A campus safety monitoring method based on image recognition technology

Abstract

This invention belongs to the field of image sensing, specifically disclosing a campus security monitoring method based on image recognition technology. The method includes: adjusting the clock deviation between target devices using a time synchronization method based on the original video frame sequence collected by multiple sensor devices to obtain a synchronization timestamp sequence; extracting personnel target boundary information from the video frames of each monitoring point based on the synchronization timestamp sequence, generating multi-dimensional feature data using a target detection method, and verifying the consistency of the cross-frame features; if the similarity between the feature data meets a preset condition, associating the corresponding personnel targets with the same identity to obtain the preliminary identity association structure; and obtaining a set of personnel trajectory data during cross-regional movement based on the preliminary identity association structure. The purpose of this invention is to solve the problems of inaccurate identity association and discontinuous trajectory caused by clock deviation, light changes, and angle differences during cross-regional movement in existing image sensing technologies.

Description

Technical Field

[0001] This invention relates to the field of image sensing technology, and more specifically to a campus security monitoring method based on image recognition technology. Background Technology

[0002] In a comprehensive surveillance network, building an accurate and dynamic personnel identification and trajectory tracking system is a core technical challenge, especially in complex scenarios involving multiple devices and multiple areas. Ensuring the continuity of personnel identification and the integrity of their trajectories is a pressing issue. Specifically, when collecting video data through multiple sensor devices, clock deviations between different devices and environmental factors such as changes in lighting and angles can cause inconsistencies in the personnel target feature data extracted from the original video frames. This can affect the accuracy of identity association, leading to situations where the same person is misidentified as different individuals when moving across areas, or different individuals are incorrectly associated with the same identity.

[0003] Furthermore, during trajectory tracking, the movement trajectory of personnel between different monitoring points may be interrupted due to data loss or feature matching failure, especially in densely populated areas or monitoring blind spots. This continuity problem directly affects the dynamic management effect of the whole-domain monitoring. Around this issue, some related minor issues have also arisen, such as how to effectively extract the consistency of cross-frame features based on synchronized timestamps, how to correct feature data affected by the environment to improve matching accuracy, and how to distinguish between normal flow and potential risk behavior in trajectory anomaly detection. These minor issues are all closely related to the core goal of improving the reliability and real-time performance of identity recognition and trajectory tracking.

[0004] Especially in real-world business scenarios, such as large public places or transportation hubs, where there is frequent movement of people and complex environments, the coverage and viewing angles of monitoring equipment vary significantly. People may quickly cross multiple monitoring areas in a short period of time. If it is not possible to accurately associate identities and fully reconstruct trajectories, it will be difficult to provide timely warnings and responses to abnormal behaviors, thereby affecting the overall security management effectiveness. Summary of the Invention

[0005] This invention provides a campus security monitoring method based on image recognition technology, aiming to solve the problems of inaccurate identity association and discontinuous trajectory caused by clock deviation, light changes and angle differences when the target moves across regions during image sensing in the existing technology.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0007] A campus security monitoring method based on image recognition technology includes: adjusting the clock deviation between target devices using a time synchronization method based on the original video frame sequence collected by multi-sensor devices to obtain a synchronization timestamp sequence; extracting personnel target boundary information from video frames of each monitoring point based on the synchronization timestamp sequence, generating multi-dimensional feature data using a target detection method, and verifying the cross-frame feature consistency; if the similarity between the feature data meets a preset condition, associating the corresponding personnel targets with the same identity to obtain the preliminary identity association structure; obtaining a set of personnel trajectory data during cross-regional movement based on the preliminary identity association structure, determining the continuity of the trajectory data set in the synchronization timestamp sequence, and obtaining the complete trajectory path; normalizing the feature data affected by light changes in the complete trajectory path based on an environmental compensation method to obtain compensated feature data; if the matching degree between the compensated feature data and the data of adjacent monitoring points meets a condition, fusing the angle difference information in the complete trajectory path to obtain the optimized identity recognition result; and generating a dynamic management structure for personnel flow in a full-domain monitoring network based on the optimized identity recognition result, detecting related trajectory anomalies, and obtaining response support information.

[0008] In one aspect of this disclosure, the step of adjusting the clock deviation between target devices using a time synchronization method based on the original video frame sequence acquired by the multi-sensor device to obtain a synchronized timestamp sequence includes:

[0009] The initial frame sequence data is obtained by collecting raw video data from multiple sensor devices.

[0010] The clock deviation between the target devices is detected to determine whether there is a time offset. If an offset is detected, the initial frame sequence data is time-calibrated by calculating the offset to obtain a calibrated frame sequence.

[0011] Based on the calibrated frame sequence, the timestamp information of the video frames acquired by each device is extracted to construct a preliminary timestamp sequence;

[0012] The initial timestamp sequence is checked for standard deviation. If outliers are found in the timestamp sequence, the outliers are corrected using a linear interpolation method to obtain the corrected timestamp sequence.

[0013] Based on the corrected timestamp sequence, data alignment processing is performed to obtain an aligned timestamp sequence;

[0014] Based on the aligned timestamp sequence, Kalman filtering is performed on the timestamp sequence to smooth it and obtain the final synchronization timestamp sequence.

[0015] By associating and mapping the final synchronization timestamp sequence with the calibrated frame sequence, a synchronized video data stream is constructed, thus completing the data integration of multiple sensor devices.

[0016] In one aspect of this disclosure, the step of extracting personnel target boundary information from video frames at each monitoring point based on the synchronization timestamp sequence, generating multi-dimensional feature data using a target detection method, and verifying the cross-frame feature consistency includes:

[0017] By obtaining video frame data corresponding to the synchronization timestamp from the video of the monitoring point, and using a preset image processing tool to perform preliminary segmentation of the video frame data, the initial boundary information of the personnel target is obtained.

[0018] Based on the initial boundary information, target detection technology is used to accurately locate personnel targets and generate multi-dimensional feature information including position and shape;

[0019] For the multidimensional feature information, feature change data across frames is obtained, and by comparing the feature changes of adjacent video frames, it is determined whether there is a deviation in feature consistency.

[0020] If there is a deviation in feature consistency, the degree of deviation is evaluated by a preset threshold to obtain a quantitative result of feature consistency.

[0021] Based on the quantization results, the corresponding timestamp data in the time series is obtained. By comparing the timestamp data with the feature consistency results, the continuity status of cross-frame features is obtained.

[0022] Based on the continuity state, a convolutional neural network is used to perform in-depth analysis of feature consistency to obtain the final cross-frame feature consistency judgment result.

[0023] Based on the final judgment result, other relevant frame data from the video data source are obtained, and the integrity assessment of feature consistency is obtained by comparing and supplementing the information.

[0024] In one aspect of this disclosure, the step of associating corresponding personnel targets with the same identity and obtaining the preliminary identity association structure if the similarity between the feature data meets a preset condition includes:

[0025] By obtaining feature data from multiple sources, comparison tools are used to perform preliminary cleaning and format unification of the feature data, thereby obtaining a processed feature set.

[0026] If the similarity between the data in the processed feature set reaches a preset standard, the comparison mechanism is triggered to obtain the preliminary matching personnel target;

[0027] Based on the initially matched personnel targets, cluster analysis is used to group the targets and determine whether there are any potential identity associations.

[0028] If there is an identity association between the grouped targets, then the grouping results are analyzed in depth using association rule mining tools to obtain the preliminary structure of the identity association.

[0029] For the identity associations in the preliminary structure, the association strength data is obtained. If the association strength data is lower than a preset threshold, weak association targets are removed to obtain an optimized association structure.

[0030] By performing multi-dimensional verification on the optimized association structure and using preset verification rules to perform secondary confirmation on the targets within the structure, the final identity association result is obtained.

[0031] If there are still unconfirmed targets in the final identity association results, additional feature data is obtained by supplementing the data source, and the comparison and association process is executed repeatedly to determine the final identity consistency.

[0032] In one aspect of this disclosure, the step of obtaining a set of personnel trajectory data during cross-regional movement based on the preliminary identity association structure, determining the continuity of the trajectory data set in the synchronization timestamp sequence, and obtaining the complete trajectory path includes:

[0033] By establishing a preliminary identity association structure, the raw data set of personnel trajectories is obtained from data sources of cross-regional movement, and preliminary data processing is completed to obtain a structured trajectory dataset.

[0034] Based on the structured trajectory dataset, time alignment processing is performed on the synchronization time and timestamp sequence, and a time window partitioning method is used to obtain the continuous distribution of data in the timestamp sequence;

[0035] If the data continuity distribution in the timestamp sequence is interrupted, interpolation is performed using historical trajectory data to complete the sequence, and the completed time series dataset is obtained. It is then determined whether its continuity meets the preset threshold requirements.

[0036] By using the completed time series dataset, we analyze the changing patterns of people's trajectories during cross-regional movement, and use trajectory smoothing to obtain smoothed trajectory data records.

[0037] Based on the smoothed trajectory data records and combined with the continuity analysis results, a complete trajectory path is constructed, and key nodes and turning information in the path are obtained.

[0038] Obtain key node data in the complete trajectory path, analyze the movement between nodes, classify the path segments, and obtain a feature set of the classified path segments.

