Cross-regional AI trajectory tracking method based on multi-information fusion

Abstract

This invention discloses a cross-regional AI trajectory tracking method based on multi-information fusion, specifically relating to the field of information fusion technology. It extracts behavioral pattern features and data acquisition environment features of the target from video surveillance, GPS, and sensor data across multiple time periods. Based on these features, it assigns accuracy weights to the feature matching algorithm for each time period, and obtains the overall algorithm recognition accuracy index through weighted averaging. The recognition accuracy index is compared with a preset threshold, automatically classifying data into accurate and inaccurate recognition. For accurately recognized data, an adaptive learning mechanism is input to optimize system parameters. For inaccurately recognized data, the recognition parameters are dynamically adjusted by analyzing deviations, further optimizing the feature matching algorithm. This reduces the risk of misjudgment and incorrect tracking, prevents unauthorized personnel from accidentally entering the monitored path, and prevents the system from adaptively learning based on erroneous data, ensuring the accuracy and consistency of tracking.

Description

Technical Field

[0001] This invention relates to the field of information fusion technology, and more specifically to a cross-regional AI trajectory tracking method based on multi-information fusion. Background Technology

[0002] Cross-regional AI trajectory tracking based on multi-information fusion is a technology that combines multiple data sources and intelligent algorithms, primarily used for accurate tracking of target trajectories across different geographical regions. The core of this technology lies in fusing heterogeneous data from multiple sources, such as video surveillance, GPS positioning, social media data, and sensor data. This data is integrated across spatiotemporal dimensions, and then valuable information is extracted through artificial intelligence algorithms. Utilizing deep learning models and data analysis techniques, the system can connect scattered and incomplete trajectory information to reconstruct the target's movement path, thereby achieving continuous tracking across regions.

[0003] The implementation of this technology typically relies on the synergy of large-scale data processing and deep learning algorithms. Specifically, the system first preprocesses monitoring data from different regions using techniques such as image recognition and pattern matching to extract distinctive feature information. Then, a multi-information fusion algorithm is used to compare and correlate these features with data from other sources, eliminating errors between different regions and devices, and achieving unified and accurate trajectory mapping. This cross-regional AI trajectory tracking technology has wide application value in fields such as intelligent traffic management and public safety monitoring, and can greatly improve the efficiency and accuracy of cross-regional target tracking.

[0004] The existing technology has the following shortcomings:

[0005] In integrating heterogeneous data, systems typically rely on feature matching algorithms to associate data from different sources with the same target. However, feature matching algorithms may misidentify the same feature as the same target, leading to incorrect identification and the tracking of unrelated individuals. This is especially problematic when dealing with large datasets, such as video surveillance and social media showing different people with similar features. The system may fail to distinguish them accurately, resulting in innocent individuals being included in the surveillance path. Such mistracking not only allows the true target to escape but also leads to the continuous monitoring of innocent individuals. Furthermore, when the system mistakenly tracks unrelated individuals, the adaptive learning mechanism of the cross-regional AI system will train and optimize based on the erroneous data, leading to further misjudgments. The system may use the feature data of unrelated individuals as target features for deep learning, adjusting recognition parameters and resulting in more subsequent misidentifications and tracking. Summary of the Invention

[0006] The purpose of this invention is to provide a cross-regional AI trajectory tracking method based on multi-information fusion to address the shortcomings of the prior art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a cross-regional AI trajectory tracking method based on multi-information fusion, comprising the following steps:

[0008] S1: Extract behavioral pattern features and data acquisition environment features of the same target from multi-source heterogeneous data collected over several time periods. The multi-source heterogeneous data includes video surveillance data, GPS data, and sensor data.

[0009] S2: Based on the extracted behavioral pattern features of the same target and the features of the data acquisition environment, determine the accuracy weight of the feature matching algorithm in identifying the same target in each time period, and calculate the algorithm recognition accuracy index by weighted averaging of the accuracy weight of the feature matching algorithm in identifying the same target in each time period.

[0010] S3: Compare and analyze the algorithm's recognition accuracy index with the preset accuracy threshold, divide the accuracy of the feature matching algorithm in recognizing the same target into different categories, and divide the recognition results into accurate recognition and inaccurate recognition respectively;

[0011] S4: For accurately identified multi-source heterogeneous data, input it into the adaptive learning mechanism to optimize the system's identification and tracking parameters for target features; for inaccurately identified multi-source heterogeneous data, analyze the deviation of multi-source heterogeneous data within a fixed time period, and dynamically adjust the identification parameters based on the analysis results to optimize the feature matching algorithm.

[0012] Preferably, in S2, after analyzing the extracted behavioral pattern features of the same target, an activity frequency fluctuation index is generated. The method for obtaining the activity frequency fluctuation index is as follows:

[0013] Time series data of the behavior patterns of the same target are collected and labeled as x(t), representing the activity intensity of the target at each time point. A window length W and an overlap O are set to ensure partial overlap between windows. The Discrete Fourier Transform (DFT) is used to transform the time series data from the time domain to the frequency domain. The expression for calculating the Fourier Transform is: Where: X(f) is a complex value in the frequency domain, N is the total number of data points in the time series, and f is the frequency, with a range of values ​​of... Where j is the imaginary unit, the formula for calculating the power spectral density is: Where: P(f) is the power spectral density at frequency f. It is the modulus of the Fourier coefficients, that is Re(X(f)) and Im(X(f)) are the real and imaginary parts of X(f), respectively; the activity frequency fluctuation index is defined as the ratio of the intensity of the high-frequency component to the low-frequency component, representing the intensity of the activity frequency fluctuation, and its calculation expression is: ;in, It is a set of frequencies in the high-frequency range. It is a set of frequencies in the low-frequency range. This is the activity frequency fluctuation index.

