Vehicle tracker detection positioning method and system based on wireless signal analysis

Abstract

The application provides a vehicle tracker detection positioning method and system based on wireless signal analysis, relates to the technical field of vehicle safety detection, and comprises the following steps: wireless signal detectors are arranged around a vehicle, signal characteristics are extracted by adopting segmented fast Fourier transform and wavelet packet decomposition, whether the vehicle is installed with a tracker is judged in combination with a long short-term memory network, and the type and accurate position of the tracker are determined through frequency domain decomposition, energy density feature map analysis and a weighted triangulation method, so that the vehicle tracker can be quickly and accurately identified, the detection positioning precision is improved, and the vehicle safety guarantee capability is enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of vehicle safety detection, and in particular to a vehicle tracker detection and positioning method and system based on wireless signal analysis. Background Art

[0002] With the rapid development of information technology, vehicle safety issues are increasingly attracting public attention. Tracking devices are usually small in size, highly concealed, can work for a long time, and transmit information such as vehicle location via wireless signals. Traditional vehicle safety inspections mainly rely on manual visual inspections or the use of simple radio frequency detectors, which are difficult to cope with the increasingly complex tracking devices.

[0003] Vehicle tracker detection technology based on wireless signal analysis has advanced in recent years. However, existing technologies still have limited ability to identify weak signals in complex electromagnetic environments. When a vehicle is exposed to significant signal interference, it is prone to missed detections or false alarms. The technology also lacks effective signal feature extraction mechanisms, making it difficult to accurately distinguish signals from legitimate communication devices and malicious trackers. This results in low recognition accuracy and poor positioning accuracy, often only providing approximate direction information but unable to precisely locate the specific installation location of the tracker on the vehicle.

[0004] Therefore, a solution is urgently needed to solve the problems existing in the prior art. Summary of the Invention

[0005] The embodiments of the present invention provide a vehicle tracker detection and positioning method and system based on wireless signal analysis, which can at least solve some of the problems existing in the prior art.

[0006] A first aspect of an embodiment of the present invention provides a vehicle tracker detection and positioning method based on wireless signal analysis, comprising:

[0007] Deploy multiple wireless signal detectors around the vehicle to be detected, divide the signal collection area into multiple frequency band sub-areas, and set sampling parameters for each frequency band sub-area;

[0008] For each frequency band sub-region, the minimum sampling rate is calculated based on Shannon's sampling theorem, and a piecewise fast Fourier transform is performed to obtain the spectrum data. The filtered spectrum data is obtained by recursively updating the filter coefficients. An eight-layer wavelet packet decomposition tree is constructed using orthogonal basis functions to obtain decomposition nodes. The energy entropy of the decomposition nodes is calculated to determine the time-frequency feature matrix, and the signal feature vector is generated through a long short-term memory network.

[0009] Based on the signal feature vector, it is determined whether the vehicle to be detected is equipped with a vehicle tracker. If the judgment result is yes, the signal feature vector is decomposed in the frequency domain to obtain multiple groups of frequency band features. The energy density distribution of each group of frequency band features is calculated to obtain a density feature map. The salient regions are extracted from the density feature map to obtain key signal features. The key signal features are adaptively segmented by threshold to obtain the vehicle tracker type and initial position information.

[0010] According to the vehicle tracker type and initial location information, combined with the signal strength difference of the wireless signal detector, the weighted triangulation method is used to calculate the installation location coordinates of the vehicle tracker and send them to the display terminal.

[0011] In an optional embodiment,

[0012] Deploy multiple wireless signal detectors around the vehicle to be detected, divide the signal collection area into multiple frequency band sub-areas, and set sampling parameters for each frequency band sub-area, including:

[0013] Deploy multiple wireless signal detectors around the vehicle to be detected;

[0014] The signal collection area is divided into three frequency band sub-areas: high frequency, medium frequency and low frequency according to the signal frequency characteristics. The frequency range of each frequency band sub-area is determined according to the working frequency band of the vehicle tracker;

[0015] Based on the frequency range and signal attenuation characteristics of each frequency band sub-region, the sampling frequency and sampling interval are adaptively calculated to generate corresponding sampling parameters, where the sampling frequency is not less than twice the highest frequency of the signal, and the sampling interval is determined based on the signal energy density distribution.

[0016] In an optional embodiment,

[0017] For each frequency band sub-region, the minimum sampling rate is calculated based on Shannon sampling theorem, and the spectrum data is obtained by performing piecewise fast Fourier transform. The filtered spectrum data is obtained by recursive iterative updating of the filter coefficients. An eight-layer wavelet packet decomposition tree is constructed using orthogonal basis functions to obtain decomposition nodes. The energy entropy value of the decomposition node is calculated to determine the time-frequency feature matrix and the signal feature vector is generated through the long short-term memory network.

[0018] For each frequency band sub-region, based on Shannon sampling theorem, the theoretical minimum sampling rate is calculated by twice the signal bandwidth. The oversampling factor is calculated according to the signal energy distribution. The theoretical minimum sampling rate is multiplied by the oversampling factor to determine the minimum sampling rate. The sampled signal is obtained based on the minimum sampling rate.

[0019] The sampled signal is segmented and processed, the window length is determined based on the signal time correlation, and the segmented signal is windowed with the cosine window function, and the segmented fast Fourier transform is performed to obtain the spectrum data;

[0020] The filter coefficients are updated through recursive iteration, the product of the input signal vector and the filter coefficient vector is calculated to obtain the predicted value, the predicted value and the expected output are calculated to obtain the error signal, and the filter coefficients are updated based on the error signal to obtain the filtered spectrum data;

[0021] Based on the filtered spectrum data, an eight-layer wavelet packet decomposition tree is constructed using orthogonal basis functions, and decomposition nodes are obtained through low-pass decomposition and high-pass decomposition;

[0022] Calculate the energy entropy value of the decomposed node, divide the square of the node coefficient by the sum of the squares of the node coefficients of the current layer to obtain the normalized probability, sum the normalized probability and the logarithmic product to obtain the energy entropy value, organize the energy entropy value into a time-frequency feature matrix according to the time-frequency distribution, and input the time-frequency feature matrix into the long short-term memory network to generate the signal feature vector.

[0023] In an optional embodiment,

[0024] Inputting the time-frequency feature matrix into the long short-term memory network to generate the signal feature vector includes:

[0025] Inputting the time-frequency feature matrix into the forward recursive branch and the backward recursive branch of the long short-term memory network respectively, and calculating the forward feature and the backward feature based on the feature state of the previous moment in the forward recursive branch and the feature state of the next moment in the backward recursive branch respectively;

[0026] The feature data of the current time step in the time-frequency feature matrix, the feature state of the previous moment in the forward recursive branch, and the feature state of the next moment in the backward recursive branch are concatenated to generate a concatenated result and the dynamic weight coefficient is obtained by mapping the result with a sigmoid function.

[0027] The feature data of the current time step in the time-frequency feature matrix is ​​concatenated with the forward feature of the previous time step and transformed to obtain the forward state; the feature data of the current time step in the time-frequency feature matrix is ​​concatenated with the backward feature of the next time step and transformed to obtain the backward state;

[0028] Multiply the forward state by the dynamic weight coefficient, multiply the backward state by the complement of the dynamic weight coefficient, add the two product results to obtain a fusion state, concatenate the fusion state with the feature data of the current time step in the time-frequency feature matrix, obtain the output gate parameters after transformation, and multiply the output gate parameters by the hyperbolic tangent function mapping result of the fusion state to obtain the signal feature vector.

[0029] In an optional embodiment,

[0030] According to the signal feature vector, it is judged whether the vehicle to be detected is installed with a vehicle tracker, when the judgment result is yes, the signal feature vector is subjected to frequency domain decomposition to obtain a plurality of groups of frequency band features, the energy density distribution of each group of frequency band features is calculated to obtain a density feature map, a salient region is extracted according to the density feature map to obtain key signal features, and the key signal features are subjected to adaptive threshold segmentation to obtain vehicle tracker type and initial position information including:

[0031] The signal feature vector is subjected to nonlinear transformation through a double-layer neural network, the first layer weight matrix of the double-layer neural network is multiplied by the signal feature vector, and after a first layer bias vector is added, the mapping result is mapped through a ReLU function, the mapping result is multiplied by a second layer weight matrix, and after a second layer bias vector is added, a discrimination result probability distribution is obtained through a softmax function;

[0032] The signal feature vector is subjected to frequency domain decomposition through a multi-bandwidth Hilbert-Huang transform, intrinsic mode components are obtained through an adaptive band-pass filter bank, the intrinsic mode components are divided into a plurality of frequency bands, the energy density distribution of the plurality of frequency bands is calculated to generate a density feature map;

[0033] A sliding window is constructed with each feature point in the density feature map as a center, the mean and standard deviation of the feature points in the sliding window are calculated, the difference between the feature points and the mean is divided by the standard deviation to obtain a local contrast, a saliency map is constructed by multiplying the local contrast and a Gaussian weighted distance of the feature points to the center feature, and a salient region is extracted as key signal features;

[0034] The key signal features are subjected to Otsu threshold segmentation, the segmented features are matched with a pre-set tracker feature template library to obtain a vehicle tracker type, and the energy centroid of the key signal features is calculated to obtain vehicle tracker initial position information.