[0039] By combining the classified path segment feature set with identity association information, the cross-regional movement pattern of personnel trajectories is analyzed, and the final trajectory path structure model is constructed.

[0040] In one aspect of this disclosure, the step of normalizing the feature data affected by changes in light along the complete trajectory path based on an environmental compensation method to obtain compensated feature data includes:

[0041] Feature data affected by changes in light are extracted from the trajectory path, and the feature data are initially corrected using a preset environmental compensation model to obtain corrected initial data.

[0042] Based on the corrected initial data, a normalization method is used to standardize and adjust the data to obtain a standardized feature dataset.

[0043] Based on the standardized feature dataset, the specific impact range of light interference on trajectory features is analyzed to obtain key data points affected by the environment.

[0044] Based on the key data points affected by the environment, a secondary compensation process is performed using data correction technology to obtain optimized trajectory feature data.

[0045] By performing path analysis on the optimized trajectory feature data, it is determined whether there are any abnormal deviations in the trajectory path. If a deviation is detected, the relevant data points are marked, and the marked path data is obtained.

[0046] Based on the marked path data, and taking into account factors such as environmental influences and changes in light, a support vector machine algorithm is used to classify the trajectory path to obtain the final compensation data result.

[0047] The final compensation data results are verified to determine whether the data processing meets the preset threshold requirements. If not, the data is corrected again to obtain a trajectory feature dataset that meets the requirements.

[0048] In one aspect of this disclosure, the step of fusing the angle difference information in the complete trajectory path to obtain the optimized identity recognition result if the matching degree between the compensated feature data and the data of neighboring monitoring points meets the condition includes:

[0049] Based on the original feature data obtained from the monitoring equipment, the data is initially cleaned and format unified to obtain the processed initial feature dataset;

[0050] If the difference between the initial feature dataset and the pre-established reference dataset exceeds a preset range, the initial feature dataset will be compensated and adjusted to obtain the compensated feature dataset.

[0051] The compensated feature dataset is compared and analyzed with the data collected from neighboring monitoring points. If the comparison result meets the preset matching degree condition, then the dataset has the basic conditions for fusion processing.

[0052] For the compensated feature dataset that meets the conditions, obtain the relevant path information in the complete trajectory path, extract the angle difference data, and obtain the angle difference information set to be fused.

[0053] The angle difference information set and the compensated feature dataset are integrated, and the integrated data are classified using a support vector machine model to obtain preliminary identity recognition results.

[0054] If the initial identity recognition result does not match the historical data to the preset threshold, the integrated data will be calibrated a second time to generate an optimized identity recognition result.

[0055] By performing a final verification between the optimized identity recognition results and supplementary data from nearby monitoring points, it is determined whether the results meet the preset accuracy standards, and a final confirmed identity recognition conclusion is obtained.

[0056] In one aspect of this disclosure, the step of generating a dynamic management structure for personnel flow in a comprehensive monitoring network based on the optimized identity recognition result, detecting related trajectory anomalies, and obtaining response support information includes:

[0057] By constructing a basic information database within the entire monitoring network using identity recognition data, the initial records of personnel movement are obtained from the basic information database to obtain the correspondence between identity and location;

[0058] Based on the aforementioned correspondence, the random forest algorithm is used to analyze the real-time changes in personnel flow and to obtain the distribution of flow trends from the random forest algorithm.

[0059] If the distribution of the flow trend exceeds the preset threshold range, the trajectory anomaly detection mechanism is triggered. The trajectory anomaly detection mechanism performs a deep comparison of the abnormal data to determine whether there is a potential risk.

[0060] Based on the assessment results of the potential risks, and combined with the resource distribution within the entire monitoring scope, priority targets are obtained from the resource distribution.

[0061] Based on the priority objects, abnormal trajectories are tracked in real time to obtain updated location information;

[0062] If the updated location information continues to show abnormalities, then integrate multi-source data within the monitoring range from the information acquisition interface, and determine the direction of subsequent processing based on the multi-source data;

[0063] Based on the aforementioned subsequent processing direction, the flow trend is optimized to obtain more accurate monitoring feedback.

[0064] In one aspect of this disclosure, in the step of performing Kalman filtering smoothing on the timestamp sequence based on the aligned timestamp sequence to obtain the final synchronization timestamp sequence, the OpenCV tool is used to perform Kalman filtering smoothing on the timestamp sequence.

[0065] In one aspect of this disclosure, in the step of obtaining key node data in the complete trajectory path, analyzing movement between nodes, classifying the path segments, and obtaining a feature set of the classified path segments, the path segments are classified using the K-means clustering algorithm.

[0066] Compared with the prior art, the present invention has the following beneficial effects:

[0067] This invention adjusts clock deviations between devices using a time synchronization method to generate a synchronized timestamp sequence, extracts personnel target boundary information and generates multi-dimensional feature data, performs preliminary identity association based on feature consistency and similarity conditions, further normalizes features affected by light through trajectory data continuity judgment and environmental compensation methods, integrates angle difference information to optimize identity recognition results, and finally generates a dynamic management structure and detects trajectory anomalies, providing response support information. This invention significantly improves the accuracy of cross-regional personnel trajectory tracking and the reliability of identity recognition, providing efficient technical support for full-domain monitoring. Attached Figure Description

[0068] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0069] Figure 1 This is a flowchart of the campus security monitoring method in this invention. Detailed Implementation

[0070] The present invention will be further described below with reference to embodiments. These embodiments are merely some, not all, of the embodiments described. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the protection scope of the present invention.

[0071] Please see Figure 1 As shown in the figure, this embodiment discloses a campus security monitoring method based on image recognition technology, which may specifically include:

[0072] Step 1: Acquire raw video data generated by the multi-sensor devices and perform preliminary data processing to obtain an initial video frame sequence. In one embodiment, the multi-sensor devices may include high-definition cameras, infrared sensors, and other video acquisition devices distributed at different monitoring points. These devices are deployed in a specific area to record personnel activities in real time. The acquisition of raw video data can be accomplished through the built-in storage modules of each device, and the data format can be a common video encoding format.

[0073] During the acquisition process, the device records the generation time of each frame, forming the initial timestamp information.

[0074] It should be noted that the preliminary processing mainly includes decoding and framing the video data, breaking down the continuous video stream into independent frame sequences for subsequent time synchronization and feature extraction.

[0075] Step 11: For the initial video frame sequence, check for time discrepancies between devices and record preliminary timestamp data. In one possible implementation, since multi-sensor devices may be manufactured by different companies or deployed in different network environments, there may be slight differences in the clocks between devices, resulting in inconsistent timestamps for the acquired video frames. To solve this problem, the time of each device can be calibrated using a network time protocol to detect any time offsets.

[0076] For example, in a monitored area, suppose there are three cameras deployed at the entrance, corridor, and exit. The timestamps of the initially collected video frames show that the time of the entrance camera is 2 seconds faster than that of the corridor camera. At this time, this offset needs to be recorded as the basis for subsequent calibration. Such a detection process can effectively avoid the problem of frame sequence misalignment caused by time deviation.

[0077] Step 12: Based on the detected time deviation, perform time calibration on the initial video frame sequence to generate a calibrated frame sequence. Specifically, time calibration can be achieved by calculating the offset between each device, adjusting the frame sequences of all devices to the same time reference.

[0078] For example, in the case of the entrance camera's time being 2 seconds fast, the timestamps of all frames from the entrance camera can be uniformly reduced by 2 seconds to align with the time reference of the corridor and exit cameras. In the calibrated frame sequence, each frame has a unified time reference point, laying the foundation for subsequent synchronization processing.

[0079] It should be noted that the calibration process can also be combined with timestamp records in the equipment log to further verify the accuracy of the calibration results.

[0080] Step 2: Construct a synchronization timestamp sequence using the calibrated frame sequence and complete the data integration of multiple sensor devices.

[0081] In one embodiment, the construction of the synchronization timestamp sequence requires extracting the timestamp information of each frame from the calibrated frame sequence to form a preliminary timestamp sequence. For this sequence, statistical analysis methods can be used to check for outliers. For example, if the timestamp of a certain frame differs too much from the timestamps of other frames, it may be marked as an anomaly. The correction of outliers can be accomplished by linear interpolation, that is, by estimating the reasonable timestamp of the abnormal frame based on the timestamp values ​​of the preceding and following frames, thereby obtaining the corrected timestamp sequence.

[0082] Step 21: Align the corrected timestamp sequence to determine the aligned timestamp sequence. Specifically, the alignment process can be achieved using a dynamic time warping tool, which can analyze the distribution of frame sequences acquired by different devices on the time axis and adjust the small time differences between the sequences.

[0083] For example, in a large shopping mall surveillance scenario, the frame sequences of the entrance and exit cameras may differ by 0.5 seconds due to network latency. By using a dynamic time warping tool, the timestamp sequences of the two cameras can be adjusted to the same timeline, ensuring that subsequent feature extraction will not result in data matching errors due to time misalignment.

[0084] It should be noted that the alignment process can be further optimized by taking into account the physical distance between devices and the signal transmission delay.

[0085] Step 22: Smooth the aligned timestamp sequence to generate the final synchronized timestamp sequence. In one possible implementation, smoothing can be achieved through filtering methods to eliminate noise or minor fluctuations that may exist in the timestamp sequence.