[0014] Preferably, in S2, after analyzing the data acquisition environment features of the extracted same target, a monitoring angle conversion anomaly index is generated. The method for obtaining the monitoring angle conversion anomaly index is as follows:

[0015] Acquire multi-source feature data of the target during the monitoring process, standardize the collected feature data, and define the standardized feature data as matrix X, where each row represents the feature data at a time point and each column represents a feature dimension; calculate the covariance matrix Σ of feature matrix X, the formula of which is: Where T is the matrix transpose and Q is the number of samples in the feature data. Let Σ be the feature vector of the i-th sample, and μ be the mean vector of the feature data. Eigenvalue decomposition is performed on the covariance matrix Σ to obtain eigenvalues ​​and eigenvectors. Principal components, i.e., the first k eigenvectors with the largest eigenvalues, are selected to construct the dimensionality reduction matrix W. The original feature data X is then projected into the dimensionality reduction space to obtain the dimensionality-reduced feature matrix. The expression is: Use the dimensionality-reduced feature data To calculate the Mahalanobis distance between each time point and the overall characteristic distribution. The formula for calculating Mahalanobis distance is: ;in, It is the feature vector of the i-th dimension-reduced sample. It is the mean vector of the feature data after dimensionality reduction. This is the covariance matrix of the feature data after dimensionality reduction. Anomalies in viewpoint transitions are detected based on Mahalanobis distance. A threshold θ is set, and samples with a Mahalanobis distance greater than the threshold θ are considered anomalies. The monitoring viewpoint transition anomaly index is calculated, expressed as: Where M is the number of samples greater than the threshold. It is the standard deviation of all Mahalanobis distances. Anomaly index for monitoring angle conversion.

[0016] Preferably, in S2, the activity frequency fluctuation index and the monitoring angle conversion anomaly index are converted into a comprehensive feature vector. The comprehensive feature vector is used as the input of the machine learning model. The machine learning model uses the prediction of the accuracy weight assignment label of the feature matching algorithm for identifying the same target in each time period as the prediction objective, and minimizes the sum of prediction errors of the accuracy weight assignment labels of the feature matching algorithm for identifying the same target in all time periods as the training objective. The machine learning model is trained until the sum of prediction errors converges and the model training stops. The accuracy weight assignment of the feature matching algorithm for identifying the same target in each time period is determined according to the model output. The machine learning model is a multinomial regression model, and the accuracy index of the algorithm is obtained by weighted averaging the accuracy weight assignments of the feature matching algorithm for identifying the same target in each time period.

[0017] Preferably, in S3, the algorithm's recognition accuracy index is compared and analyzed with a preset accuracy threshold. The accuracy of the feature matching algorithm in recognizing the same target is divided into different categories, and the recognition results are divided into accurate recognition and inaccurate recognition, specifically:

[0018] The calculated algorithm recognition accuracy index is compared with the preset accuracy threshold. If the algorithm recognition accuracy index is greater than or equal to the preset accuracy threshold, it means that the feature matching algorithm has high accuracy in recognizing the same target. In this case, no warning signal is generated, and the recognition result is classified as accurate recognition. If the algorithm recognition accuracy index is less than the preset accuracy threshold, it means that the feature matching algorithm has low accuracy in recognizing the same target. In this case, a warning signal is generated, and the recognition result is classified as inaccurate recognition.

[0019] Preferably, in S4, for inaccurately identified multi-source heterogeneous data, i.e., the algorithm recognition accuracy index generated within a fixed time period is less than a preset accuracy threshold, the algorithm recognition accuracy index generated within subsequent fixed time periods that is less than the preset accuracy threshold is collected, and a corresponding dataset is established. The mean and standard deviation of the dataset are calculated, and the deviation of the multi-source heterogeneous data is judged after analysis. The recognition parameters are dynamically adjusted to optimize the feature matching algorithm.

[0020] Preferably, if the mean of the algorithm recognition accuracy index in the dataset is greater than or equal to the reference threshold of the mean of the algorithm recognition accuracy index, and the standard deviation of the algorithm recognition accuracy index is less than the reference threshold of the standard deviation of the algorithm recognition accuracy index, the overall recognition accuracy is good and stable. In this case, the recognition parameters can be adjusted to improve the fault tolerance of the system.

[0021] If the mean of the algorithm's recognition accuracy index is greater than or equal to the reference threshold for the mean of the algorithm's recognition accuracy index, and the standard deviation of the algorithm's recognition accuracy index is greater than or equal to the reference threshold for the standard deviation of the algorithm's recognition accuracy index, it indicates that the accuracy index fluctuates greatly and the system performs inconsistently in different scenarios. The recognition weights should be adjusted according to environmental conditions to enhance the algorithm's adaptability.

[0022] If the mean of the algorithm's recognition accuracy index is less than the reference threshold for the mean of the algorithm's recognition accuracy index, and the standard deviation of the algorithm's recognition accuracy index is greater than or equal to the reference threshold for the standard deviation of the algorithm's recognition accuracy index, the accuracy is low and fluctuates greatly, indicating that the recognition effect is not ideal and unstable. The algorithm should be re-optimized and the matching parameters adjusted.