[0035] In an optional implementation,

[0036] A saliency map is constructed by multiplying the local contrast and a Gaussian weighted distance of the feature points to the center feature, and a salient region is extracted as key signal features, including:

[0037] A local feature manifold structure of the density feature map is constructed, a neighborhood set of each feature point in the density feature map is constructed into an adjacency matrix through a Gaussian kernel function, a geodesic distance on the manifold is calculated based on the adjacency matrix through a shortest path algorithm, and a neighborhood feature of each feature point is exponentially weighted and summed according to the geodesic distance to obtain a reconstructed feature;

[0038] Adaptively fusing the reconstructed feature with the local contrast to obtain enhanced local contrast, wherein the adaptive fusion is achieved by dynamically adjusting a fusion coefficient, the fusion coefficient is determined according to the local feature variance, and the Gaussian weighted distance from each feature point in the density feature map to the central feature is calculated by a negative exponential function of the Euclidean distance between the feature point value and the central feature;

[0039] A saliency map is constructed by multiplying the enhanced local contrast with the Gaussian weighted distance, and a salient region is extracted from the saliency map as a key signal feature by a non-maximum suppression algorithm.

[0040] In an optional embodiment,

[0041] Based on the vehicle tracker type and initial location information, combined with the signal strength difference of the wireless signal detector, the weighted triangulation method is used to calculate the installation location coordinates of the vehicle tracker and send them to the display terminal, including:

[0042] Obtaining the type and initial location information of the vehicle tracker, and collecting the wireless signal strength emitted by the vehicle tracker within the detection range of each wireless signal detector;

[0043] calculating a signal strength difference between adjacent wireless signal detectors, and determining a weight coefficient of the vehicle tracker relative to the wireless signal detector based on the signal strength difference;

[0044] Based on the known coordinates and weight coefficients of the wireless signal detector, a triangular network with the wireless signal detector as the vertex is established through weighted triangulation. The coordinates of each triangle vertex are weighted according to the weight coefficient. The centroid coordinates of the weighted triangular network are solved by the least squares method. The centroid coordinates are used as the installation position coordinates of the vehicle tracker, and the installation position coordinates are sent to the display terminal.

[0045] A second aspect of an embodiment of the present invention provides a vehicle tracker detection and positioning system based on wireless signal analysis, comprising:

[0046] The first unit is used to deploy multiple wireless signal detectors around the vehicle to be detected, divide the signal collection area into multiple frequency band sub-areas and set sampling parameters for each frequency band sub-area;

[0047] The second unit is used to calculate the minimum sampling rate for each frequency band sub-region based on Shannon sampling theorem, perform piecewise fast Fourier transform to obtain spectrum data, recursively update the filter coefficients to obtain filtered spectrum data, construct an eight-layer wavelet packet decomposition tree using orthogonal basis functions to obtain decomposition nodes, calculate the energy entropy of the decomposition nodes to determine the time-frequency feature matrix, and generate signal feature vectors through long short-term memory network;

[0048] The third unit is used to determine whether the detected vehicle is equipped with a vehicle tracker based on the signal feature vector. If the determination result is yes, the signal feature vector is decomposed in the frequency domain to obtain multiple sets of frequency band features. The energy density distribution of each set of frequency band features is calculated to obtain a density feature map. The salient regions are extracted from the density feature map to obtain key signal features. The key signal features are segmented by adaptive thresholds to obtain the vehicle tracker type and initial position information.

[0049] The fourth unit is used to calculate the installation position coordinates of the vehicle tracker by using a weighted triangulation method according to the type and initial position information of the vehicle tracker and the difference in signal strength of the wireless signal detector, and send the calculated coordinates to the display terminal.

[0050] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including:

[0051] A processor and a memory for storing processor-executable instructions, wherein the processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0052] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0053] The present invention deploys multiple wireless signal detectors around the vehicle and employs segmented spectrum analysis and wavelet packet decomposition techniques to effectively identify the signal characteristics of vehicle trackers, thereby improving detection accuracy, reducing false alarm rates, adapting to signal characteristics of different frequency bands, and addressing interference issues in complex electromagnetic environments. Long-short-term memory networks are used to process signal feature vectors, combined with an adaptive threshold segmentation method, to accurately determine the type and initial position of the vehicle tracker, improving recognition accuracy and robustness. The method can also adapt to the characteristic differences of trackers of different models, providing a reliable foundation for subsequent precise positioning. Weighted triangulation is used to calculate the installation coordinates of the vehicle tracker, and the signal strength differences of multiple detectors are comprehensively utilized to overcome the insufficient accuracy of traditional positioning methods in complex vehicle structures. Centimeter-level positioning accuracy can be achieved, helping users quickly and accurately find and remove illegal tracking devices, effectively protecting personal privacy and driving safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 Schematic diagram of the flow of a vehicle tracker detection and positioning method based on wireless signal analysis according to an embodiment of the present invention;

[0055] Figure 2 Flowchart for generating signal feature vectors by inputting the time-frequency feature matrix into the long short-term memory network. DETAILED DESCRIPTION

[0056] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0057] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0058] Figure 1 FIG. 1 is a flow chart of a vehicle tracker detection and positioning method based on wireless signal analysis according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0059] Deploy multiple wireless signal detectors around the vehicle to be detected, divide the signal collection area into multiple frequency band sub-areas, and set sampling parameters for each frequency band sub-area;

[0060] For each frequency band sub-region, the minimum sampling rate is calculated based on Shannon's sampling theorem, and a piecewise fast Fourier transform is performed to obtain the spectrum data. The filtered spectrum data is obtained by recursively updating the filter coefficients. An eight-layer wavelet packet decomposition tree is constructed using orthogonal basis functions to obtain decomposition nodes. The energy entropy of the decomposition nodes is calculated to determine the time-frequency feature matrix, and the signal feature vector is generated through a long short-term memory network.

[0061] Based on the signal feature vector, it is determined whether the vehicle to be detected is equipped with a vehicle tracker. If the judgment result is yes, the signal feature vector is decomposed in the frequency domain to obtain multiple groups of frequency band features. The energy density distribution of each group of frequency band features is calculated to obtain a density feature map. The salient regions are extracted from the density feature map to obtain key signal features. The key signal features are adaptively segmented by threshold to obtain the vehicle tracker type and initial position information.

[0062] According to the vehicle tracker type and initial location information, combined with the signal strength difference of the wireless signal detector, the weighted triangulation method is used to calculate the installation location coordinates of the vehicle tracker and send them to the display terminal.

[0063] In an optional embodiment,

[0064] Deploy multiple wireless signal detectors around the vehicle to be detected, divide the signal collection area into multiple frequency band sub-areas, and set sampling parameters for each frequency band sub-area, including:

[0065] Deploy multiple wireless signal detectors around the vehicle to be detected;

[0066] The signal collection area is divided into three frequency band sub-areas: high frequency, medium frequency and low frequency according to the signal frequency characteristics. The frequency range of each frequency band sub-area is determined according to the working frequency band of the vehicle tracker;

[0067] Based on the frequency range and signal attenuation characteristics of each frequency band sub-region, the sampling frequency and sampling interval are adaptively calculated to generate corresponding sampling parameters, where the sampling frequency is not less than twice the highest frequency of the signal, and the sampling interval is determined based on the signal energy density distribution.

[0068] Multiple wireless signal detectors are deployed around the vehicle to be inspected. Select at least four wireless signal detectors and place them on the front, rear, left, and right sides of the vehicle, forming a detection network surrounding the vehicle. Each detector can be set to a distance of 0.5 to 2 meters from the vehicle, depending on the specific vehicle model and site conditions. The detectors use either omnidirectional or directional antennas. In the case of directional antennas, the main lobe of the antenna is directed toward the vehicle. These detectors are connected to a central processing unit via wired or wireless connections, enabling real-time data transmission and processing.

