuth and pitch angles of the photoelectric turntable, the azimuth and pitch angle changes that the turntable needs to rotate are calculated through inverse kinematics. At the same time, the focal length adjustment of the zoom motor is calculated based on the radial distance change in the predicted displacement. S42. Establish a cost function that includes the horizontal mechanical limit angle, pitch mechanical limit angle, maximum angular acceleration, and servo bandwidth of the photoelectric turntable. Perform a feasibility pre-check on the calculated turntable attitude to determine whether the required rotation angle touches the mechanical limit and whether the angular velocity exceeds the servo rated upper limit. S43. If all the turntable attitude parameters are within the preset threshold range, the servo driver outputs the drive current to control the photoelectric turntable to rotate to the predicted orientation in advance, and at the same time drives the zoom motor to preset the focal length, so that the photoelectric system is in the preset state before the target UAV enters the field of view.

9. The active photoelectric detection and tracking method for unmanned aerial vehicles based on radio reverse protocol parsing technology according to claim 8, characterized in that: The cost function is expressed as: ; in, These are the control parameters for azimuth, elevation, and focal length. and The horizontal and pitch mechanical limit thresholds for the photoelectric turntable. Let ω be the angular velocity of the turntable. α, β, and γ are the upper limit of the servo bandwidth and the penalty coefficients. When J(θ) > 0, the turntable posture is determined to exceed the preset threshold, the system prohibits the execution of the drive and maintains the current state; when J(θ) = 0, the posture is determined to be feasible and the preset drive is allowed to be executed.

10. The active photoelectric detection and tracking method for unmanned aerial vehicles based on radio reverse protocol parsing technology according to claim 8, characterized in that: Step S5 specifically includes: S51. After the optoelectronic system reaches the preset state, the visible light or infrared sensor on the optoelectronic turntable deployed in the airport airspace or the security area of ​​a large event is activated to continuously collect video streams in the low-altitude airspace and obtain first visual images containing intruding drone targets. S52. An object detection model based on attention mechanism and feature pyramid network is adopted to perform end-to-end detection of drone targets in the first visual image, accurately segment the area where the drone is located from the complex background, and output a second visual image with target bounding boxes. S53. Extract the appearance and motion features of the UAV from the second visual image using a twin network. Combine the world coordinate extrinsic parameters of the camera's internal components and installation position pre-calibrated by the optoelectronic system to calculate the precise pixel coordinates of the UAV in the image coordinate system. Then, through coordinate transformation, map the UAV to a three-dimensional spatial position in the world coordinate system with the optoelectronic turntable as the origin. S54. The calculated precise spatial position is fed back to the photoelectric servo control system through the real-time communication interface, driving the turntable to continuously align with the UAV, so that the UAV is always in the center of the photoelectric field, achieving continuous and stable preset tracking in low-altitude airspace.