[0023] If the mean of the algorithm's recognition accuracy index is less than the reference threshold for the mean of the algorithm's recognition accuracy index, and the standard deviation of the algorithm's recognition accuracy index is less than the reference threshold for the standard deviation of the algorithm's recognition accuracy index, it indicates that the overall recognition accuracy is low and stable. The algorithm parameters should be re-evaluated, and the feature extraction method should be strengthened to improve the recognition accuracy.

[0024] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0025] 1. This invention achieves precise tracking of cross-regional targets by integrating heterogeneous data from multiple sources such as video surveillance, GPS, and sensors, and combining behavioral pattern characteristics and data acquisition environment characteristics. This method dynamically calculates and evaluates the recognition accuracy of the feature matching algorithm, comparing the algorithm's recognition accuracy index with a preset threshold to classify the recognition results as accurate or inaccurate. For accurately identified data, the system further optimizes parameters through an adaptive learning mechanism; for inaccurately identified data, it analyzes the deviation and dynamically adjusts the parameters to avoid the accumulation of mistracking and misjudgment. This mechanism significantly improves the system's accuracy and robustness, preventing unauthorized personnel from accidentally entering the tracking path and preventing erroneous data from interfering with the system's learning.

[0026] 2. This invention accurately measures changes in target behavior and environmental characteristics through an activity frequency fluctuation index and a monitoring angle conversion anomaly index. It then uses a multinomial regression model to train these comprehensive feature vectors using machine learning to minimize identification errors and dynamically optimize matching parameters. For inaccurately identified datasets, the system dynamically adjusts parameters through mean and standard deviation analysis, improving the system's fault tolerance and adaptability based on the volatility of the accuracy index and overall performance. Ultimately, this effectively solves the problems of misjudgment and mistracking in cross-regional tracking of existing systems, enhances adaptability to different scenarios and environmental conditions, and ensures the accuracy and consistency of target identification. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0028] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] For examples, please refer to Figure 1 As shown in this embodiment, the cross-regional AI trajectory tracking method based on multi-information fusion includes the following steps:

[0031] S1: Extract behavioral pattern features and data acquisition environment features of the same target from multi-source heterogeneous data collected over several time periods. The multi-source heterogeneous data includes video surveillance data, GPS data, and sensor data.

[0032] S2: Based on the extracted behavioral pattern features of the same target and the features of the data acquisition environment, determine the accuracy weight of the feature matching algorithm in identifying the same target in each time period, and calculate the algorithm recognition accuracy index by weighted averaging of the accuracy weight of the feature matching algorithm in identifying the same target in each time period.

[0033] S3: Compare and analyze the algorithm's recognition accuracy index with the preset accuracy threshold, divide the accuracy of the feature matching algorithm in recognizing the same target into different categories, and divide the recognition results into accurate recognition and inaccurate recognition respectively;

[0034] S4: For accurately identified multi-source heterogeneous data, input it into the adaptive learning mechanism to optimize the system's identification and tracking parameters for target features; for inaccurately identified multi-source heterogeneous data, analyze the deviation of multi-source heterogeneous data within a fixed time period, and dynamically adjust the identification parameters based on the analysis results to optimize the feature matching algorithm.

[0035] In S1, behavioral pattern features of the same target and data acquisition environment features are extracted from multi-source heterogeneous data collected over several time periods. The multi-source heterogeneous data includes video surveillance data, GPS data, and sensor data. Specifically:

[0036] Video surveillance data typically contains visual features and behavioral patterns of the target. Using computer vision technology, the following features can be extracted from video data:

[0037] Appearance features: These include physical characteristics such as clothing color, shape, and accessories. This information helps identify targets and distinguish them from other individuals. Common extraction methods include convolutional neural networks (CNNs) and feature encoding methods in deep learning, which transform images into embedded features.

[0038] Motion patterns: By utilizing the target's motion trajectory, behavioral pattern features, such as direction of travel, speed, and gait characteristics, can be extracted. Target tracking algorithms (such as optical flow, KLT feature point tracking, or detection-tracking algorithms like YOLO and DeepSort) can be used to obtain the continuous positional changes of the target in video frames, thereby calculating its speed and direction.

[0039] Actions and postures: Pose recognition algorithms can identify a target's action state (such as standing, walking, running, etc.) and the posture of its arms and legs. This information helps analyze the target's behavioral habits. For example, open-source pose estimation algorithms such as OpenPose or MediaPipe can track key points on the human body in real time.

[0040] GPS data provides accurate geographic location information and is an important source for determining target location and behavioral patterns. GPS data mainly includes the following characteristics:

[0041] Location coordinates: The absolute location of a target can be obtained using latitude and longitude data from a GPS device. By mapping GPS coordinates onto a map, the specific geographical location of the target, such as streets and buildings, can be identified.

[0042] Movement speed: GPS devices typically collect data every few seconds. By calculating the distance and time between adjacent locations, the movement speed of the target can be estimated, helping to identify any abnormal behavior (such as sudden acceleration, sudden stop, etc.).

[0043] Stops and routes: By analyzing the frequency with which a target stays in a certain place for an extended period or repeatedly traverses a certain route, the target's habitual activity areas and typical paths can be identified. This information helps to construct the target's daily activity patterns.