[0086] For example, within a monitored area, occasional interruptions in signal transmission between devices may cause slight jumps in some timestamps. Filtering can smooth out and make the timestamp sequence more continuous. A smoothed timestamp sequence provides a reliable time reference for constructing subsequent video data streams. This processing method helps improve the accuracy of data integration and avoids frame matching problems caused by discontinuous timestamps.

[0087] Step 23 involves associating the final synchronization timestamp sequence with the calibrated frame sequence to construct a synchronized video data stream. In one embodiment, the association mapping process involves binding each frame with its corresponding synchronization timestamp to form a unified time-frame correspondence.

[0088] For example, in an office building monitoring scenario, the synchronized timestamp sequence shows that three cameras simultaneously recorded video frames at a certain point in time. Through association mapping, the frame data from these three cameras can be integrated into a unified video data stream, completing the data integration of multiple sensor devices.

[0089] It should be noted that the integrated data stream can provide a consistent data foundation for subsequent personnel target detection, avoiding analysis errors caused by inconsistencies in data between devices.

[0090] Step 3: Extract personnel target boundary information from the synchronized video data stream, generate multi-dimensional feature data, and determine cross-frame feature consistency.

[0091] In one possible implementation, the extraction of personnel targets first requires preliminary segmentation from the video frame data corresponding to the synchronization timestamp to obtain the initial boundary information of the personnel targets. Preliminary segmentation can be achieved using image processing tools;

[0092] For example, by using background modeling methods to distinguish people in the foreground from the background area, the initial boundary information after segmentation usually includes the approximate outline and position of the people in the video frame, which provides a basis for subsequent accurate positioning.

[0093] Step 31: Based on the initial boundary information, accurately locate the personnel target and generate multi-dimensional feature information containing position and shape.

[0094] Specifically, precise positioning can be achieved through target detection technology, which can further optimize the initial boundary and extract the precise location and shape features of personnel targets.

[0095] For example, in a subway station monitoring scenario, the initial boundary information may only identify the general area of ​​the target person, but through target detection technology, the outline of specific parts of the target person, such as the head and limbs, can be further determined to form multi-dimensional feature information that includes position coordinates and shape description.

[0096] It should be noted that multidimensional feature information can also include auxiliary features such as color and texture, which can be used to enhance the discriminative power of targets.

[0097] Step 32: Analyze the feature changes across frames for multi-dimensional feature information and determine whether there is a deviation in feature consistency.

[0098] In one embodiment, the analysis of cross-frame feature changes can be achieved by comparing feature data of the same target in adjacent video frames.

[0099] For example, if the coordinates of a person's position change drastically within a short period of time in consecutive video frames, exceeding a reasonable range, it may be judged as a deviation in feature consistency. This analysis process can help identify feature extraction errors caused by factors such as changes in lighting or occlusion, providing a basis for subsequent deviation correction. It should be noted that feature consistency assessment can also incorporate timestamp data to ensure that the analysis results maintain consistency with the continuity of the time series.

[0100] Step 33: If there is a deviation in feature consistency, the degree of deviation is evaluated by a preset threshold to obtain a quantitative result of feature consistency.

[0101] Specifically, the degree of deviation can be assessed by comparing the value of the feature change with a preset threshold.

[0102] For example, in a campus surveillance scenario, if the positional offset of a target person in adjacent frames exceeds a preset threshold range, the degree of deviation can be quantified as high, indicating poor feature consistency. The quantification results can provide data support for subsequent feature correction, avoiding target recognition errors caused by accumulated deviations. It should be noted that the threshold setting can be adjusted according to the complexity of the surveillance scenario; for example, in densely populated areas, the threshold can be appropriately relaxed.

[0103] Step 34: Based on the quantization results and timestamp data, determine the continuity status of cross-frame features.

[0104] In one possible implementation, determining the continuity state requires comparing the quantization results with the synchronous timestamp sequence to analyze whether the characteristic changes are consistent with the continuity of the time series.

[0105] For example, if the quantization results show that the characteristics of a certain person target deviate continuously over a certain period of time, while the timestamp sequence shows that the video frames are continuous and uninterrupted during that period of time, it can be determined that the cross-frame feature continuity of the target is poor and may require further correction.

[0106] This analytical approach can effectively identify inconsistencies in features caused by data processing or environmental interference.

[0107] Step 35: Based on the continuous state, a deep analysis method is used to further judge the feature consistency and obtain the final cross-frame feature consistency result.

[0108] Specifically, in-depth analysis can be achieved by constructing time series models of feature changes to analyze the patterns of feature changes over long periods of time.

[0109] For example, in a commercial street surveillance scenario, if the characteristics of a certain person repeatedly deviate over multiple time periods, but the overall trend of change conforms to a reasonable movement pattern, it can be determined that the consistency of its cross-frame features is still acceptable. The final judgment result can provide a reliable basis for subsequent identity association and avoid the impact of short-term deviations on the overall recognition effect.

[0110] Step 36: Based on the final judgment result, obtain other relevant frame data from the video data source and perform a complete assessment of feature consistency.

[0111] In one embodiment, integrity assessment can be accomplished by comparing supplementary information, such as extracting relevant data within the same time period from video frames of other monitoring points to verify whether the target features are consistent.

[0112] For example, if the consistency of a person's characteristics at the main monitoring point is judged to be poor, but supplementary data from nearby monitoring points shows that the changes in their characteristics follow a reasonable pattern, then the consistency assessment result can be corrected to an acceptable state.

[0113] This evaluation method can improve the comprehensiveness of feature consistency judgment and reduce the possibility of misjudgment.

[0114] In one possible implementation, the process of accurately locating personnel targets in step 31 can be further refined into multiple sub-steps to adapt to the needs of different monitoring scenarios.

[0115] For example, in a surveillance scenario in an airport security checkpoint, the accurate location of personnel targets needs to take into account characteristics such as dense crowds and rapid movement.

[0116] Step 311: First, noise filtering is performed on the initial boundary information to eliminate the boundary blurring problem caused by background interference.

[0117] Step 312: Based on the filtered boundary information, a region segmentation method is used to further separate the personnel target from the background and obtain clearer contour data.

[0118] Step 313: Based on the separated contour data, the parameters of the positioning algorithm are adjusted in combination with the personnel's movement speed and direction information to ensure that the target position can still be accurately captured under rapid movement. Such refinement can significantly improve the positioning accuracy in complex scenes and lay a solid foundation for subsequent feature extraction.

[0119] In another embodiment, the analysis of cross-frame feature changes in step 32 can be explored in depth from different perspectives.

[0120] For example, in a surveillance scenario in a train station waiting hall, the features of people may change significantly due to changes in lighting or obstruction by crowds.

[0121] Step 321: First, extract the multidimensional feature data of the personnel target in each frame, including position, shape and color information.

[0122] Step 322: For the feature data of adjacent frames, calculate the position offset and shape change rate to determine whether there are abnormal fluctuations.

[0123] Step 323: If abnormal fluctuations are detected, analyze whether the fluctuations are consistent with the movement patterns of people by combining the timestamp data. For example, running fast may cause a large positional shift, but it is within a reasonable range of change.

[0124] Step 324: Based on the comprehensive analysis results, a preliminary report on cross-frame feature changes is generated to provide data support for subsequent consistency judgment. This multi-angle analysis can effectively address the feature change problem in complex environments and ensure the reliability of the judgment results.

[0125] In one possible implementation, the evaluation of the degree of feature consistency deviation in step 33 can be optimized by combining the parameters with the specific scenario.

[0126] For example, in a nighttime parking lot monitoring scenario, insufficient light may lead to significant deviations in feature extraction, requiring a more lenient selection of preset thresholds.

[0127] Step 331: First, adjust the threshold range of deviation assessment dynamically according to the lighting conditions of the monitoring environment. For example, appropriately increase the position offset threshold when the light is dim.

[0128] Step 332: For the adjusted threshold, calculate the characteristic deviation value of each person's target and generate quantitative data on the degree of deviation.

[0129] Step 333: Combine quantitative data to analyze whether the deviation is concentrated in a specific time period or a specific monitoring point. For example, if the deviation value is generally higher in a certain area at night, it may be necessary to further optimize the camera settings. Such parameter optimization can make the deviation assessment more in line with the actual scenario and improve the practicality of the assessment results.

[0130] In another embodiment, multi-dimensional data can be introduced for auxiliary analysis to determine the cross-frame feature continuity state in step 34.

[0131] For example, in a large-scale exhibition monitoring scenario, the movement paths of personnel are complex and easily affected by occlusion. Relying solely on feature change data may not be sufficient to accurately determine the continuity status.

[0132] Step 341: First, extract the time period data of the target occurrence from the synchronization timestamp sequence and analyze whether its time distribution is continuous.

[0133] Step 342: Combine the time distribution data to check the change pattern of the target features in each time period. For example, if the features of a certain target change significantly in a certain time period, but then return to normal, it may be due to temporary occlusion.

[0134] Step 343: By integrating time distribution and characteristic change data, a comprehensive evaluation result of the continuous state is generated to guide subsequent in-depth analysis. This multi-dimensional analysis method can effectively reduce misjudgments caused by insufficient data sources.

[0135] In one possible implementation, the in-depth analysis of feature consistency in step 35 can be refined by constructing a time series model.