[0069] When dividing the signal collection area into three frequency bands based on signal frequency characteristics: high frequency, medium frequency, and low frequency, the low frequency sub-area is defined as 30MHz to 300MHz, the medium frequency sub-area is defined as 300MHz to 1GHz, and the high frequency sub-area is defined as 1GHz to 6GHz, based on the operating frequency bands of common vehicle trackers on the market. For example, the low frequency region can cover traditional VHF trackers, the medium frequency region covers UHF and some cellular network bands, and the high frequency region includes the frequency bands used by modern cellular networks, WiFi, Bluetooth, and other technologies.

[0070] The sampling frequency is adaptively calculated, ensuring it's no less than twice the highest signal frequency, based on the Nyquist sampling theorem. For the low-frequency sub-region (30MHz to 300MHz), the sampling frequency is set to at least 600MHz; for the mid-frequency sub-region (300MHz to 1GHz), the sampling frequency is set to at least 2GHz; and for the high-frequency sub-region (1GHz to 6GHz), the sampling frequency is set to at least 12GHz. In practical applications, to improve signal reconstruction quality, the sampling frequency is set to 2.5 times or higher of the highest signal frequency to avoid edge effects and undersampling.

[0071] The determination of the sampling interval is based on the adaptive adjustment of the signal energy density distribution, and the energy distribution profile of each frequency band is obtained by fast scanning. For the frequency band with higher energy density, a denser sampling interval is used; for the frequency band with lower energy density, a relatively sparse sampling interval is used. For example, when a signal with an energy density exceeding -70 dBm / Hz is detected in the low-frequency sub-region, the sampling interval can be set to 10 kHz; when the energy density is between -70 dBm / Hz and -90 dBm / Hz, the sampling interval can be set to 50 kHz; when the energy density is lower than -90 dBm / Hz, the sampling interval can be set to 100 kHz. Similarly, the corresponding sampling interval is set according to the energy density distribution in the medium-frequency and high-frequency sub-regions.

[0072] In actual application scenarios, when the system starts, the four detectors start working at the same time. Each detector performs full-band fast scanning to identify the frequency band where the signal may exist. Assuming that in the fast scanning, a signal with an energy density of -65 dBm / Hz is detected in the range of 120 MHz to 140 MHz in the low-frequency sub-region, the sampling interval of this frequency band is set to 10 kHz to achieve fine scanning; a signal with an energy density of -80 dBm / Hz is detected in the range of 850 MHz to 900 MHz in the medium-frequency sub-region, and the sampling interval of this frequency band is set to 100 kHz; a signal with an energy density of -75 dBm / Hz is detected in the range of 2.4 GHz to 2.5 GHz in the high-frequency sub-region, and the sampling interval of this frequency band is set to 200 kHz.

[0073] For different vehicle models, the sampling parameters are further optimized according to the vehicle material and structure. For example, for vehicle models with more metal shells, the low-frequency signal has stronger penetration ability, and the sampling density of the low-frequency sub-region is accordingly increased; for vehicle models using a large amount of composite materials, the high-frequency signal has less attenuation, and the scanning frequency of the high-frequency sub-region is increased.

[0074] When the detector collects signal data, the data of each detector is time-synchronized and spatially correlated for analysis. By comparing the time difference and intensity difference of the same signal received by different detectors, the approximate position of the signal source in the vehicle is calculated. For example, if the signal received by the left detector among the four detectors is the strongest and the earliest in time, the signal source is likely to be located in the left part of the vehicle. The positioning accuracy is controlled within 30 centimeters, which is sufficient to help users quickly find hidden trackers.

[0075] In this embodiment, the deployment of multiple detectors achieves all-round signal coverage of the space around the vehicle, avoiding the blind spots of single-point detection, improving the reliability and integrity of signal detection, dividing the signal collection area into three frequency band sub-areas: high frequency, medium frequency, and low frequency. The signal characteristics of different frequency bands are processed in a targeted manner, and the sampling interval is dynamically adjusted based on the signal energy density distribution. While ensuring the sampling quality, the utilization of system resources is optimized, achieving efficient and accurate positioning of the vehicle tracker. At the same time, it has strong versatility and scalability and can adapt to the needs of different application scenarios.

[0076] In an optional embodiment,

[0077] For each frequency band sub-region, the minimum sampling rate is calculated based on Shannon sampling theorem, and the spectrum data is obtained by performing piecewise fast Fourier transform. The filtered spectrum data is obtained by recursive iterative updating of the filter coefficients. An eight-layer wavelet packet decomposition tree is constructed using orthogonal basis functions to obtain decomposition nodes. The energy entropy value of the decomposition node is calculated to determine the time-frequency feature matrix and the signal feature vector is generated through the long short-term memory network.

[0078] For each frequency band sub-region, based on Shannon sampling theorem, the theoretical minimum sampling rate is calculated by twice the signal bandwidth. The oversampling factor is calculated according to the signal energy distribution. The theoretical minimum sampling rate is multiplied by the oversampling factor to determine the minimum sampling rate. The sampled signal is obtained based on the minimum sampling rate.

[0079] The sampled signal is segmented and processed, the window length is determined based on the signal time correlation, and the segmented signal is windowed with the cosine window function, and the segmented fast Fourier transform is performed to obtain the spectrum data;

[0080] The filter coefficients are updated through recursive iteration, the product of the input signal vector and the filter coefficient vector is calculated to obtain the predicted value, the predicted value and the expected output are calculated to obtain the error signal, and the filter coefficients are updated based on the error signal to obtain the filtered spectrum data;

[0081] Based on the filtered spectrum data, an eight-layer wavelet packet decomposition tree is constructed using orthogonal basis functions, and decomposition nodes are obtained through low-pass decomposition and high-pass decomposition;

[0082] Calculate the energy entropy value of the decomposed node, divide the square of the node coefficient by the sum of the squares of the node coefficients of the current layer to obtain the normalized probability, sum the normalized probability and the logarithmic product to obtain the energy entropy value, organize the energy entropy value into a time-frequency feature matrix according to the time-frequency distribution, and input the time-frequency feature matrix into the long short-term memory network to generate the signal feature vector.

[0083] For each frequency sub-region, the minimum sampling rate is calculated based on the Shannon sampling theorem. For a given signal, its frequency range is divided into multiple sub-regions, for example, a signal with a frequency range of 0-1000 Hz is divided into five sub-regions: 0-200 Hz, 200-400 Hz, 400-600 Hz, 600-800 Hz, and 800-1000 Hz. For each sub-region, according to the Shannon sampling theorem, twice the bandwidth of the sub-region signal is calculated as the theoretical minimum sampling rate. For example, for the 200-400 Hz sub-region, the bandwidth is 200 Hz, and the theoretical minimum sampling rate is 400 Hz. Considering the signal quality requirements in practical applications, an oversampling factor is introduced for adjustment. By analyzing the signal energy distribution, the oversampling factor value can be determined, usually set between 1.2 and 1.5. If the oversampling factor is selected as 1.25, the actual minimum sampling rate of the sub-region is 400 Hz x 1.25 = 500 Hz. Based on this sampling rate, the signal is sampled to obtain the sampled signal.

[0084] To obtain the frequency spectrum data, the sampled signal is segmented and fast Fourier transform is performed. The appropriate window length is determined according to the time correlation of the signal, for example, a longer window (such as 512 points) can be selected for a steady-state signal, while a shorter window (such as 128 points) can be selected for a transient signal. A cosine window function is selected to window the segmented signal to reduce spectral leakage. Fast Fourier transform is performed on the windowed signal segment to obtain the initial frequency spectrum data. For example, performing fast Fourier transform on a signal segment with a sampling rate of 500 Hz and a window length of 256 points can obtain frequency spectrum data with a frequency resolution of about 1.95 Hz in the range of 0-250 Hz.

[0085] To improve the quality of the frequency spectrum, the frequency spectrum data is optimized by adaptive filtering. The filter coefficients are updated using a recursive iteration method. The filter coefficient vector with a length of 32 is initialized, and all elements are set to 0.01. In each iteration, the product of the input signal vector and the current filter coefficient vector is calculated to obtain the predicted value. The predicted value is compared with the actual expected output to calculate the error signal. Based on the error signal and the learning rate (set to 0.05), the filter coefficients are updated. For example, for the update of the kth coefficient, the original coefficient is added to the product of the learning rate and the error and the corresponding input signal value. After 50 iterations, the filter coefficients obtained by the last update are used to process the original frequency spectrum data to obtain the filtered frequency spectrum data.