[0044] Sensor data (such as accelerometer, gyroscope, RFID, Bluetooth beacon, etc.) provides additional information about the microscopic behavioral patterns of a target. Here are some key characteristics of sensor data:

[0045] Acceleration and angular velocity: Accelerometers and gyroscopes can record changes in the target's acceleration and angular velocity. This data can be used to analyze the target's motion state (such as walking, running, turning), and by analyzing the patterns of acceleration changes, it can also be determined whether the target has made any sudden movements (such as falling or stopping abruptly).

[0046] Near-field location characteristics: Near-field communication sensors such as RFID and Bluetooth beacons are used to identify the near-field location of a target. For example, RFID devices can be used to accurately track subtle changes in the target's position within a specific area, which is very useful when the target frequently enters or leaves a small area.

[0047] Environmental data: Some sensors may provide environmental condition data (such as temperature, humidity, and light intensity), which can help analyze the external environmental characteristics of the target. For example, through a light sensor, the system can identify whether the target is indoors, outdoors, or in an environment with varying light intensity.

[0048] Feature data from different data sources needs to be integrated to form a holistic target behavior pattern. The main methods include:

[0049] Spatiotemporal alignment: Align all data sources according to the timeline and integrate spatial locations. The absolute position of GPS, the relative position of video surveillance, and the near-field position of sensors need to be transformed into a unified coordinate system to ensure the integrity of behavioral patterns.

[0050] Behavioral pattern construction: By associating features such as speed, acceleration, and movement from different data sources, a complete behavioral pattern can be generated. For example, combining gait information from a video with GPS walking speed can more accurately describe the target's walking style.

[0051] Environmental characteristics describe the contextual information of data collection, which may affect the target's behavior and the usability of the data. These mainly include the following aspects:

[0052] Lighting conditions: Video surveillance and certain sensors are sensitive to lighting conditions; the clarity of video data may decrease under low light conditions. This characteristic can be extracted from brightness information in video images or from light sensors to determine whether the current acquisition environment is suitable for visual feature recognition.

[0053] Weather conditions: GPS and sensor data are sensitive to weather conditions; for example, heavy rain may affect GPS signals. Combining this data with external meteorological data can help determine the weather conditions at the time of data collection and aid in the analysis of potential data errors.

[0054] Terrain and Built Environment: Location data from GPS and video surveillance can determine the geographical environment of a target, such as whether it is located in an area with dense high-rise buildings or an indoor location. This information is helpful in identifying signal interference, positioning errors, and behavioral patterns.

[0055] Data acquisition device status: The hardware status of the device (such as camera resolution and GPS signal strength) is also an environmental characteristic. For example, under conditions of low battery, weak signal, or unstable sensors, the acquired data may contain noise or be incomplete, and needs to be marked for processing during data fusion.

[0056] S2: Based on the extracted behavioral pattern features of the same target and the features of the data acquisition environment, determine the accuracy weight of the feature matching algorithm in identifying the same target in each time period, and calculate the algorithm recognition accuracy index by weighted averaging of the accuracy weight of the feature matching algorithm in identifying the same target in each time period.

[0057] After analyzing the extracted behavioral pattern features of the same target, an activity frequency fluctuation index is generated. The method for obtaining the activity frequency fluctuation index is as follows:

[0058] Collect time-series data on the behavioral patterns of the same target, including changes in movement speed, step frequency, or displacement. Label this as x(t), representing the intensity of the target's activity at each time point.

[0059] To analyze short-term activity frequency changes, long-term series can be divided into multiple short time windows (e.g., 10 seconds or 1 minute). Performing a Fourier transform on each window separately can better capture frequency fluctuations. Setting the window length W and overlap O ensures partial overlap between windows, guaranteeing the continuity of frequency changes.

[0060] The Discrete Fourier Transform (DFT) is used to transform time series data from the time domain to the frequency domain. The expression for calculating the Fourier Transform is as follows: Where: X(f) is a complex value in the frequency domain (Fourier coefficients), reflecting the amplitude and phase of different frequency components; N is the total number of data points in the time series; and f is the frequency, with a value range of... Where j is the imaginary unit, the power spectral density is calculated. The power spectral density reflects the energy distribution of each frequency component and can be used to measure the intensity of activity at different frequencies. The calculation formula is: Where: P(f) is the power spectral density at frequency f. It is the magnitude (amplitude) of the Fourier coefficients, i.e. Re(X(f)) and Im(X(f)) are the real and imaginary parts of X(f), respectively.

[0061] The activity frequency fluctuation index is used to quantify the variation of activity frequency across different frequency bands. It can be defined as the ratio of the relative intensity of high-frequency components to low-frequency components, representing the intensity of activity frequency fluctuations. The calculation expression is: ;in, It is a set of frequencies in the high-frequency range, usually corresponding to faster frequency components. For example, it can be defined as components with frequencies greater than a certain threshold (such as 1 Hz). It is a set of frequencies in the low-frequency range, typically corresponding to slower frequency components. For example, it can be defined as components with frequencies less than or equal to this threshold. This is the activity frequency fluctuation index.

[0062] The magnitude of the activity frequency fluctuation index can be used to assess the accuracy of feature matching algorithms in identifying the same target. A low activity frequency fluctuation index indicates that the target's behavior pattern is relatively stable, with its activity frequency predominating in the low-frequency range. This suggests that the target's action rhythm, path, and behavior pattern remain consistent across different time periods. Such stability typically implies higher accuracy in identifying the same target across different data sources, as the consistency of target behavior enhances the reliability of feature matching. Furthermore, a smaller fluctuation index indicates less noise influence, reducing the possibility of false matches, thus allowing the system to track the target's continuous activity more accurately.