[0136] For example, in a city square surveillance scenario, the movement trajectory of a person may be affected by a variety of factors, and in-depth analysis requires comprehensive consideration of multiple variables.

[0137] Step 351: First, construct a time series model of the target features, arranging the feature data such as location and shape in chronological order to form feature change curves.

[0138] Step 352: Analyze the fluctuation frequency and amplitude of the characteristic change curve to determine whether there are periodic or sudden changes. For example, periodic changes may be related to people repeatedly passing through a certain area.

[0139] Step 353: Based on the fluctuation analysis results, adjust the consistency judgment criteria, for example, appropriately relax the consistency requirements for targets with large periodic changes.

[0140] Step 354 generates the final deep analysis report, providing comprehensive support for cross-frame feature consistency results. This deep analysis method can significantly improve the accuracy of feature consistency judgment in complex scenarios.

[0141] In another embodiment, the supplementary data comparison for the feature consistency integrity assessment in step 36 can be optimized from the perspective of data integration from different monitoring points.

[0142] For example, in a shopping mall surveillance scenario, a person might have poor feature consistency assessment at the main monitoring point due to obstruction, but a neighboring monitoring point might record clearer data.

[0143] Step 361: First, filter the time period data related to the target from the synchronization timestamp sequence to determine the range of frames that need to be supplemented for comparison.

[0144] Step 362: For the selected frame range, extract the feature information of the corresponding time period from the video data stream of the nearby monitoring points.

[0145] Step 363: Combine the feature information of the main monitoring point and neighboring monitoring points to analyze whether the target features are consistent across multiple data points, such as whether the location trajectory is continuous and whether the shape features are similar.

[0146] Step 364 involves synthesizing the results of multi-point data analysis and updating the integrity assessment of feature consistency to ensure a more comprehensive and reliable assessment result. This multi-point data integration method effectively compensates for the deficiencies of data from a single monitoring point, improving the accuracy of the overall assessment.

[0147] It should be noted that the implementation of each of the above steps and embodiments is centered on the scenario of personnel trajectory recognition and monitoring, and revolves around the data synchronization and feature extraction of multi-sensor devices.

[0148] To cater to the characteristics of different monitoring scenarios, such as population density, lighting conditions, and movement speed, the parameters and processing methods of each step can be flexibly adjusted. For example, in densely populated scenarios, the redundancy of feature extraction can be increased, and the target discrimination capability can be improved through multi-dimensional feature data; in scenarios with poor lighting, the image processing algorithm can be optimized to enhance the robustness of feature extraction. This flexible adjustment method enables the technical solution of this invention to adapt to various practical application needs.

[0149] In one possible implementation, the process of constructing the synchronized timestamp sequence in step 2 can be further optimized in combination with specific scenarios.

[0150] For example, in a port terminal monitoring scenario, multi-sensor devices may be affected by environmental factors such as sea breeze and tides, resulting in unstable timestamp data collection.

[0151] Step 221: First, environmental noise is filtered from the preliminary timestamp sequence to eliminate abnormal timestamp values ​​caused by equipment jitter or signal interference.

[0152] Step 222: For the filtered timestamp sequence, adjust the parameters of the dynamic time warping tool based on the physical distance between devices and signal transmission delay to ensure alignment accuracy.

[0153] Step 223: After alignment, the timestamp sequence is further optimized using a smoothing method, such as eliminating small fluctuations by using a sliding window averaging method.

[0154] Step 224 generates the final synchronization timestamp sequence and associates it with the calibrated frame sequence to form a unified video data stream. This optimization method can effectively address the time synchronization challenges in complex environments and ensure the reliability of data integration.

[0155] In another embodiment, the extraction of personnel target boundary information in step 3 can be explored in depth from the perspective of image processing.

[0156] For example, in a factory workshop monitoring scenario, personnel targets may be mixed with the background of mechanical equipment, making boundary extraction difficult.

[0157] Step 311: First, the video frame data is preprocessed to enhance image contrast and highlight the difference between the person target and the background.

[0158] Step 312: Based on the preprocessed image data, the foreground target is separated using a background modeling method to generate initial boundary information.

[0159] Step 313: Based on the initial boundary information, and combined with the motion characteristics of the human target, such as gait frequency and direction of movement, further optimize the boundary contour to avoid misjudging background noise as the target boundary.

[0160] Step 314 generates optimized boundary information to provide reliable data support for subsequent precise positioning. This image processing optimization method can significantly improve the accuracy of boundary extraction, especially in scenes with complex backgrounds.

[0161] In one possible implementation, the parameters for the acquisition and preliminary processing of the raw video data in step 1 can be adjusted based on the device performance.

[0162] For example, in a highway tollbooth monitoring scenario, multi-sensor equipment needs to operate continuously for extended periods, resulting in a massive amount of data collection. Step 111: First, adjust the sampling frequency of the video data based on the equipment's storage capacity and processing speed, for example, increasing the sampling rate during peak hours to capture more details.

[0163] Step 112: Perform preliminary compression processing on the collected raw video data to reduce storage pressure while retaining key frame information.

[0164] Step 113: After compression, the video data is decomposed into a frame sequence, and the initial timestamp data of each frame is recorded.

[0165] Step 114 generates an initial video frame sequence, providing a data foundation for subsequent time calibration. This parameter adjustment method effectively balances device performance and data processing requirements, ensuring the stability of the acquisition process.

[0166] In another embodiment, the alignment process of the timestamp sequence in step 21 can be refined from the perspective of data consistency verification.

[0167] For example, in a stadium monitoring scenario, multi-sensor devices have a wide coverage area, and the data transmission latency between devices varies greatly.

[0168] Step 211: First, perform a consistency check on the preliminary timestamp sequence and analyze whether the timestamp distribution of each device conforms to the expected pattern.

[0169] Step 212: Based on the consistency check results, identify devices that may have significant delays, such as a device whose timestamp is always lagging behind other devices.

[0170] Step 213: Based on the recognition results, adjust the alignment parameters of the dynamic time warping tool, for example, by increasing the time compensation value for devices with large delays.

[0171] Step 214: After completing the alignment process, generate the aligned timestamp sequence and record the parameter adjustment logs during the alignment process for subsequent verification. This verification and adjustment method can effectively improve the accuracy of timestamp sequence alignment and ensure the interoperability of data from multiple devices.

[0172] It should be noted that the detailed process of the above embodiments and steps revolves around the application of multi-sensor devices in monitoring scenarios, aiming to provide reliable support for personnel trajectory recognition and identity association through data synchronization and feature extraction. For the specific needs of different scenarios, such as the number of devices, coverage area, and environmental interference, the implementation of each step can be further optimized.

[0173] For example, in scenarios with a large number of devices, a distributed processing mechanism can be introduced to improve the efficiency of data synchronization; in scenarios with significant environmental interference, a data preprocessing step can be added to enhance data robustness. This flexibility enables the technical solution of this invention to play a role in a variety of practical applications.

[0174] In one possible implementation, the time calibration process in step 12 can be verified from multiple dimensions based on the specific scenario. For example, in a hospital lobby monitoring scenario, the time deviation of multi-sensor devices may fluctuate significantly due to network congestion.

[0175] Step 121: First, record the initial time offset between each device and generate an offset distribution map to analyze the deviation pattern.

[0176] Step 122: Based on the offset distribution map, identify devices that may be abnormal, such as a device whose time offset differs significantly from other devices.

[0177] Step 123: Based on the anomaly identification results, perform preliminary calibration on the initial frame sequence and adjust the timestamp data of the abnormal device.

[0178] Step 124: After completing the initial calibration, use cross-validation to check the calibration results. For example, compare the timestamp sequences of adjacent devices to confirm whether the calibrated data is consistent.

[0179] Step 125 generates the final calibrated frame sequence and records the anomaly handling log during the calibration process for subsequent optimization. This multi-dimensional verification method can effectively improve the reliability of time calibration and avoid data misalignment problems caused by the inadequacy of a single calibration method.

[0180] In another embodiment, the smoothing of the timestamp sequence in step 22 can be optimized from the perspective of noise sources.

[0181] For example, in an industrial park monitoring scenario, multi-sensor devices may be affected by mechanical vibration or electromagnetic interference, resulting in large fluctuations in timestamp data.

[0182] Step 221: First, analyze the fluctuation characteristics of the timestamp sequence to identify whether the fluctuation is related to a specific device or time period.

[0183] Step 222: Adjust the smoothing parameters according to the fluctuation characteristics, for example, increase the smoothing window size for time periods with large fluctuations.

[0184] Step 223: After completing the parameter adjustment, use a smoothing method to process the timestamp sequence and eliminate small-scale noise.

[0185] Step 224 generates a smoothed timestamp sequence and verifies the smoothing effect by comparing it with the original sequence, for example, checking whether the smoothed sequence is more continuous. This noise source analysis and parameter optimization method can significantly improve the targeting of the smoothing process and ensure the quality of the synchronized timestamp sequence.

[0186] In one possible implementation, the association mapping between the synchronized timestamp sequence and the frame sequence in step 23 can be combined with a data integrity check. For example, in a tourist attraction monitoring scenario, multi-sensor devices cover a wide area and generate a large amount of data, and data omissions may occur during the association mapping process.