[0086] Based on the filtered spectral data, wavelet packet decomposition is used to extract time-frequency features. An eight-layer wavelet packet decomposition tree is constructed using orthogonal basis functions (such as the Daubechies-4 wavelet). Starting from the root node, each layer of nodes is decomposed using low-pass and high-pass filters, generating child nodes for the next layer. After eight layers of complete decomposition, a total of 255 decomposed nodes are obtained. For nodes in a given layer, low-pass filtering and downsampling are used to obtain low-frequency child nodes, while high-pass filtering and downsampling are used to obtain high-frequency child nodes. For example, a node in the third layer is processed through a 32nd-order low-pass filter and a high-pass filter, resulting in two fourth-layer nodes.

[0087] Energy entropy is calculated for each decomposed node to characterize the signal. The square of each wavelet coefficient in the node is divided by the sum of the squares of the coefficients of all nodes in the current layer to obtain a normalized probability. The product of each normalized probability and its natural logarithm is summed and negated to obtain the energy entropy value for that node. For example, a node contains 64 coefficients. The square of the coefficients is divided by the sum of the squares of the coefficients of all nodes in that layer to obtain 64 probability values. The sum of the products of the probabilities and their logarithms is negated to obtain the energy entropy value of the node, which is 4.83. The energy entropy values ​​of all nodes are organized into a time-frequency feature matrix based on their time-frequency distribution. This matrix reflects the energy distribution characteristics of the signal in different time-frequency domains.

[0088] The time-frequency feature matrix is ​​fed into a long short-term memory network to extract the signal feature vector. The network structure consists of an LSTM layer with 128 units and a fully connected layer with 64 neurons. The time-frequency feature matrix is ​​fed into the LSTM layer in time series to capture the temporal dependency of the features. The LSTM layer output is fed into the fully connected layer to obtain a 64-dimensional feature vector as the signal feature representation.

[0089] In this embodiment, the theoretical minimum sampling rate is calculated based on the Shannon sampling theorem, and an oversampling factor is introduced in combination with the signal energy distribution characteristics for dynamic adjustment, which not only ensures the sampling quality but also avoids resource waste. Through adaptive window division and cosine window function based on signal time correlation, the spectrum leakage phenomenon is effectively suppressed and the accuracy of spectrum analysis is improved. A recursive iterative adaptive filtering algorithm is adopted, and the filter coefficients are continuously optimized through the error signal, so that the filtering process can adaptively track the dynamic changes of signal characteristics, which not only ensures the accuracy of signal processing but also improves the reliability of feature extraction, laying a solid foundation for subsequent signal analysis and recognition.

[0090] In an optional embodiment,

[0091] Inputting the time-frequency feature matrix into the long short-term memory network to generate the signal feature vector includes:

[0092] Inputting the time-frequency feature matrix into the forward recursive branch and the backward recursive branch of the long short-term memory network respectively, and calculating the forward feature and the backward feature based on the feature state of the previous moment in the forward recursive branch and the feature state of the next moment in the backward recursive branch respectively;

[0093] The feature data of the current time step in the time-frequency feature matrix, the feature state of the previous moment in the forward recursive branch, and the feature state of the next moment in the backward recursive branch are concatenated to generate a concatenated result and the dynamic weight coefficient is obtained by mapping the result with a sigmoid function.

[0094] The feature data of the current time step in the time-frequency feature matrix is ​​concatenated with the forward feature of the previous time step and transformed to obtain the forward state; the feature data of the current time step in the time-frequency feature matrix is ​​concatenated with the backward feature of the next time step and transformed to obtain the backward state;

[0095] Multiply the forward state by the dynamic weight coefficient, multiply the backward state by the complement of the dynamic weight coefficient, add the two product results to obtain a fusion state, concatenate the fusion state with the feature data of the current time step in the time-frequency feature matrix, obtain the output gate parameters after transformation, and multiply the output gate parameters by the hyperbolic tangent function mapping result of the fusion state to obtain the signal feature vector.

[0096] Figure 2 A flowchart illustrating the process of inputting a time-frequency feature matrix into a long short-term memory network to generate signal eigenvectors. Before entering the long short-term memory network, the time-frequency feature matrix is ​​normalized to ensure that its values ​​are within an appropriate range and to avoid instability in network training due to excessively large or small values. Normalization uses a maximum-minimum scaling method to map the eigenvalues ​​to a range between zero and one, preserving the relative distribution of the original data. The normalized time-frequency feature matrix is ​​then fed into a bidirectional long short-term memory network consisting of a forward recursive branch and a backward recursive branch.

[0097] The forward recursive branch starts from the beginning of the time series and gradually advances the processing towards the end. When processing reaches a certain time step, the frequency feature vector corresponding to the time step is read, and the information of the previous moment stored in the internal state is used at the same time. The internal state includes long-term memory units and short-term memory units. The long-term memory unit stores the important features of the long-term span in the sequence, while the short-term memory unit reflects the changing trend of the recent time step. Through a complex gating mechanism, the frequency features of the current input are fused and calculated with the internal state of the previous moment, the long-term and short-term memory units are updated, and the forward feature output of the current time step is generated. Each row in the time-frequency feature matrix is ​​processed in turn to ensure that the forward feature contains the accumulated information from the beginning of the sequence to the current time step.

[0098] The backward recursive branch uses the opposite processing order as the forward recursive branch, starting from the end of the time series and gradually moving toward the starting point. The processing mechanism is similar to that of the forward branch. It also uses an internal gating structure to fuse the frequency features of the current time step with the state information of the next moment to generate backward features that reflect future information. This reverse processing method can capture the trend of sequence changes after the current time step, providing future time context for feature extraction. After the forward and backward recursive branches complete the processing of the entire time-frequency feature matrix, they generate corresponding forward and backward features for each time step.

[0099] The feature fusion phase processes each time step. For a specific time step, the corresponding frequency eigenvector in the time-frequency feature matrix is ​​extracted. The feature state at the previous moment in the forward recursive branch and the feature state at the next moment in the backward recursive branch are also obtained. These three sets of vectors are concatenated into a single long vector, with the frequency eigenvector first, the forward feature state in the middle, and the backward feature state at the end.

[0100] The concatenated long vectors are fed into a dedicated weight generation network layer, which uses a fully connected transformation to map the input into a set of weight coefficients. To ensure that the weight values ​​are between zero and one and have probabilistic meaning, a nonlinear activation function is applied to the transformed result to generate the final dynamic weight coefficients. The dynamic weight coefficients reflect the respective weights of historical and future information at the current time step.

[0101] The forward state and backward state are constructed separately. When constructing the forward state, the frequency feature vector of the current time step is concatenated with the forward feature generated in the previous time step. The concatenation result is processed through a dedicated state transition layer, which performs a linear transformation and maps the concatenated vector to an appropriate dimensional space. A nonlinear activation function is applied to enhance the expressive power, generating a forward state containing historical information. When constructing the backward state, the frequency feature vector of the current time step is concatenated with the backward feature generated in the next time step and processed through a state transition layer of the same structure to obtain a backward state containing future information.

[0102] The forward state is multiplied element by element by the dynamic weight coefficient to achieve weighted adjustment of historical information; the backward state is multiplied element by element by the complement of the dynamic weight coefficient (i.e. one minus the weight coefficient) to achieve weighted adjustment of future information.

[0103] The fused state is concatenated with the original frequency feature vector of the current time step and processed through the output transformation layer. The output transformation layer, which involves two steps, linear transformation and nonlinear activation, generates gating parameters used to control the final output. A hyperbolic tangent function is applied to the fused state for nonlinear transformation, mapping the value range to between -1 and +1 to obtain candidate output values. The gating parameters are element-wise multiplied by the candidate output values ​​to selectively retain the candidate outputs and generate a signal feature vector that reflects the comprehensive characteristics of the current time step.

[0104] This process is repeated for each time step in the time-frequency feature matrix, resulting in a series of signal feature vectors. These vectors are concatenated to form a feature sequence. In practical applications, the feature vector of the last time step can be used as the representation of the entire sequence. Alternatively, the feature vectors of all time steps can be combined into a unified representation vector through some aggregation method (such as average pooling or attention weighting) for subsequent vehicle tracker detection and localization tasks.

[0105] For example, in a vehicle tracker detection, a time-frequency feature matrix containing five time points and four frequency features is obtained. Each row in the matrix represents a time point, each column represents a frequency feature, and the value represents the corresponding energy entropy value. The first row: [0.2, 0.5, 0.3, 0.1], the second row: [0.3, 0.6, 0.2, 0.2], the third row: [0.5, 0.4, 0.4, 0.3], the fourth row: [0.4, 0.3, 0.6, 0.4], the fifth row: [0.2, 0.2, 0.5, 0.6];

[0106] Assuming that the hidden state dimension of the LSTM network is six, the forward recursive branch first processes the first row [0.2, 0.5, 0.3, 0.1] and generates the initial forward features [0.15, 0.22, 0.18, 0.25, 0.30, 0.20].