[0063] When the activity frequency fluctuation index is high, the target's behavior pattern changes significantly at high frequencies, indicating frequent behavioral changes, acceleration, sudden stops, or turns within a short period. This high volatility reduces the accuracy of feature matching algorithms because the irregularity of behavior can lead to inconsistencies in features across different time periods, increasing the risk of mismatches. For example, frequent changes in direction or gait by the target can cause significant deviations in feature matching from video surveillance, GPS, or sensor data, making the system more likely to misidentify it as a different target. Therefore, a high volatility index suggests that the algorithm needs further analysis to determine the possibility of mismatches, especially in the fusion of multi-source heterogeneous data, where matching results should be judged with greater caution.

[0064] After analyzing the data acquisition environment features of the extracted same target, a monitoring angle conversion anomaly index is generated. The method for obtaining the monitoring angle conversion anomaly index is as follows:

[0065] Acquire multi-source feature data of the target during the monitoring process, including video features from different angles (such as position, pose, appearance, etc.). Standardize the collected feature data (e.g., zero-mean and unit variance) to ensure that different features are relatively consistent in numerical scale, thus avoiding bias introduced by different feature scales. Assume the standardized feature data is a matrix X, where each row represents the feature data at a time point, and each column represents a feature dimension.

[0066] The covariance matrix Σ of the characteristic matrix X is calculated using the following formula: Where T is the matrix transpose and Q is the number of samples in the feature data. Let Σ be the feature vector of the i-th sample, and μ be the mean vector of the feature data. Eigenvalue decomposition is performed on the covariance matrix Σ to obtain eigenvalues ​​and eigenvectors. Principal components, i.e., the first k eigenvectors with the largest eigenvalues, are selected to construct the dimensionality reduction matrix W. The original feature data X is then projected into the dimensionality reduction space to obtain the dimensionality-reduced feature matrix. The expression is: Use the dimensionality-reduced feature data To calculate the Mahalanobis distance between each time point and the overall characteristic distribution. The formula for calculating Mahalanobis distance is: ;in, It is the feature vector of the i-th dimension-reduced sample. It is the mean vector of the feature data after dimensionality reduction. This is the covariance matrix of the feature data after dimensionality reduction. Anomalies in viewpoint transitions are detected based on Mahalanobis distance. A threshold θ is set, and samples with a Mahalanobis distance greater than the threshold are considered anomalies. The monitoring viewpoint transition anomaly index is calculated, expressed as: Where M is the number of samples exceeding the threshold. It is the standard deviation of all Mahalanobis distances. Anomaly index for monitoring angle conversion.

[0067] The magnitude of the anomaly index during monitoring angle transitions directly affects the accuracy of feature matching algorithms in identifying the same target. A low anomaly index indicates that the change in monitoring perspective does not significantly affect the target's features, meaning that the target's appearance, posture, and position remain highly consistent across different viewpoints. This consistency typically implies higher accuracy for feature matching algorithms in identifying the same target across multiple viewpoints, as the minimal change in features makes it easier for the matching algorithm to identify identical target features. Therefore, a lower anomaly index indicates stronger stability and higher accuracy in identifying the same target.

[0068] A high anomaly index during monitoring angle transitions indicates that the target's features change significantly from different perspectives. This could be due to changes in appearance features caused by angle changes, occlusion, or differences in lighting. A high anomaly index typically means that the target's features are unlikely to match those from other perspectives at the current angle, increasing the risk of misidentification for the feature matching algorithm. For example, if the target is partially occluded or distorted after a viewpoint change, the algorithm might mistakenly identify it as another object. Therefore, a high anomaly index indicates decreased accuracy in identifying the same target, requiring additional matching strategies or auxiliary data to enhance the algorithm's reliability.

[0069] The activity frequency fluctuation index and the monitoring angle conversion anomaly index are converted into a comprehensive feature vector. This comprehensive feature vector is used as the input to a machine learning model. The machine learning model uses the accuracy weight assignment label of the feature matching algorithm for identifying the same target in each time period as the prediction objective, and minimizes the sum of prediction errors of the accuracy weight assignment labels of the feature matching algorithm for identifying the same target in all time periods as the training objective. The machine learning model is trained until the sum of prediction errors converges, at which point the model training stops. The accuracy weight assignment of the feature matching algorithm for identifying the same target in each time period is determined based on the model output. The machine learning model is a multinomial regression model, and the accuracy index of the algorithm is obtained by weighted averaging the accuracy weight assignments of the feature matching algorithm for identifying the same target in each time period.

[0070] The method for obtaining the accuracy weights for feature matching algorithms in identifying the same target across different time periods is as follows: Obtain the corresponding function expression from the comprehensive feature vector training data of the trained machine learning model. In the formula, This is the model's output function, where WE is the activity frequency fluctuation index and SD is the monitoring angle conversion anomaly index. Assign weights to the accuracy of feature matching algorithms in identifying the same target within each time period.

[0071] S3: Compare and analyze the algorithm's recognition accuracy index with the preset accuracy threshold, divide the accuracy of the feature matching algorithm in recognizing the same target into different categories, and divide the recognition results into accurate recognition and inaccurate recognition respectively.

[0072] The calculated algorithm recognition accuracy index is compared with the preset accuracy threshold. If the algorithm recognition accuracy index is greater than or equal to the preset accuracy threshold, it means that the feature matching algorithm has high accuracy in recognizing the same target. In this case, no warning signal is generated, and the recognition result is classified as accurate recognition. If the algorithm recognition accuracy index is less than the preset accuracy threshold, it means that the feature matching algorithm has low accuracy in recognizing the same target. In this case, a warning signal is generated, and the recognition result is classified as inaccurate recognition.