[0187] Step 231: First, perform an integrity check on the synchronized timestamp sequence to confirm whether there are any missing or duplicate timestamp data.

[0188] Step 232: Based on the inspection results, fix any potential data problems, such as by using interpolation to fill in missing timestamps.

[0189] Step 233: After completing the data repair, the synchronization timestamp sequence is matched one-to-one with the calibrated frame sequence to form an associated mapping relationship.

[0190] Step 234: Generate the synchronized video data stream and record the data repair log during the mapping process for subsequent verification. This integrity check and repair method can effectively avoid mapping errors caused by incomplete data and ensure the continuity of the video data stream.

[0191] In another embodiment, the extraction of personnel target boundary information in step 3 can be refined from the perspective of image quality optimization.

[0192] For example, in a monitoring scenario in an underground parking lot, the lighting conditions are poor, and the quality of video frame images may be low.

[0193] Step 301: First, perform image enhancement processing on the video frame data, such as adjusting brightness and contrast to highlight the outline features of the person target.

[0194] Step 302: For the enhanced image data, a background modeling method is used to separate the foreground target and generate initial boundary information.

[0195] Step 303: Adjust the background modeling parameters based on the lighting conditions, such as reducing the sensitivity to changes in lighting to avoid misjudgment.

[0196] Step 304 generates optimized initial boundary information, providing reliable support for subsequent precise positioning. This image quality optimization method can effectively address the challenges of boundary extraction in low-light environments and improve the accuracy of target recognition.

[0197] It should be noted that the core objective of all the above steps and embodiments is to improve the accuracy of data synchronization and feature extraction, aiming to provide a reliable data foundation for subsequent identity association and trajectory construction. The parameters and processing methods of each step can be flexibly adjusted to meet the specific needs of different monitoring scenarios. For example, in scenarios with large amounts of data, a parallel processing mechanism can be introduced to improve processing efficiency; in scenarios with poor image quality, an image preprocessing step can be added to enhance data quality. This flexibility enables the technical solution of this invention to play a role in various practical applications, providing strong support for personnel trajectory recognition and dynamic management.

[0198] This invention provides a detailed implementation of a method for personnel trajectory recognition and dynamic management based on multi-sensor devices. Starting from step 4, the specific implementation process of subsequent steps is further described to make the purpose, technical solution and advantages of this invention clearer.

[0199] The following describes the specific implementation methods of each step in the logical order of technical implementation, focusing on core aspects such as personnel identity association, trajectory construction and optimized recognition.

[0200] Step 4: If the similarity between feature data meets the preset conditions, the corresponding personnel targets are associated with the same identity to obtain the preliminary identity association structure.

[0201] In one possible implementation, the identity association process first requires acquiring feature data from multi-source data, including multi-dimensional feature information of personnel targets extracted from different monitoring points. This feature data can be pre-processed using tools to clean it, such as removing duplicate or inconsistently formatted data, ensuring the accuracy of subsequent comparisons. The cleaned feature set provides a unified data foundation for subsequent similarity calculations.

[0202] Step 41: Calculate the similarity of the processed feature set to determine whether the similarity between the data meets the preset standard.

[0203] Specifically, similarity calculation can be achieved by comparing the position, shape, and other feature data of each target in the feature set.

[0204] For example, in a large shopping mall surveillance scenario, if two individuals have similar coordinates and shape features across different monitoring points, their similarity can be preliminarily determined to meet a preset standard, triggering a subsequent comparison mechanism. It should be noted that the preset standard can be adjusted based on the complexity of the surveillance scenario; for instance, in densely populated areas, the similarity standard can be appropriately relaxed.

[0205] Step 42: Based on the similarity calculation results, determine the preliminary matched personnel targets and classify the targets using a grouping method.

[0206] In one embodiment, initially matched personnel targets can be filtered out through a comparison mechanism. For example, targets with similarity higher than a preset threshold are marked as potential matches. These potential matches can then be grouped using clustering analysis to determine if there is a potential identity association. For instance, if targets with similar characteristics appear in the data from multiple monitoring points, and these targets are continuous in time and space, they can be grouped into the same group and preliminarily identified as having the same identity. This grouping method can effectively reduce the complexity of subsequent analysis.

[0207] Step 43: If there is an identity association between the grouped targets, the grouping results are further confirmed by a deep analysis tool to obtain the preliminary structure of the identity association.

[0208] Specifically, deep analysis can be achieved by mining the association rules in the grouping results. For example, analyzing the temporal order and positional relationship of targets appearing at different monitoring points can confirm whether they conform to the movement patterns of the same identity. For instance, in an office building monitoring scenario, if a group of targets appears sequentially at the entrance, corridor, and exit monitoring points with reasonable time intervals, it can be confirmed that they share the same identity, generating a preliminary identity association structure.

[0209] It should be noted that in-depth analysis can also be verified using historical data to ensure the reliability of the correlation results.

[0210] Step 44: For the initial identity association structure, obtain association strength data, optimize the association results, and remove weakly associated targets. In one possible implementation, association strength data can be determined by calculating the stability of feature similarity between targets. For example, if the feature similarity of a target fluctuates significantly across multiple monitoring points, its association strength may be low. Targets with association strength below a preset threshold can be removed from the association structure, generating an optimized association structure. For instance, in a subway station monitoring scenario, if a target is similar in features to other targets only at one monitoring point, but differs significantly in features at other monitoring points, it can be identified as a weakly associated target and removed. This optimization method can improve the accuracy of the identity association structure.

[0211] Step 45 involves a secondary verification of the optimized association structure using a multi-dimensional verification method to obtain the final identity association result. Specifically, multi-dimensional verification can be achieved by comprehensively analyzing data from multiple perspectives, including time, space, and features. For example, if an identity in the optimized association structure is continuous in time, reasonable in space, and has stable feature data, it can be confirmed as the final identity association result. It should be noted that if anomalies are found during the verification process, such as unreasonable spatial jumps in an identity, these can be marked as targets to be confirmed, providing a basis for subsequent supplementary data processing.

[0212] Step 46: If unconfirmed targets still exist in the final identity association results, additional feature data is obtained through supplementary data sources, and the comparison and association process is repeated. In one embodiment, the supplementary data source may include video data or historical data from other monitoring points. By extracting additional feature data, unconfirmed targets can be compared again. For example, in a campus monitoring scenario, if the identity association result of a target at the main monitoring point is unclear, its feature information can be extracted from the supplementary data of neighboring monitoring points, and similarity calculation and association analysis can be repeated until its identity consistency is confirmed. This cyclical processing method can effectively reduce the number of unconfirmed targets and improve the completeness of identity association.

[0213] Step 5: Through the preliminary identity association structure, obtain the set of personnel trajectory data when moving across regions, determine its continuity in the synchronous timestamp sequence, and construct a complete trajectory path.

[0214] In one possible implementation, acquiring trajectory data first requires extracting a raw dataset from data sources involving cross-regional movement. This dataset includes personnel movement information recorded at different monitoring points. The raw dataset can then be preliminarily organized, for example, by arranging the records from each monitoring point in chronological order, forming a structured trajectory dataset that provides a foundation for subsequent analysis.

[0215] Step 51: Based on the structured trajectory dataset, perform time alignment processing to determine the data continuity distribution in the timestamp sequence.

[0216] Specifically, time alignment can be achieved by matching trajectory data with a synchronization timestamp sequence, using a time window division method to analyze the distribution of data on the time axis.

[0217] For example, in a city park surveillance scenario, if the trajectory data of a certain individual is recorded completely within one time window but missing in another, it can be determined that there is an interruption in its continuous distribution. It should be noted that the size of the time window can be adjusted according to the scope of the monitored area and the speed of personnel movement.

[0218] Step 52: If the continuity of data distribution in the timestamp sequence is interrupted, interpolation is performed using historical trajectory data to complete the time series dataset. In one embodiment, interpolation completion can be achieved by analyzing the movement patterns of historical trajectory data, for example, estimating the possible positions within the interrupted time period based on the preceding and following positions of the personnel target, thus completing the missing data.

[0219] For example, in a commercial street surveillance scenario, if the trajectory data of a certain individual is missing within a certain time period due to a surveillance blind spot, interpolation can be performed based on the direction and speed of movement in the preceding and following time periods to generate a complete dataset. This completion method can effectively improve the continuity of trajectory data.

[0220] Step 53: Analyze the trajectory change patterns of people moving across regions using the completed time series dataset, perform trajectory smoothing processing, and obtain smoothed trajectory data records.

[0221] Specifically, trajectory smoothing can be achieved through filtering methods to eliminate small fluctuations or noise in trajectory data.

[0222] For example, if the trajectory data of a certain identity shows unreasonable position jumps in a short period of time, its trajectory curve can be adjusted through smoothing to make it more consistent with the actual movement pattern.

[0223] It should be noted that the smoothing process can also be optimized by combining information on the speed and direction of personnel movement to ensure the rationality of the processing results.

[0224] Step 54: Based on the smoothed trajectory data records and the results of continuity analysis, construct a complete trajectory path and determine the key nodes and turning information in the path.

[0225] In one possible implementation, the complete trajectory path can be constructed by connecting trajectory data points from different time periods and extracting key nodes.

[0226] For example, points where personnel change direction of movement or enter new areas.