[0107] When processing the second row, the forward features of the first row are combined to generate the forward features of the second row [0.20, 0.28, 0.25, 0.30, 0.35, 0.22], and so on, to obtain the forward features corresponding to each row.

[0108] The backward recursive branch first processes the fifth row [0.2, 0.2, 0.5, 0.6] to generate the initial backward features [0.18, 0.25, 0.32, 0.28, 0.20, 0.15]. When processing the fourth row, the backward features of the fifth row are combined to generate the backward features of the fourth row [0.22, 0.30, 0.36, 0.32, 0.24, 0.18]. And so on, the backward features corresponding to each row are obtained.

[0109] Take the third row as an example for feature fusion: take the third row features [0.5, 0.4, 0.4, 0.3], the forward features of the previous time step [0.20, 0.28, 0.25, 0.30, 0.35, 0.22], and the backward features of the next time step [0.22, 0.30, 0.36, 0.32, 0.24, 0.18], and concatenate them into a long vector [0.5, 0.4, 0.4, 0.3, 0.20, 0.28, 0.25, 0.30, 0.35, 0.22, 0.22, 0.30, 0.36, 0.32, 0.24, 0.18].

[0110] Through weight generation network processing, the dynamic weight coefficients [0.6, 0.55, 0.48, 0.52, 0.58, 0.62] are obtained.

[0111] Concatenate and transform the third row of features with the forward features to obtain the forward state [0.35, 0.42, 0.38, 0.45, 0.52, 0.40].

[0112] Concatenate and transform the third row of features with the backward features to obtain the backward state [0.32, 0.38, 0.44, 0.40, 0.36, 0.30].

[0113] Applying dynamic weight coefficients to the forward state yields [0.21, 0.23, 0.18, 0.23, 0.30, 0.25].

[0114] Applying weight complement to the backward state yields [0.13, 0.17, 0.23, 0.19, 0.15, 0.11].

[0115] Adding the two together to get the fused state gives [0.34, 0.40, 0.41, 0.42, 0.45, 0.36].

[0116] Concatenate and transform the fused state with the third row of features to obtain the output gate parameters [0.52, 0.58, 0.56, 0.60, 0.62, 0.54], and apply the hyperbolic tangent function to the fused state to obtain [0.33, 0.38, 0.39, 0.40, 0.42, 0.35].

[0117] Multiplying the output gate parameters with the transformed fusion state yields the signal feature vector [0.17, 0.22, 0.22, 0.24, 0.26, 0.19].

[0118] Repeat the above process for all five time points to obtain five signal feature vectors. The feature vector of the fifth time point is selected as the signal feature vector, or the five vectors are averaged to obtain the signal feature vector for subsequent vehicle tracker detection and positioning.

[0119] In this embodiment, the forward and backward bidirectional feature extraction mechanism significantly reduces the risk of feature loss and improves the integrity and accuracy of feature extraction. The dynamic weight coefficient is generated by sigmoid function mapping, so that the feature fusion process can be adaptively adjusted according to the signal characteristics, reducing information redundancy in the feature fusion process and improving computational efficiency. The feature fusion mechanism based on the forward state and backward state enhances the system's anti-noise ability and can still maintain a high feature extraction accuracy in a low signal-to-noise ratio environment.

[0120] In an optional embodiment,

[0121] The signal feature vector is used to determine whether the vehicle to be detected is equipped with a vehicle tracker. If the judgment result is yes, the signal feature vector is decomposed in the frequency domain to obtain multiple sets of frequency band features. The energy density distribution of each set of frequency band features is calculated to obtain a density feature map. The key signal features are extracted from the density feature map to obtain the key signal features. The key signal features are segmented by adaptive threshold to obtain the vehicle tracker type and initial position information, including:

[0122] The signal feature vector is nonlinearly transformed through a two-layer neural network. The first-layer weight matrix of the two-layer neural network is multiplied by the signal feature vector and the first-layer bias vector is added, and then the result is mapped by the ReLU function. The mapping result is multiplied by the second-layer weight matrix and the second-layer bias vector is added, and then the result is processed by the softmax function to obtain the probability distribution of the discrimination result;

[0123] The signal eigenvector is decomposed in the frequency domain using a multi-bandwidth Hilbert-Huang transform. The intrinsic modal components are obtained through an adaptive bandpass filter bank. The intrinsic modal components are divided into multiple frequency bands. The multiple frequency bands are Fourier transformed and the energy density distribution is calculated to generate a density feature map.

[0124] A sliding window is constructed with each feature point in the density feature map as the center. The mean and standard deviation of the feature points in the sliding window are calculated. The difference between the feature point and the mean is divided by the standard deviation to obtain the local contrast. The local contrast is multiplied by the Gaussian weighted distance from the feature point to the central feature to construct a saliency map. The salient area is extracted as the key signal feature.

[0125] The key signal features are segmented using the maximum inter-class variance method, and the segmented features are matched with a preset tracker feature template library to obtain the vehicle tracker type. The energy centroid of the key signal features is calculated to obtain the initial position information of the vehicle tracker.

[0126] The signal feature vector contains electromagnetic signal data collected from the vehicle's surroundings. It is typically an 8192-dimensional feature vector, with each dimension corresponding to the signal strength value at a frequency point. Signal acquisition uses a broadband receiver with a frequency range of 300 MHz to 2.5 GHz and a sampling rate of 10 MHz.

[0127] After calculating the signal feature vector, a two-layer neural network performs a nonlinear transformation to determine whether the vehicle is equipped with a tracker. The first layer of the two-layer neural network contains 128 neurons, with a weight matrix of 8192×128 and a bias vector of 128×1. The signal feature vector is multiplied by the first-layer weight matrix to produce a 128-dimensional intermediate result, which is then added to the first-layer bias vector. The ReLU activation function is applied to this intermediate result, which takes the maximum value (0, x) for each element to produce a non-negative output. The second layer contains two neurons, with a weight matrix of 128×2 and a bias vector of 2×1. The ReLU output is multiplied by the second-layer weight matrix, added to the second-layer bias vector, and then converted into a probability distribution using the softmax function. The softmax function converts the two-dimensional output into two probability values ​​in the interval [0, 1] that sum to 1, representing the probabilities of "no tracker" and "presence of a tracker," respectively. When the probability of "presence of a tracker" is greater than 0.75, the vehicle is considered to be equipped with a tracker.

[0128] After determining that a vehicle is equipped with a tracker, the signal eigenvector is normalized and the amplitude is adjusted to the range of [-1, 1]. Using the Hilbert-Huang transform algorithm, the signal is decomposed into 10 intrinsic modal components, each of which contains signal characteristics of a specific frequency band. Each intrinsic modal component is filtered using an adaptive bandpass filter bank. The filter bank contains 10 bandpass filters with different center frequencies: 350MHz, 450MHz, 700MHz, 900MHz, 1.2GHz, 1.4GHz, 1.6GHz, 1.8GHz, 2.1GHz, and 2.4GHz, with a bandwidth of 10% of the center frequency. The resulting intrinsic modal components are divided into five frequency bands: low frequency (300-600MHz), mid-low frequency (600-1000MHz), mid-frequency (1000-1500MHz), mid-high frequency (1500-2000MHz), and high frequency (2000-2500MHz). A 512-point fast Fourier transform (FFT) is performed on the natural mode components within each frequency band to calculate the energy density distribution. The energy density is calculated as the square of the spectrum amplitude divided by the bandwidth. This yields an energy density feature map for each frequency band. Five density feature maps are generated, each with a size of 512 × 1.

[0129] Based on the density feature map, a saliency region is constructed to extract key signal features. A sliding window of size 21 is constructed with each feature point in the density feature map as the center. The mean and standard deviation of all feature points in the sliding window are calculated. The number of feature points in the window is 21. For each feature point, the difference between it and the window mean is calculated, and then divided by the standard deviation in the window to obtain the local contrast value of the feature point. The distance from the feature point to the central feature is calculated and weighted by a Gaussian function. The standard deviation of the Gaussian function is set to 7. The local contrast is multiplied by the Gaussian weighted distance to construct a saliency map. The area with a value greater than 0.6 in the saliency map is considered a saliency region and extracted as a key signal feature. Key signal features are usually located in the frequency band where the energy density mutates, accounting for 5%-15% of the original features, and retaining the main characteristics of the tracker signal.