[0073] S4: For accurately identified multi-source heterogeneous data, input it into the adaptive learning mechanism to optimize the system's identification and tracking parameters for target features; for inaccurately identified multi-source heterogeneous data, analyze the deviation of multi-source heterogeneous data within a fixed time period, and dynamically adjust the identification parameters based on the analysis results to optimize the feature matching algorithm.

[0074] First, the accurately identified data undergoes denoising and smoothing to ensure data stability. For example, noise in video data is removed, and transient biases in GPS location data are smoothed to reduce the impact of outliers on model training. Core features of the target (such as appearance, location information, acceleration, and behavioral patterns) are extracted from data from video surveillance, GPS, and sensors, and these features are standardized to ensure consistency in feature scale across different data sources, facilitating subsequent adaptive learning processing.

[0075] To accurately identify the target, labels are generated from the data. These labels can be based on known target IDs from historical data or newly generated identifiers, ensuring that the adaptive learning mechanism can be accurately trained on specific targets. For features from different perspectives or sources in multi-source data, the target labels are kept consistent, and features of the same target are calibrated so that the adaptive learning mechanism can handle a consistent feature set.

[0076] Accurately identified data is input into the deep learning or machine learning model (such as neural networks, decision trees, support vector machines, etc.) within the adaptive learning mechanism for further training on the target features. This allows the system to progressively enhance its ability to recognize target features, particularly in terms of behavior, posture, and appearance, thereby improving its feature recognition capabilities. After training the model based on accurate data, the system updates key tracking parameters, such as optimizing the similarity threshold in video tracking and adjusting the distance threshold in GPS tracking. This process helps the system maintain high recognition accuracy and robustness even when features are inconsistent or the environment changes.

[0077] During adaptive learning, the system should verify the model update effect in real time. The model's recognition accuracy is tested using accurately identified data. If the model performs well, the current update is retained; if the recognition accuracy does not improve or decreases, retraining or adjusting parameters such as the learning rate may be necessary. When target features change over time (such as clothing, gait, and movement speed), the adaptive learning mechanism dynamically adjusts feature weights based on the latest accurate data, ensuring the model can adapt to changes in the target in real time, thereby improving the continuity of tracking.

[0078] The adaptively learned and optimized recognition model is deployed into the real-time tracking system, making its recognition and tracking of the same target more accurate in practical applications. Based on the updated model obtained from adaptive learning, the system can optimize multi-source heterogeneous data fusion strategies, such as automatically selecting the best camera angle and adjusting the sensor sampling frequency, thereby enhancing the system's recognition and tracking capabilities in complex scenes.

[0079] For multi-source heterogeneous data that is not accurately identified, i.e., the algorithm identification accuracy index generated within a fixed time period is less than the preset accuracy threshold, the algorithm identification accuracy index generated within subsequent fixed time periods that is less than the preset accuracy threshold is collected, and a corresponding dataset is established. The mean and standard deviation of the dataset are calculated, and the bias of the multi-source heterogeneous data is determined after analysis. The identification parameters are then dynamically adjusted to optimize the feature matching algorithm.

[0080] If the mean of the algorithm's recognition accuracy index within the dataset is greater than or equal to a reference threshold, and the standard deviation of the algorithm's recognition accuracy index is less than a reference threshold for its standard deviation, it indicates that the average value of the algorithm's recognition accuracy index is high, meaning that most data points are close to or exceed the reference threshold. Simultaneously, a smaller standard deviation implies lower data volatility and more stable changes in the accuracy index. In this case, the consistency of multi-source heterogeneous data is high, indicating that the current feature matching algorithm performs stably. In this situation, the algorithm's tolerance can be appropriately increased, and the accuracy requirements of the recognition parameters can be slightly relaxed. For example, increasing the matching threshold or reducing the sensitivity to feature similarity can improve the system's fault tolerance. This adjustment helps reduce unnecessary false alarms while maintaining stability.

[0081] If the mean of the algorithm's recognition accuracy index is greater than or equal to a reference threshold, and the standard deviation of the algorithm's recognition accuracy index is greater than or equal to a reference threshold for the standard deviation of the algorithm's recognition accuracy index, the high mean but large standard deviation indicates high volatility in the dataset, suggesting inconsistent algorithm performance under different monitoring perspectives or environmental conditions. For example, the feature matching algorithm may perform poorly under certain specific perspectives, leading to fluctuations in recognition accuracy. In this case, an environment-related weight adjustment mechanism can be introduced. The feature matching parameters can be dynamically adjusted based on different perspectives or specific conditions (such as changes in lighting or location occlusion) to improve recognition stability. Different feature matching weights or tolerances can be set for periods of unstable performance to reduce the impact of fluctuations on overall recognition accuracy.