[0227] For example, in an airport monitoring scenario, if the trajectory of a certain person shows that they move from the entrance to the security check area and then turn to the boarding gate, the security check area and the boarding gate can be marked as key nodes. This construction method can clearly show the entire process of personnel movement.

[0228] Step 55: Obtain key node data in the complete trajectory path, classify path segments, analyze cross-regional movement patterns, and construct the final trajectory path structure model.

[0229] Specifically, path segment classification can be achieved by analyzing the movement characteristics between key nodes, such as classifying path segments into types like straight-line movement and turning movement, and then analyzing the overall pattern of cross-regional movement.

[0230] For example, in a train station monitoring scenario, if the trajectory of a certain identity shows that they frequently switch between the waiting area and the platform, it can be determined that their movement mode is a waiting mode, and a corresponding structural model can be constructed.

[0231] It should be noted that the structural model can also be optimized by combining identity association information to ensure that the trajectory path is consistent with the identity.

[0232] Step 6: The feature data affected by light variations in the complete trajectory path are normalized using an environmental compensation method to determine the compensated feature data. In one embodiment, environmental compensation first requires extracting feature data affected by light variations from the trajectory path. For example, feature data recorded in dimly lit or overly bright areas may contain biases. For these data, a pre-set environmental compensation model can be used for preliminary correction, adjusting the brightness or contrast of the feature data to obtain corrected initial data, providing a basis for subsequent processing.

[0233] Step 61: Standardize the corrected initial data to obtain a standardized feature dataset. Specifically, standardization can be achieved through normalization methods to adjust the numerical range of the feature data to a uniform interval, avoiding uneven data distribution caused by changes in lighting conditions.

[0234] For example, in a monitoring scenario in an underground parking lot, if the characteristic data of a certain identity is low due to insufficient lighting, its value range can be improved through standardization adjustments to maintain consistency with data from other areas.

[0235] It should be noted that the standardized processing parameters can also be adjusted based on the degree of change in light.

[0236] Step 62: Based on the standardized feature dataset, analyze the specific impact range of light interference on trajectory features and determine the key data points affected by the environment.

[0237] In one possible implementation, the analysis of the scope of influence can be achieved by comparing the variation patterns of feature data under different lighting conditions.

[0238] For example, if the trajectory characteristics of a certain identity deviate significantly after entering a darker area, the data points in that area can be marked as key data points affected by the environment. This analysis method can provide accurate targets for subsequent compensation processing.

[0239] Step 63: Perform secondary compensation processing on key data points affected by the environment to obtain optimized trajectory feature data.

[0240] Specifically, secondary compensation can be achieved through data correction techniques, such as adjusting the feature values ​​of key data points according to the pattern of light changes to make them consistent with the data under normal lighting conditions.

[0241] For example, in a nighttime street surveillance scenario, if the trajectory characteristics of a certain identity deviate in an area with insufficient street lighting, the stability of its feature values ​​can be improved through secondary compensation.

[0242] It should be noted that secondary compensation can also be verified by combining the feature values ​​of neighboring data points to ensure the rationality of the processing results.

[0243] Step 64: By performing path analysis on the optimized trajectory feature data, determine whether there are any abnormal deviations in the trajectory path, and mark the relevant data points.

[0244] In one embodiment, path analysis can be achieved by checking the continuity and rationality of the trajectory path. For example, if an identity's trajectory path shows unreasonable position jumps in a certain area, it can be judged as an abnormal deviation, and the relevant data points can be marked.

[0245] For example, in a factory workshop monitoring scenario, if the trajectory path of a certain identity shows that it suddenly jumps from one end to the other, the jump point can be marked as abnormal. This marking method can provide a clear target for subsequent correction.

[0246] Step 65: Based on the marked path data, combined with environmental influences and changes in light, perform classification processing to obtain the final compensation data results.

[0247] Specifically, classification processing can be achieved by analyzing the environmental characteristics of the labeled data points. For example, data points that are greatly affected by changes in light can be grouped into one category, and targeted compensation can be applied to this category of data points.

[0248] For example, in a port terminal monitoring scenario, if the trajectory path of a certain identity shows multiple abnormal deviations in the nighttime area, these data points can be classified and compensated to generate the final compensated data result.

[0249] It should be noted that the classification process can be further optimized by incorporating time period information to ensure the comprehensiveness of the compensation results.

[0250] Step 66: Verify the final compensation data results to determine whether the data processing meets the preset threshold requirements. If not, return for further correction.

[0251] In one possible implementation, the verification process can be achieved by comparing the feature distributions of the compensated data and the normal data. For example, if the difference between the compensated feature data distribution and the normal data is small, it can be judged as meeting the requirements; otherwise, it is necessary to return to the secondary compensation step for further correction.

[0252] For example, in a stadium monitoring scenario, if the compensated trajectory feature data still has a large deviation, the compensation parameters can be adjusted and the relevant data points can be reprocessed until they meet the requirements.

[0253] Step 7: If the matching degree between the compensated feature data and the data of the neighboring monitoring points meets the conditions, then the angle difference information in the complete trajectory path is fused to obtain the optimized identity recognition result.

[0254] In one embodiment, the matching degree determination first requires obtaining the original feature data from the monitoring device, and performing preliminary cleaning and format unification processing to obtain the processed initial feature dataset. These data can be compared with a pre-established reference dataset. If the difference exceeds the preset range, compensation adjustment is required to generate a compensated feature dataset.

[0255] Step 71: Based on the compensated feature dataset, compare and analyze it with the data collected from neighboring monitoring points to determine whether the matching degree conditions are met.

[0256] Specifically, comparative analysis can be achieved by calculating the similarity between feature datasets. For example, if the similarity between the compensated feature data and the data from neighboring monitoring points is higher than a preset threshold, it can be determined that the matching conditions are met. For instance, in a shopping mall monitoring scenario, if the feature data of a certain identity maintains a high similarity between the main monitoring point and neighboring monitoring points, it can be determined that it meets the basic conditions for fusion processing.

[0257] It should be noted that the matching conditions can be adjusted according to the distribution density of the monitoring points.

[0258] Step 72: For the compensated feature dataset that meets the conditions, extract the angle difference data in the complete trajectory path and perform data integration processing.

[0259] In one possible implementation, angular difference data can be extracted by analyzing the turning information of key nodes in the trajectory path, such as recording changes in the movement direction of a person between different monitoring points. This data can then be integrated with a compensated feature dataset to form a comprehensive dataset.

[0260] For example, in a subway station monitoring scenario, if the trajectory of a certain identity shows that they make multiple turns in the transfer passage, these turning angle data can be extracted and integrated with feature data to support subsequent classification processing.

[0261] Step 73: The integrated data is analyzed using a classification processing method to obtain preliminary identity recognition results, and then a secondary calibration process is performed.

[0262] Specifically, classification processing can be achieved by constructing a comprehensive model of feature and perspective differences, analyzing the correlation between data, and generating preliminary identity recognition results.

[0263] If the initial results do not match the historical data to the preset threshold, a second calibration is required, such as adjusting feature weights or supplementing data sources, until an optimized identity recognition result is generated.

[0264] For example, in an airport surveillance scenario, if the initial identification results show that a certain identity does not match the historical records, it can be calibrated by supplementing data from other monitoring points.

[0265] Step 74: Perform a final verification by comparing the optimized identity recognition results with supplementary data from nearby monitoring points to determine whether the results meet the preset accuracy standards.

[0266] In one embodiment, final verification can be achieved through cross-comparison of multi-point data. For example, if the optimized identity recognition result remains consistent across multiple monitoring points, it can be confirmed that it meets the accuracy standard, and a final confirmed identity recognition conclusion can be generated.

[0267] It should be noted that if there are still doubts about the verification results, the identity can be marked as pending confirmation to provide a basis for subsequent processing.

[0268] Step 8: Based on the optimized identity recognition results, generate a dynamic management structure for personnel flow in the full-domain monitoring network, detect related trajectory anomalies, and obtain response support information.

[0269] In one possible implementation, the generation of the dynamic management structure first requires building a basic information database within the global monitoring network to record identity recognition data and location information, thus forming a correspondence between identity and location.

[0270] Based on this data, we can analyze the real-time changes in population flow and obtain the distribution of flow trends.

[0271] Step 81: Based on the correspondence between identity and location, analyze the real-time changes in personnel flow and determine whether the flow trend exceeds a preset threshold range. Specifically, the flow trend analysis can be achieved by statistically analyzing the distribution density and movement speed of personnel in different areas;

[0272] For example, if the population density in a certain area suddenly increases and exceeds a preset threshold, it can be judged as an abnormal flow trend. This analysis method can provide data support for subsequent anomaly detection.

[0273] For example, in a large-scale exhibition monitoring scenario, if the flow of people in a certain exhibition area shows abnormal clustering, further detection mechanisms can be triggered.

[0274] Step 82: If the flow trend exceeds the threshold range, the trajectory anomaly detection mechanism is triggered to perform in-depth comparison of the abnormal data and determine whether there are potential risks.

[0275] In one embodiment, trajectory anomaly detection can be achieved by analyzing the trajectory paths of individuals within an abnormal area. For example, if the trajectory path of a certain identity shows that they repeatedly wander within an abnormal area, it can be determined that there is a potential risk. It should be noted that deep comparison can also be combined with historical data for verification, such as checking whether the identity has exhibited similar abnormal behavior in the past.