[0130] Key signal features are segmented using the maximum inter-class variance method to determine the vehicle tracker type and initial location information. The maximum inter-class variance method automatically calculates the optimal segmentation threshold for key signal features, with a typical threshold range of 0.3-0.7. The segmented features are matched against a pre-set tracker feature template library, which contains feature templates for 50 common vehicle trackers. Cosine similarity is used to calculate the similarity between the features and the templates. The tracker type corresponding to the highest matching template with a similarity greater than 0.85 is the detection result. The energy centroid of the key signal features is calculated. The energy centroid is expressed as the energy-weighted average position of the feature points. It is obtained by multiplying the position of each feature point by its energy value and dividing it by the total energy. The position of the energy centroid corresponds to the initial location information of the vehicle tracker.

[0131] In this embodiment, by constructing a multi-level signal feature analysis system, accurate discrimination from feature vectors to tracker type identification is achieved. The nonlinear mapping of the two-layer neural network enhances the discrimination ability of feature classification. The ReLU activation function and the softmax output layer effectively improve the classification accuracy and robustness. The frequency domain decomposition method improves the discrimination of signal features. The density feature map generated by the energy density distribution highlights the key feature structure of the signal. The saliency detection mechanism constructed based on local contrast and Gaussian weighted distance improves the extraction accuracy of key signal features. Through the synergistic effect of multiple feature analysis and matching strategies, the accuracy of tracker identification and positioning is significantly improved, and it has strong environmental adaptability and anti-interference ability.

[0132] Table 1 is a system performance comparison table:

[0133]

[0134] As shown in Table 1, our method achieved a detection accuracy of 94.6%, significantly outperforming the 85.3% achieved by traditional spectrum analysis and the 92.1% achieved by deep learning methods, representing improvements of 9.3% and 2.5%, respectively. Our method also outperformed the other two methods in type recognition at 93.7%, a 12.5% ​​improvement compared to traditional methods.

[0135] Positioning accuracy is a key indicator of system practicality. This method achieved an error of only 14.9 cm, significantly superior to the 25.7 cm achieved by traditional methods and the 18.3 cm achieved by deep learning methods. In anti-interference testing (with a signal-to-noise ratio of 5 dB), this method maintained 88.4% performance, demonstrating strong environmental adaptability.

[0136] It is worth noting that in terms of computational complexity and processing time, the traditional spectrum analysis method performed best, with a processing time of only 150ms; this method was in the middle, requiring 215ms; while the deep learning method had high computational complexity and a processing time of up to 480ms.

[0137] In an optional embodiment,

[0138] The saliency map is constructed by multiplying the local contrast with the Gaussian weighted distance from the feature point to the central feature. The salient areas are extracted as key signal features, including:

[0139] Constructing a local feature manifold structure of the density feature map, constructing an adjacency matrix for the neighbor set of each feature point in the density feature map through a Gaussian kernel function, calculating the geodesic distance on the manifold based on the adjacency matrix through a shortest path algorithm, and performing an exponentially weighted summation of the neighborhood features of each feature point according to the geodesic distance to obtain a reconstructed feature;

[0140] Adaptively fusing the reconstructed feature with the local contrast to obtain enhanced local contrast, wherein the adaptive fusion is achieved by dynamically adjusting a fusion coefficient, the fusion coefficient is determined according to the local feature variance, and the Gaussian weighted distance from each feature point in the density feature map to the central feature is calculated by a negative exponential function of the Euclidean distance between the feature point value and the central feature;

[0141] A saliency map is constructed by multiplying the enhanced local contrast with the Gaussian weighted distance, and a salient region is extracted from the saliency map as a key signal feature by a non-maximum suppression algorithm.

[0142] Determine the nearest neighbor set for each feature point. Taking a 32×32 density feature map as an example, for the feature point at coordinates (15, 16), calculate the Euclidean distance between this feature point and all other feature points. Select the K points with the smallest distance as its nearest neighbor set, where K is 8. For example, the nearest neighbor set for this feature point might include feature points at coordinates (14, 16), (15, 15), (16, 16), (15, 17), (14, 15), (16, 15), (14, 17), and (16, 17). Construct the adjacency matrix W using the Gaussian kernel function. Each element W[i][j] in the matrix represents the similarity between feature points i and j, calculated as the square of the difference in eigenvalues ​​divided by the square of the parameter σ, where σ is 0.5 and then taken the negative exponential. For example, if the value of feature point i is 0.75 and the value of feature point j is 0.82, then W[i][j] is 0.9802. Based on the adjacency matrix, the Dijkstra algorithm is used to calculate the geodesic distance between any two points on the manifold. For each feature point in the density feature map, the reconstructed feature is obtained by performing an exponentially weighted sum of the features of its neighborhood based on the geodesic distance. If the neighborhood of feature point i contains 8 feature points, and the geodesic distances are 0.1, 0.15, 0.2, 0.25, 0.3, 0.35, 0.4, and 0.45, respectively, and the corresponding eigenvalues ​​are 0.78, 0.73, 0.81, 0.69, 0.77, 0.82, 0.75, and 0.74, then the reconstructed eigenvalue of feature point i is 0.766.

[0143] Adaptive fusion enhances local contrast. The local contrast of each feature point is calculated. For the feature point with coordinates (15, 16) in the density feature map, its local contrast is calculated based on the difference between the feature point and its neighboring features. The absolute average of the differences between the feature point value and the values ​​of all feature points in its 5×5 neighborhood is calculated. Assuming the calculated local contrast is 0.08, the reconstructed features are adaptively fused with the local contrast to obtain the enhanced local contrast. The fusion coefficient α is determined by the local feature variance. A larger local feature variance indicates a smaller fusion coefficient, indicating that the reconstructed feature should contribute less. For example, if the variance of the features in the 5×5 neighborhood is 0.02, α can be set to 1-0.02×10=0.8. The enhanced local contrast is calculated as α×local contrast+(1-α)×reconstructed feature, that is, 0.8×0.08+(1-0.8)×0.766=0.217. The Gaussian weighted distance from each feature point in the density feature map to the central feature is calculated. Assuming that the central feature of the density feature map is located at coordinates (16, 16) with a feature value of 0.85, and the feature point value at coordinates (15, 16) is 0.75, then the Euclidean distance between the two points is 0.1, and the Gaussian weighted distance is the negative exponential function e^(-0.1 2 / 0.3 2 )=0.89, where 0.3 is the width parameter of the Gaussian kernel.

[0144] A saliency map was constructed and salient regions were extracted. The saliency map was constructed by multiplying the enhanced local contrast with the Gaussian weighted distance. For the feature point at coordinates (15, 16), its saliency value was 0.217 × 0.89 = 0.193. After the saliency map was constructed, the salient regions were extracted from the saliency map as key signal features using a non-maximum suppression algorithm. A 3×3 sliding window was set. If the saliency value of the window center point was greater than the saliency values ​​of all other points within the window and greater than a preset threshold (e.g., 0.15), the point was marked as a key signal feature point. To enhance the robustness of the algorithm, a minimum distance constraint was set to ensure that the Euclidean distance between any two key feature points was no less than a preset value (e.g., 5 pixels). Twelve key signal feature points were extracted from a 32×32 density feature map. These key signal feature points can represent the important structural and variation characteristics of the signal.

[0145] In this embodiment, by constructing a local feature manifold structure and an adaptive feature fusion mechanism, accurate reconstruction and significance enhancement of features are achieved. The geodesic distance calculated by the shortest path algorithm truly reflects the geometric structure of the feature on the manifold, so that the feature reconstruction process better maintains the intrinsic structural characteristics of the data. The reconstructed features are combined with the local contrast through the adaptive fusion mechanism, and the fusion coefficient can be dynamically adjusted according to the local feature variance, thereby enhancing the adaptability of the feature fusion process to local structural changes. The multi-level feature enhancement and screening mechanism significantly improves the extraction accuracy of key signal features and enhances the system's ability to detect weak signals and suppress noise.

[0146] In an optional embodiment,

[0147] Based on the vehicle tracker type and initial location information, combined with the signal strength difference of the wireless signal detector, the weighted triangulation method is used to calculate the installation location coordinates of the vehicle tracker and send them to the display terminal, including:

[0148] Obtaining the type and initial location information of the vehicle tracker, and collecting the wireless signal strength emitted by the vehicle tracker within the detection range of each wireless signal detector;

[0149] calculating a signal strength difference between adjacent wireless signal detectors, and determining a weight coefficient of the vehicle tracker relative to the wireless signal detector based on the signal strength difference;

[0150] Based on the known coordinates and weight coefficients of the wireless signal detectors, a triangular network with the wireless signal detectors as the vertices is established by a weighted triangulation method, each triangular vertex coordinate is weighted according to the weight coefficients, the centroid coordinates of the weighted triangular network are solved by a least square method, the centroid coordinates are taken as the installation position coordinates of the vehicle tracker, and the installation position coordinates are sent to a display terminal.