[0082] If the mean of the algorithm's recognition accuracy index is less than the reference threshold for the mean of the algorithm's recognition accuracy index, and the standard deviation of the algorithm's recognition accuracy index is greater than or equal to the reference threshold for the standard deviation of the algorithm's recognition accuracy index, then the algorithm's recognition accuracy index is low and its standard deviation is large. This indicates that not only is the overall recognition accuracy low, but the data also exhibits high volatility and significant instability. This suggests that the current feature matching algorithm's recognition performance is unsatisfactory for most time periods, and the recognition performance varies significantly across different time periods. This may be due to large biases between multi-source data or significant environmental noise, leading to inconsistent algorithm recognition accuracy. In this case, substantial parameter optimization and algorithm adjustments are required. For example, the reliance on noisy data sources can be reduced by lowering their weights, or more data cleaning and preprocessing steps can be used to reduce data bias. In addition, the matching parameters of key features (such as matching thresholds, similarity calculation methods, etc.) can be re-optimized to better adapt to recognition needs under different environmental conditions.

[0083] If the mean of the algorithm's recognition accuracy index is less than the reference threshold for the mean of the algorithm's recognition accuracy index, and the standard deviation of the algorithm's recognition accuracy index is less than the reference threshold for the standard deviation of the algorithm's recognition accuracy index, the low mean and small standard deviation of the algorithm's recognition accuracy index indicate that the system's recognition accuracy is generally poor and has low volatility, exhibiting consistent low accuracy. This suggests that the feature matching algorithm is struggling to accurately identify the same target under the current parameter settings, possibly because the algorithm's parameter configuration is unsuitable for the characteristics of the current multi-source data, or because there is a systematic bias between different data sources. In this case, the overall configuration of the feature matching algorithm needs to be re-evaluated and optimized. The weights of each data source can be reallocated, or a new feature matching method can be introduced to improve recognition accuracy. Furthermore, it is recommended to increase the amount of data used for model training, especially data containing different scenes and perspectives, to improve the model's generalization ability in different environments. Simultaneously, improvements to noise filtering and feature extraction methods can be considered to ensure that the algorithm can more effectively capture the core features of the target.

[0084] In this embodiment, firstly, behavioral pattern features and data acquisition environment features of the same target are extracted from video surveillance, GPS, and sensor data collected over multiple time periods. Next, based on these features, accuracy weights are assigned to the feature matching algorithm for each time period, and a weighted average is calculated to obtain the overall algorithm recognition accuracy index. Then, the recognition accuracy index is compared with a preset threshold, and the results are categorized as accurate or inaccurate recognition. Finally, for accurately recognized data, an adaptive learning mechanism is input to optimize the system's recognition and tracking parameters. For inaccurately recognized data, multi-source data bias is analyzed, and recognition parameters are dynamically adjusted to optimize the feature matching algorithm, thereby improving the system's recognition accuracy and stability in changing environments.

[0085] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0086] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0087] It should be understood that the term "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone, where A and B can be singular or plural. Additionally, the character " / " in this document generally indicates an "or" relationship between the preceding and following related objects, but it may also indicate an "and / or" relationship; please refer to the context for specific understanding. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0088] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A cross-regional AI trajectory tracking method based on multi-information fusion, characterized by: Includes the following steps; S1: Extract behavioral pattern features and data acquisition environment features of the same target from multi-source heterogeneous data collected over several time periods. The multi-source heterogeneous data includes video surveillance data, GPS data, and sensor data. S2: Based on the extracted behavioral pattern features of the same target and the features of the data acquisition environment, determine the accuracy weight of the feature matching algorithm in identifying the same target in each time period, and calculate the algorithm recognition accuracy index by weighted averaging of the accuracy weight of the feature matching algorithm in identifying the same target in each time period. Among them, the activity frequency fluctuation index is generated after analyzing the extracted behavioral pattern features of the same target. The method for obtaining the activity frequency fluctuation index is as follows: Time series data of the behavior patterns of the same target are collected and labeled as x(t), representing the activity intensity of the target at each time point. A window length W and an overlap O are set to ensure partial overlap between windows. The Discrete Fourier Transform (DFT) is used to transform the time series data from the time domain to the frequency domain. The expression for calculating the Fourier Transform is: Where: X(f) is a complex value in the frequency domain, N is the total number of data points in the time series, and f is the frequency, with a range of values ​​of... Where j is the imaginary unit, the formula for calculating the power spectral density is: Where: P(f) is the power spectral density at frequency f. It is the modulus of the Fourier coefficients, that is Re(X(f)) and Im(X(f)) are the real and imaginary parts of X(f), respectively; the activity frequency fluctuation index is defined as the ratio of the intensity of the high-frequency component to the low-frequency component, representing the intensity of the activity frequency fluctuation, and its calculation expression is: ;in, It is a set of frequencies in the high-frequency range. It is a set of frequencies in the low-frequency range. This is an index representing the frequency fluctuation of activity. S3: Compare and analyze the algorithm's recognition accuracy index with the preset accuracy threshold, divide the accuracy of the feature matching algorithm in recognizing the same target into different categories, and divide the recognition results into accurate recognition and inaccurate recognition respectively; S4: For accurately identified multi-source heterogeneous data, input it into the adaptive learning mechanism to optimize the system's identification and tracking parameters for target features; for inaccurately identified multi-source heterogeneous data, analyze the deviation of multi-source heterogeneous data within a fixed time period, and dynamically adjust the identification parameters based on the analysis results to optimize the feature matching algorithm.