[0276] Step 83: Based on the assessment results of potential risks and the resource distribution within the entire monitoring area, determine the priority targets for handling.

[0277] Specifically, prioritizing individuals can be achieved by analyzing risk levels and resource availability. For example, high-risk individuals can be marked as priority targets, and the nearest monitoring resources can be allocated for tracking. In a city square monitoring scenario, if an individual is judged to be high-risk, cameras at nearby monitoring points can be prioritized for real-time tracking. This prioritization approach can effectively improve response efficiency.

[0278] Step 84: For priority processing objects, perform real-time tracking, obtain updated location information, and determine whether the anomaly persists.

[0279] In one possible implementation, real-time tracking can be achieved by continuously updating the target's location data. For example, this can be done by integrating data from multiple monitoring points to record the target's latest movement trajectory. If the updated location information shows persistent anomalies, such as the target continuously operating in a high-risk area, further integration of multi-source data is needed to determine the next steps. It should be noted that real-time tracking can also be optimized by incorporating timestamp data to ensure the continuity of the tracking process.

[0280] Step 85: Optimize the flow trend analysis through subsequent processing to obtain more accurate monitoring feedback. Specifically, optimization analysis can be achieved by classifying the flow trends.

[0281] For example, personnel movement can be divided into two categories: normal movement and abnormal movement. Detailed monitoring and feedback reports can be generated for abnormal movement.

[0282] For example, in a train station monitoring scenario, if the flow trend in a certain area continues to be abnormal, a feedback report containing the time, location, and identity information of the abnormality can be generated to support management decisions. This optimization method can significantly improve the accuracy of monitoring feedback.

[0283] In one possible implementation, the parameters for calculating the similarity of feature data in step 41 can be optimized in combination with the specific scenario.

[0284] For example, in a hospital lobby surveillance scenario, people may have low distinguishability due to similar clothing.

[0285] Step 411: First, adjust the weight parameters of similarity calculation based on the population density of the monitored environment, such as increasing the weight of location features and reducing the influence of shape features.

[0286] Step 412: For the adjusted parameters, calculate the similarity values ​​between each target in the feature set and generate a similarity distribution map.

[0287] Step 413: Based on the distribution map, select target pairs with similarity higher than a preset threshold as preliminary matching objects.

[0288] Step 414: Record the parameter adjustment log during the similarity calculation process for subsequent verification and optimization. This parameter optimization method can effectively improve the adaptability of similarity calculation and ensure the accuracy of matching results.

[0289] In another embodiment, the process of time alignment of trajectory data in step 51 can be refined from the perspective of data continuity verification.

[0290] For example, in a port terminal monitoring scenario, the movement trajectory of personnel may be discontinuous due to monitoring blind spots.

[0291] Step 511: First, perform timestamp matching on the structured trajectory dataset to check for any missing data within a time window.

[0292] Step 512: For cases with missing data, analyze the distribution of trajectory data in the time windows before and after the missing data to determine whether the missing data affects the overall continuity.

[0293] Step 513: Based on the analysis results, adjust the time alignment parameters, for example, by increasing the interpolation weight for time windows with more missing data.

[0294] Step 514: Generate aligned time series data and record missing data handling logs during the alignment process for subsequent optimization. This verification and adjustment method can effectively improve the accuracy of time alignment and ensure the continuity of trajectory data.

[0295] In one possible implementation, the process of standardizing and adjusting the feature data in step 61 can be optimized by incorporating the dynamic characteristics of light changes.

[0296] For example, in a nighttime street surveillance scenario, lighting conditions may change constantly with time and location.

[0297] Step 611: First, analyze the light distribution characteristics of the feature dataset to identify time periods and areas with significant light changes.

[0298] Step 612: Based on the recognition results, dynamically adjust the normalization parameters of the standardization process, for example, increase the adjustment range for feature data in darker areas.

[0299] Step 613: After completing the parameter adjustment, perform standardization processing to generate a standardized feature dataset.

[0300] Step 614 involves recording parameter adjustment logs during the standardization process for subsequent verification and optimization. This dynamic adjustment method effectively addresses the complexity of lighting variations and ensures the reliability of the standardization results.

[0301] In another embodiment, the process of judging the matching degree of feature data in step 71 can be refined from the perspective of multi-point data integration.

[0302] For example, in a commercial street surveillance scenario, the data from neighboring surveillance points may not be completely consistent due to differences in viewing angles.

[0303] Step 711: First, obtain the compensated feature dataset from the nearby monitoring points and record the acquisition perspective and time information of each dataset.

[0304] Step 712: Adjust the parameters for matching degree calculation based on viewpoint differences, for example, reduce the weight of shape features for datasets with large viewpoint differences.

[0305] Step 713: Calculate the matching degree value based on the adjusted parameters and determine whether the preset conditions are met.

[0306] Step 714 generates a matching degree calculation report, recording parameter adjustments and comparison results for subsequent verification. This multi-point data integration method can effectively reduce matching errors caused by differences in perspective and improve the accuracy of the judgment results.

[0307] In one possible implementation, the process of analyzing personnel flow trends in step 81 can be analyzed from multiple dimensions in combination with the characteristics of the scenario.

[0308] For example, in a large stadium monitoring scenario, the flow of people may be affected by the timing and geographical distribution of events.

[0309] Step 811: First, collect population density data for different time periods and regions to generate a flow trend distribution map.

[0310] Step 812: Analyze the changing patterns of the flow trend based on the distribution map. For example, the density of the entrance area increases before the start of the event and the density of the exit area increases after the end of the event.

[0311] Step 813: Based on the analysis results, adjust the threshold range of the flow trend, for example, appropriately increase the density threshold during the peak of the event.

[0312] Step 814 generates a flow trend analysis report, recording the threshold adjustment and change patterns for subsequent anomaly detection. This multi-dimensional analysis method can effectively improve the adaptability of flow trend judgment and ensure that the detection results meet the needs of actual scenarios.

[0313] In another embodiment, the real-time tracking process in step 84 can be optimized from the perspective of data update frequency.

[0314] For example, in a city square surveillance scenario, high-risk individuals may move quickly, requiring frequent updates to their location data.

[0315] Step 841: First, adjust the real-time tracking data update frequency according to the risk level of the target. For example, update the location data once per second for high-risk targets.

[0316] Step 842: Optimize the data integration method for multiple monitoring points based on the update frequency, such as prioritizing the processing of data from neighboring monitoring points.

[0317] Step 843: Combine the integrated data to generate the target's real-time trajectory path and determine whether the anomaly continues.

[0318] Step 844 records the update frequency and data integration logs during real-time tracking for subsequent optimization. This frequency optimization method can effectively improve the response speed of real-time tracking and ensure that anomalies are handled promptly.

[0319] It should be noted that the implementation of each step and embodiment described above revolves around personnel trajectory recognition and dynamic management, focusing on data processing and identity association using multi-sensor devices. The parameters and processing methods of each step can be flexibly adjusted to suit the characteristics of different monitoring scenarios, such as personnel density, environmental conditions, and movement patterns. For example, in densely populated scenarios, the redundancy of feature matching can be increased to improve the accuracy of identity association through multi-dimensional data; in scenarios with complex environmental conditions, environmental compensation methods can be optimized to enhance the robustness of feature data. This flexible adjustment approach enables the technical solution of this invention to adapt to various practical application needs, providing strong support for the dynamic management of a comprehensive monitoring network.

[0320] In the description of this invention, it should be understood that the terms "coaxial," "bottom," "one end," "top," "middle," "other end," "upper," "side," "top," "inner," "front," "center," "both ends," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0321] Furthermore, the terms “first,” “second,” “third,” and “fourth” are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as “first,” “second,” “third,” or “fourth” may explicitly or implicitly include at least one of those features.

[0322] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "setting," "connection," "fixing," "screw connection," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal connection of two components or the interaction between two components. Unless otherwise explicitly limited, those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0323] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A campus security monitoring method based on image recognition technology, characterized in that, include: Based on the original video frame sequence collected by multiple sensor devices, a time synchronization method is used to adjust the clock deviation between target devices to obtain a synchronization timestamp sequence. Based on the synchronization timestamp sequence, personnel target boundary information is extracted from video frames of each monitoring point, and multi-dimensional feature data is generated using a target detection method to verify cross-frame feature consistency. If the similarity between the feature data meets the preset conditions, the corresponding personnel targets will be associated with the same identity to obtain a preliminary identity association structure. Based on the preliminary identity association structure, obtain the personnel trajectory data set when moving across regions, determine the continuity of the trajectory data set in the synchronous timestamp sequence, and obtain the complete trajectory path; Based on the environmental compensation method, the feature data affected by changes in light in the complete trajectory path are normalized to obtain the compensated feature data. If the matching degree between the compensated feature data and the data of the neighboring monitoring points meets the conditions, then the angle difference information in the complete trajectory path is fused to obtain the optimized identity recognition result; Based on the optimized identity recognition results, a dynamic management structure for personnel movement in the full-domain monitoring network is generated, relevant trajectory anomalies are detected, and response support information is obtained; The step of fusing the angle difference information in the complete trajectory path to obtain the optimized identity recognition result if the matched degree between the compensated feature data and the data of neighboring monitoring points meets the condition includes: Raw feature data is obtained from monitoring equipment, and the data is initially cleaned and formatted to obtain the processed initial feature dataset. If the difference between the initial feature dataset and the pre-established reference dataset exceeds a preset range, the initial feature dataset will be compensated and adjusted to obtain the compensated feature dataset. The compensated feature dataset is compared and analyzed with the data collected from neighboring monitoring points. If the comparison result meets the preset matching degree condition, then the dataset has the basic conditions for fusion processing. For the compensated feature dataset that meets the conditions, obtain the relevant path information in the complete trajectory path, extract the angle difference data, and obtain the angle difference information set to be fused. The angle difference information set and the compensated feature dataset are integrated, and the integrated data are classified using a support vector machine model to obtain preliminary identity recognition results. If the initial identity recognition result does not match the historical data to the preset threshold, the integrated data will be calibrated a second time to generate an optimized identity recognition result. By performing a final verification on the optimized identity recognition results and supplementary data from nearby monitoring points, it is determined whether the results meet the preset accuracy standards, and a final confirmed identity recognition conclusion is obtained.