[0151] Type information and initial position information of the vehicle tracker are acquired, the vehicle tracker types include GPS locators, Bluetooth trackers, wireless radio frequency identification trackers and the like, different types correspond to different signal transmission frequencies and power characteristics. Taking the wireless radio frequency identification tracker as an example, the working frequency is 433 MHz, the transmission power is 10 mW, and the initial position information can be the approximate coordinates of the vehicle parking area, such as the center point coordinates (100, 150) meters of the parking lot. In actual application, the vehicle tracker type and the initial position information can be acquired from a vehicle management database or provided by manual input.

[0152] A plurality of wireless signal detectors are deployed in the parking lot, each detector has a fixed known coordinate position. For example, in a rectangular parking lot, five wireless signal detectors are installed at coordinate points (0, 0), (0, 200), (200, 0), (200, 200), (100, 100) meters, respectively, and are numbered as detectors A, B, C, D and E. When the vehicle tracker is installed on the vehicle and activated, each wireless signal detector starts to collect the wireless signal strength value transmitted by the vehicle tracker, the collection frequency is 10 times per second, and the collection duration is 100 milliseconds each time.

[0153] For each wireless signal detector, 600 signal strength values of sampling points are acquired by continuous collection for 60 seconds, and stable signal strength readings are obtained by removing outliers and calculating the average value. For example, the signal strength values measured by detectors A, B, C, D and E are -65 dBm, -70 dBm, -75 dBm, -80 dBm and -60 dBm, respectively.

[0154] The signal strength difference between adjacent wireless signal detectors is calculated. Taking detectors A and B as an example, the signal strength difference is -65 dBm-(-70 dBm)=5 dBm. Similarly, the signal strength differences between A and C, A and D, A and E are 10 dBm, 15 dBm and -5 dBm, respectively; the signal strength differences between B and C, B and D, B and E are 5 dBm, 10 dBm and -10 dBm, respectively; the signal strength differences between C and D, C and E are 5 dBm and -15 dBm, respectively; and the signal strength difference between D and E is -20 dBm.

[0155] The weight coefficients are assigned to each wireless signal detector based on the signal strength difference. The weight assignment uses a non-linear mapping method. The higher the signal strength, the greater the weight coefficient, indicating a higher likelihood that the vehicle tracker is closer to the detector. After normalizing the signal strength of each detector, the weight values are converted through an exponential function. For example, the weakest signal of -80 dBm is mapped to a weight of 0.2, and the strongest signal of -60 dBm is mapped to a weight of 0.9. Other signals are proportionally assigned weights. The weight coefficients of detectors A, B, C, D, and E are 0.7, 0.6, 0.5, 0.2, and 0.9, respectively.

[0156] A triangular network is established with the wireless signal detectors as vertices. For five detectors, multiple triangles can be formed, including triangles ABC, ACD, ADE, ABE, and BCE. For each triangle, the weighted centroid coordinates are calculated using the known coordinates and weight coefficients of the vertices. Taking triangle ABC as an example, vertex A has coordinates (0, 0) and a weight of 0.7, vertex B has coordinates (0, 200) and a weight of 0.6, and vertex C has coordinates (200, 0) and a weight of 0.5. After weighting, the centroid coordinates of triangle ABC are (52.8, 80.0) meters.

[0157] The weighted centroid coordinates of all triangles are calculated. The weighted centroid coordinates of triangle ACD are (47.1, 28.6) meters, the weighted centroid coordinates of triangle ADE are (68.3, 120.5) meters, the weighted centroid coordinates of triangle ABE are (50.0, 97.2) meters, and the weighted centroid coordinates of triangle BCE are (85.0, 125.0) meters.

[0158] The weighted centroid coordinates of all triangles are collected as observation samples, and the optimal estimated position is solved by least squares method. The error sum of squares between each sample point and the point to be solved is calculated, and the point to be solved is adjusted through iteration to minimize the error sum of squares. After calculation, the installation position coordinates of the vehicle tracker are (60.5, 90.3) meters.

[0159] To improve positioning accuracy, historical position data and vehicle movement patterns of the vehicle tracker can also be combined to perform position prediction and correction through Kalman filtering algorithm. For example, if the historical data shows that the vehicle is usually parked near coordinates (65, 95) meters, the calculation result will be adjusted appropriately to obtain a coordinate value (63.2, 92.5) meters that is more consistent with the actual situation.

[0160] The finalized tracker installation coordinates (63.2, 92.5 meters) are transmitted via a network interface to a display terminal, which can be a computer screen at the monitoring center, a security officer's handheld device, or a vehicle owner's mobile app. After receiving the coordinate data, the display terminal marks the tracker's location on the parking lot floor plan and provides interactive functions such as zooming in and out and viewing historical track records to help users intuitively understand the tracker's installation location.

[0161] In this embodiment, multiple wireless signal detectors simultaneously collect the signal strength emitted by the tracker, avoiding the positioning error caused by single-point measurement and improving the reliability of signal collection. The weighted triangulation method converts the detector layout into a triangular network. The introduction of weight coefficients compensates for the attenuation error during signal propagation. The least squares method of solving the center of mass coordinates reduces the impact of measurement noise, improves the accuracy of the positioning results, and achieves precise positioning of the vehicle tracker installation position.

[0162] A second aspect of an embodiment of the present invention provides a vehicle tracker detection and positioning system based on wireless signal analysis, comprising:

[0163] The first unit is used to deploy multiple wireless signal detectors around the vehicle to be detected, divide the signal collection area into multiple frequency band sub-areas and set sampling parameters for each frequency band sub-area;

[0164] The second unit is used to calculate the minimum sampling rate for each frequency band sub-region based on Shannon sampling theorem, perform piecewise fast Fourier transform to obtain spectrum data, recursively update the filter coefficients to obtain filtered spectrum data, construct an eight-layer wavelet packet decomposition tree using orthogonal basis functions to obtain decomposition nodes, calculate the energy entropy of the decomposition nodes to determine the time-frequency feature matrix, and generate signal feature vectors through long short-term memory network;

[0165] The third unit is used to determine whether the detected vehicle is equipped with a vehicle tracker based on the signal feature vector. If the determination result is yes, the signal feature vector is decomposed in the frequency domain to obtain multiple sets of frequency band features. The energy density distribution of each set of frequency band features is calculated to obtain a density feature map. The salient regions are extracted from the density feature map to obtain key signal features. The key signal features are segmented by adaptive thresholds to obtain the vehicle tracker type and initial position information.

[0166] The fourth unit is used to calculate the installation position coordinates of the vehicle tracker by using a weighted triangulation method according to the type and initial position information of the vehicle tracker and the difference in signal strength of the wireless signal detector, and send the calculated coordinates to the display terminal.

[0167] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including:

[0168] A processor and a memory for storing processor-executable instructions, wherein the processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0169] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0170] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0171] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A vehicle tracker detection and positioning method based on wireless signal analysis, characterized in that: include: Deploy multiple wireless signal detectors around the vehicle to be detected, divide the signal collection area into multiple frequency band sub-areas, and set sampling parameters for each frequency band sub-area; For each frequency band sub-region, the minimum sampling rate is calculated based on the Shannon sampling theorem, and the spectrum data is obtained by performing a piecewise fast Fourier transform. The filtered spectrum data is obtained by recursively updating the filter coefficients. An eight-layer wavelet packet decomposition tree is constructed using orthogonal basis functions to obtain decomposition nodes. The energy entropy of the decomposition nodes is calculated to determine the time-frequency feature matrix and generate the signal feature vector through the long short-term memory network, including: For each frequency band sub-region, based on Shannon sampling theorem, the theoretical minimum sampling rate is calculated by twice the signal bandwidth. The oversampling factor is calculated according to the signal energy distribution. The theoretical minimum sampling rate is multiplied by the oversampling factor to determine the minimum sampling rate. The sampled signal is obtained based on the minimum sampling rate. The sampled signal is segmented and processed, the window length is determined based on the signal time correlation, and the segmented signal is windowed with the cosine window function, and the segmented fast Fourier transform is performed to obtain the spectrum data; The filter coefficients are updated through recursive iteration, the product of the input signal vector and the filter coefficient vector is calculated to obtain the predicted value, the predicted value and the expected output are calculated to obtain the error signal, and the filter coefficients are updated based on the error signal to obtain the filtered spectrum data; Based on the filtered spectrum data, an eight-layer wavelet packet decomposition tree is constructed using orthogonal basis functions, and decomposition nodes are obtained through low-pass decomposition and high-pass decomposition; Calculate the energy entropy of the decomposed node, divide the square of the node coefficient by the sum of the squares of the node coefficients in the current layer to obtain the normalized probability, sum the normalized probability and the logarithm product to obtain the energy entropy value, organize the energy entropy value into a time-frequency feature matrix according to the time-frequency distribution, and input the time-frequency feature matrix into the long short-term memory network to generate the signal feature vector; Based on the signal feature vector, it is determined whether the vehicle to be detected is equipped with a vehicle tracker. If the judgment result is yes, the signal feature vector is decomposed in the frequency domain to obtain multiple groups of frequency band features. The energy density distribution of each group of frequency band features is calculated to obtain a density feature map. The salient regions are extracted from the density feature map to obtain key signal features. The key signal features are adaptively segmented by threshold to obtain the vehicle tracker type and initial position information. According to the vehicle tracker type and initial location information, combined with the signal strength difference of the wireless signal detector, the weighted triangulation method is used to calculate the installation location coordinates of the vehicle tracker and send them to the display terminal.