2. The cross-regional AI trajectory tracking method based on multi-information fusion according to claim 1, characterized in that: In S2, after analyzing the data acquisition environment features of the extracted same target, a monitoring angle conversion anomaly index is generated. The method for obtaining the monitoring angle conversion anomaly index is as follows: Acquire multi-source feature data of the target during the monitoring process, standardize the collected feature data, and set the standardized feature data as matrix X, where each row represents the feature data of a time point and each column represents a feature dimension. The covariance matrix Σ of the characteristic matrix X is calculated using the following formula: Where T is the matrix transpose and Q is the number of samples in the feature data. Let Σ be the feature vector of the i-th sample, and μ be the mean vector of the feature data. Eigenvalue decomposition is performed on the covariance matrix Σ to obtain eigenvalues ​​and eigenvectors. Principal components, i.e., the first k eigenvectors with the largest eigenvalues, are selected to construct the dimensionality reduction matrix W. The original feature data X is then projected into the dimensionality reduction space to obtain the dimensionality-reduced feature matrix. The expression is: Use the dimensionality-reduced feature data To calculate the Mahalanobis distance between each time point and the overall characteristic distribution. The formula for calculating Mahalanobis distance is: ;in, It is the feature vector of the i-th dimension-reduced sample. It is the mean vector of the feature data after dimensionality reduction. This is the covariance matrix of the feature data after dimensionality reduction. Anomalies in viewpoint transitions are detected based on Mahalanobis distance. A threshold θ is set, and samples with a Mahalanobis distance greater than the threshold θ are considered anomalies. The monitoring viewpoint transition anomaly index is calculated, expressed as: Where M is the number of samples greater than the threshold. It is the standard deviation of all Mahalanobis distances. Anomaly index for monitoring angle conversion.

3. The cross-regional AI trajectory tracking method based on multi-information fusion according to claim 2, characterized in that: In S2, the activity frequency fluctuation index and the monitoring angle conversion anomaly index are converted into a comprehensive feature vector. The comprehensive feature vector is used as the input of the machine learning model. The machine learning model uses the accuracy weight assignment label of the feature matching algorithm for identifying the same target in each time period as the prediction objective, and minimizes the sum of prediction errors of the accuracy weight assignment labels of the feature matching algorithm for identifying the same target in all time periods as the training objective. The machine learning model is trained until the sum of prediction errors converges and the model training stops. The accuracy weight assignment of the feature matching algorithm for identifying the same target in each time period is determined based on the model output. The machine learning model is a multinomial regression model, and the accuracy index of the algorithm is obtained by weighted averaging the accuracy weight assignments of the feature matching algorithm for identifying the same target in each time period.

4. The cross-regional AI trajectory tracking method based on multi-information fusion according to claim 3, characterized in that: In S3, the algorithm's recognition accuracy index is compared and analyzed with a preset accuracy threshold. The accuracy of the feature matching algorithm in recognizing the same target is divided into different categories, and the recognition results are divided into accurate recognition and inaccurate recognition, specifically: The calculated algorithm recognition accuracy index is compared with the preset accuracy threshold. If the algorithm recognition accuracy index is greater than or equal to the preset accuracy threshold, it means that the feature matching algorithm has high accuracy in recognizing the same target. At this time, no warning signal is generated, and the recognition result is classified as accurate recognition. If the algorithm's recognition accuracy index is less than the preset accuracy threshold, it indicates that the feature matching algorithm has low accuracy in recognizing the same target. At this time, a warning signal is generated, and the recognition result is classified as inaccurate recognition.

5. The cross-regional AI trajectory tracking method based on multi-information fusion according to claim 1, characterized in that: In S4, for inaccurately identified multi-source heterogeneous data, i.e., the algorithm recognition accuracy index generated within a fixed time period is less than the preset accuracy threshold, the algorithm recognition accuracy index generated within subsequent fixed time periods that is less than the preset accuracy threshold is collected, and a corresponding dataset is established. The mean and standard deviation of the dataset are calculated, and the deviation of the multi-source heterogeneous data is judged after analysis. The recognition parameters are dynamically adjusted to optimize the feature matching algorithm.

6. The cross-regional AI trajectory tracking method based on multi-information fusion according to claim 5, characterized in that: If the mean of the algorithm recognition accuracy index in the dataset is greater than or equal to the reference threshold of the mean of the algorithm recognition accuracy index, and the standard deviation of the algorithm recognition accuracy index is less than the reference threshold of the standard deviation of the algorithm recognition accuracy index, the overall recognition accuracy is good and stable. At this time, the recognition parameters can be adjusted to improve the fault tolerance of the system. If the mean of the algorithm's recognition accuracy index is greater than or equal to the reference threshold for the mean of the algorithm's recognition accuracy index, and the standard deviation of the algorithm's recognition accuracy index is greater than or equal to the reference threshold for the standard deviation of the algorithm's recognition accuracy index, it indicates that the accuracy index fluctuates greatly and the system performs inconsistently in different scenarios. The recognition weights should be adjusted according to environmental conditions to enhance the algorithm's adaptability. If the mean of the algorithm's recognition accuracy index is less than the reference threshold for the mean of the algorithm's recognition accuracy index, and the standard deviation of the algorithm's recognition accuracy index is greater than or equal to the reference threshold for the standard deviation of the algorithm's recognition accuracy index, the accuracy is low and fluctuates greatly, indicating that the recognition effect is not ideal and unstable. The algorithm should be re-optimized and the matching parameters adjusted. If the mean of the algorithm's recognition accuracy index is less than the reference threshold for the mean of the algorithm's recognition accuracy index, and the standard deviation of the algorithm's recognition accuracy index is less than the reference threshold for the standard deviation of the algorithm's recognition accuracy index, it indicates that the overall recognition accuracy is low and stable. The algorithm parameters should be re-evaluated, and the feature extraction method should be strengthened to improve the recognition accuracy.

Images