2. The campus security monitoring method based on image recognition technology according to claim 1, characterized in that: The step of adjusting the clock deviation between target devices using a time synchronization method based on the original video frame sequence acquired by the multi-sensor devices to obtain a synchronized timestamp sequence includes: The initial frame sequence data is obtained by collecting raw video data from multiple sensor devices. The clock deviation between the target devices is detected to determine whether there is a time offset. If an offset is detected, the initial frame sequence data is time-calibrated by calculating the offset to obtain a calibrated frame sequence. Based on the calibrated frame sequence, the timestamp information of the video frames acquired by each device is extracted to construct a preliminary timestamp sequence; The standard deviation of the initial timestamp sequence is checked. If outliers are found in the timestamp sequence, the outliers are corrected using a linear interpolation method to obtain the corrected timestamp sequence. Based on the corrected timestamp sequence, data alignment processing is performed to obtain an aligned timestamp sequence; Based on the aligned timestamp sequence, Kalman filtering is performed on the timestamp sequence to smooth it and obtain the final synchronization timestamp sequence. By associating and mapping the final synchronization timestamp sequence with the calibrated frame sequence, a synchronized video data stream is constructed, completing the data integration of multiple sensor devices.

3. The campus security monitoring method based on image recognition technology according to claim 1, characterized in that: The steps of extracting personnel target boundary information from video frames of each monitoring point based on the synchronization timestamp sequence, generating multi-dimensional feature data using a target detection method, and verifying the cross-frame feature consistency include: By obtaining video frame data corresponding to the synchronization timestamp from the video of the monitoring point, and using a preset image processing tool to perform preliminary segmentation of the video frame data, the initial boundary information of the personnel target is obtained. Based on the initial boundary information, target detection technology is used to accurately locate personnel targets and generate multi-dimensional feature information including position and shape; For the multidimensional feature information, feature change data across frames is obtained, and by comparing the feature changes of adjacent video frames, it is determined whether there is a deviation in feature consistency. If there is a deviation in feature consistency, the degree of deviation is evaluated by a preset threshold to obtain a quantitative result of feature consistency. Based on the quantization results, the corresponding timestamp data in the time series is obtained. By comparing the timestamp data with the feature consistency results, the continuity status of cross-frame features is obtained. Based on the continuity state, a convolutional neural network is used to perform in-depth analysis of feature consistency to obtain the final cross-frame feature consistency judgment result. Based on the final judgment result, other relevant frame data from the video data source are obtained, and the integrity assessment of feature consistency is obtained by comparing and supplementing the information.

4. The campus security monitoring method based on image recognition technology according to claim 1, characterized in that: The step of associating corresponding personnel targets with the same identity and obtaining the preliminary identity association structure if the similarity between the feature data meets a preset condition includes: By obtaining feature data from multiple sources, comparison tools are used to perform preliminary cleaning and format unification of the feature data, thereby obtaining a processed feature set. If the similarity between the data in the processed feature set reaches a preset standard, the comparison mechanism is triggered to obtain the preliminary matching personnel target; Based on the initially matched personnel targets, cluster analysis is used to group the targets and determine whether there are any potential identity associations. If there is an identity association between the grouped targets, then the grouping results are analyzed in depth using association rule mining tools to obtain the preliminary structure of the identity association. For the identity associations in the preliminary structure, the association strength data is obtained. If the association strength data is lower than a preset threshold, weak association targets are removed to obtain an optimized association structure. By performing multi-dimensional verification on the optimized association structure and using preset verification rules to perform secondary confirmation on the targets within the structure, the final identity association result is obtained. If there are still unconfirmed targets in the final identity association results, additional feature data is obtained by supplementing the data source, and the comparison and association process is repeated to determine the final identity consistency.

5. A campus security monitoring method based on image recognition technology according to claim 1, characterized in that: The step of obtaining a set of personnel trajectory data during cross-regional movement based on the preliminary identity association structure, determining the continuity of the trajectory data set in the synchronized timestamp sequence, and obtaining the complete trajectory path includes: By establishing a preliminary identity association structure, the raw data set of personnel trajectories is obtained from data sources of cross-regional movement, and preliminary data processing is completed to obtain a structured trajectory dataset. Based on the structured trajectory dataset, time alignment processing is performed on the synchronization time and timestamp sequence, and a time window partitioning method is used to obtain the continuous distribution of data in the timestamp sequence; If the data continuity distribution in the timestamp sequence is interrupted, interpolation is performed using historical trajectory data to complete the sequence, and the completed time series dataset is obtained. It is then determined whether its continuity meets the preset threshold requirements. By using the completed time series dataset, we analyze the changing patterns of people's trajectories during cross-regional movement, and use trajectory smoothing to obtain smoothed trajectory data records. Based on the smoothed trajectory data records and combined with the continuity analysis results, a complete trajectory path is constructed, and key nodes and turning information in the path are obtained. Obtain key node data in the complete trajectory path, analyze the movement between nodes, classify the path segments, and obtain the feature set of the classified path segments. By combining the classified path segment feature set with identity association information, the cross-regional movement pattern of personnel trajectories is analyzed, and the final trajectory path structure model is constructed.

6. The campus security monitoring method based on image recognition technology according to claim 1, characterized in that: The step of normalizing the feature data affected by light changes in the complete trajectory path based on the environmental compensation method to obtain the compensated feature data includes: Feature data affected by changes in light are extracted from the trajectory path, and the feature data are initially corrected using a preset environmental compensation model to obtain corrected initial data. Based on the corrected initial data, a normalization method is used to standardize and adjust the data to obtain a standardized feature dataset. Based on the standardized feature dataset, the specific impact range of light interference on trajectory features is analyzed to obtain key data points affected by the environment. Based on the key data points affected by the environment, a secondary compensation process is performed using data correction technology to obtain optimized trajectory feature data. By performing path analysis on the optimized trajectory feature data, it is determined whether there are any abnormal deviations in the trajectory path. If a deviation is detected, the relevant data points are marked, and the marked path data is obtained. Based on the marked path data, and taking into account factors such as environmental influences and changes in light, a support vector machine algorithm is used to classify the trajectory path to obtain the final compensation data result. The final compensation data results are verified to determine whether the data processing meets the preset threshold requirements. If not, the data is corrected again to obtain a trajectory feature dataset that meets the requirements.

7. A campus security monitoring method based on image recognition technology according to claim 1, characterized in that: The steps of generating a dynamic management structure for personnel movement in the overall monitoring network based on the optimized identity recognition results, detecting related trajectory anomalies, and obtaining response support information include: By constructing a basic information database within the entire monitoring network using identity recognition data, the initial records of personnel movement are obtained from the basic information database to obtain the correspondence between identity and location; Based on the aforementioned correspondence, the random forest algorithm is used to analyze the real-time changes in personnel flow and to obtain the distribution of flow trends from the random forest algorithm. If the distribution of the flow trend exceeds the preset threshold range, the trajectory anomaly detection mechanism is triggered. The trajectory anomaly detection mechanism performs a deep comparison of the abnormal data to determine whether there is a potential risk. Based on the assessment results of the potential risks, and combined with the resource distribution within the entire monitoring scope, priority targets are obtained from the resource distribution. Based on the priority objects, abnormal trajectories are tracked in real time to obtain updated location information; If the updated location information continues to show abnormalities, then integrate multi-source data within the monitoring range from the information acquisition interface, and determine the direction of subsequent processing based on the multi-source data; Based on the aforementioned subsequent processing direction, the flow trend is optimized to obtain more accurate monitoring feedback.

8. A campus security monitoring method based on image recognition technology according to claim 2, characterized in that: In the step of performing Kalman filtering smoothing on the timestamp sequence based on the aligned timestamp sequence to obtain the final synchronization timestamp sequence, the OpenCV tool is used to perform Kalman filtering smoothing on the timestamp sequence.

9. A campus security monitoring method based on image recognition technology according to claim 5, characterized in that: In the steps of obtaining key node data in the complete trajectory path, analyzing movement between nodes, classifying the path segments, and obtaining a feature set of the classified path segments, the path segments are classified using the K-means clustering algorithm.

Images