2. The method according to claim 1, characterized in that Deploy multiple wireless signal detectors around the vehicle to be detected, divide the signal collection area into multiple frequency band sub-areas, and set sampling parameters for each frequency band sub-area, including: Deploy multiple wireless signal detectors around the vehicle to be detected; The signal collection area is divided into three frequency band sub-areas: high frequency, medium frequency and low frequency according to the signal frequency characteristics. The frequency range of each frequency band sub-area is determined according to the working frequency band of the vehicle tracker; Based on the frequency range and signal attenuation characteristics of each frequency band sub-region, the sampling frequency and sampling interval are adaptively calculated to generate corresponding sampling parameters, where the sampling frequency is not less than twice the highest frequency of the signal, and the sampling interval is determined based on the signal energy density distribution.

3. The method according to claim 1, characterized in that Inputting the time-frequency feature matrix into the long short-term memory network to generate the signal feature vector includes: Inputting the time-frequency feature matrix into the forward recursive branch and the backward recursive branch of the long short-term memory network respectively, and calculating the forward feature and the backward feature based on the feature state of the previous moment in the forward recursive branch and the feature state of the next moment in the backward recursive branch respectively; The feature data of the current time step in the time-frequency feature matrix, the feature state of the previous moment in the forward recursive branch, and the feature state of the next moment in the backward recursive branch are concatenated to generate a concatenated result and the dynamic weight coefficient is obtained by mapping the result with a sigmoid function. The feature data of the current time step in the time-frequency feature matrix is ​​concatenated with the forward feature of the previous time step and transformed to obtain the forward state; the feature data of the current time step in the time-frequency feature matrix is ​​concatenated with the backward feature of the next time step and transformed to obtain the backward state; Multiply the forward state by the dynamic weight coefficient, multiply the backward state by the complement of the dynamic weight coefficient, add the two product results to obtain a fusion state, concatenate the fusion state with the feature data of the current time step in the time-frequency feature matrix, obtain the output gate parameters after transformation, and multiply the output gate parameters by the hyperbolic tangent function mapping result of the fusion state to obtain the signal feature vector.

4. The method according to claim 1, wherein The signal feature vector is used to determine whether the vehicle to be detected is equipped with a vehicle tracker. If the judgment result is yes, the signal feature vector is decomposed in the frequency domain to obtain multiple sets of frequency band features. The energy density distribution of each set of frequency band features is calculated to obtain a density feature map. The key signal features are extracted from the density feature map to obtain the key signal features. The key signal features are segmented by adaptive threshold to obtain the vehicle tracker type and initial position information, including: The signal feature vector is nonlinearly transformed through a two-layer neural network. The first-layer weight matrix of the two-layer neural network is multiplied by the signal feature vector and the first-layer bias vector is added, and then the result is mapped by the ReLU function. The mapping result is multiplied by the second-layer weight matrix and the second-layer bias vector is added, and then the result is processed by the softmax function to obtain the probability distribution of the discrimination result; The signal eigenvector is decomposed in the frequency domain using a multi-bandwidth Hilbert-Huang transform. The intrinsic modal components are obtained through an adaptive bandpass filter bank. The intrinsic modal components are divided into multiple frequency bands. The multiple frequency bands are Fourier transformed and the energy density distribution is calculated to generate a density feature map. A sliding window is constructed with each feature point in the density feature map as the center. The mean and standard deviation of the feature points in the sliding window are calculated. The difference between the feature point and the mean is divided by the standard deviation to obtain the local contrast. The local contrast is multiplied by the Gaussian weighted distance from the feature point to the central feature to construct a saliency map. The salient area is extracted as the key signal feature. The key signal features are segmented using the maximum inter-class variance method, and the segmented features are matched with a preset tracker feature template library to obtain the vehicle tracker type. The energy centroid of the key signal features is calculated to obtain the initial position information of the vehicle tracker.

5. The method according to claim 4, characterized in that The saliency map is constructed by multiplying the local contrast with the Gaussian weighted distance from the feature point to the central feature. The salient areas are extracted as key signal features, including: Constructing a local feature manifold structure of the density feature map, constructing an adjacency matrix for the neighbor set of each feature point in the density feature map through a Gaussian kernel function, calculating the geodesic distance on the manifold based on the adjacency matrix through a shortest path algorithm, and performing an exponentially weighted summation of the neighborhood features of each feature point according to the geodesic distance to obtain a reconstructed feature; Adaptively fusing the reconstructed feature with the local contrast to obtain enhanced local contrast, wherein the adaptive fusion is achieved by dynamically adjusting a fusion coefficient, the fusion coefficient is determined according to the local feature variance, and the Gaussian weighted distance from each feature point in the density feature map to the central feature is calculated by a negative exponential function of the Euclidean distance between the feature point value and the central feature; A saliency map is constructed by multiplying the enhanced local contrast with the Gaussian weighted distance, and a salient region is extracted from the saliency map as a key signal feature by a non-maximum suppression algorithm.

6. The method according to claim 1, characterized in that Based on the vehicle tracker type and initial location information, combined with the signal strength difference of the wireless signal detector, the weighted triangulation method is used to calculate the installation location coordinates of the vehicle tracker and send them to the display terminal, including: Obtaining the type and initial location information of the vehicle tracker, and collecting the wireless signal strength emitted by the vehicle tracker within the detection range of each wireless signal detector; calculating a signal strength difference between adjacent wireless signal detectors, and determining a weight coefficient of the vehicle tracker relative to the wireless signal detector based on the signal strength difference; Based on the known coordinates and weight coefficients of the wireless signal detector, a triangular network with the wireless signal detector as the vertex is established through weighted triangulation. The coordinates of each triangle vertex are weighted according to the weight coefficient. The centroid coordinates of the weighted triangular network are solved by the least squares method. The centroid coordinates are used as the installation position coordinates of the vehicle tracker, and the installation position coordinates are sent to the display terminal.

7. A vehicle tracker detection and positioning system based on wireless signal analysis, used to implement the method according to any one of claims 1 to 6, characterized in that: include: The first unit is used to deploy multiple wireless signal detectors around the vehicle to be detected, divide the signal collection area into multiple frequency band sub-areas and set sampling parameters for each frequency band sub-area; The second unit is used to calculate the minimum sampling rate for each frequency band sub-region based on Shannon sampling theorem, perform piecewise fast Fourier transform to obtain spectrum data, recursively update the filter coefficients to obtain filtered spectrum data, construct an eight-layer wavelet packet decomposition tree using orthogonal basis functions to obtain decomposition nodes, calculate the energy entropy of the decomposition nodes to determine the time-frequency feature matrix, and generate signal feature vectors through long short-term memory network; The third unit is used to determine whether the detected vehicle is equipped with a vehicle tracker based on the signal feature vector. If the determination result is yes, the signal feature vector is decomposed in the frequency domain to obtain multiple sets of frequency band features. The energy density distribution of each set of frequency band features is calculated to obtain a density feature map. The salient regions are extracted from the density feature map to obtain key signal features. The key signal features are segmented by adaptive thresholds to obtain the vehicle tracker type and initial position information. The fourth unit is used to calculate the installation position coordinates of the vehicle tracker by using a weighted triangulation method according to the type and initial position information of the vehicle tracker and the difference in signal strength of the wireless signal detector, and send the calculated coordinates to the display terminal.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images