A dynamic bright spot target statistics and analysis method based on visual tracking

Abstract

The present application relates to the technical field of visual target analysis, and discloses a dynamic bright spot target statistics and analysis method based on visual tracking. The method collects a reference video sequence to construct a background model database, acquires a real-time video stream data set, detects visual bright spot targets and calculates average brightness values as detection reference thresholds. The visual bright spot targets are analyzed in combination with the background model database to identify and mark potential abnormal bright spot targets. When there are potential abnormal bright spot targets, their spatial position information and motion trajectory features are extracted to determine whether they are real abnormal bright spot targets. After being determined as real abnormal bright spot targets, initial risk assessment values are calculated according to the occurrence frequencies, shape contour and size change features are extracted, and a feature matching algorithm is used for similarity grouping analysis. According to the analysis result, it is determined whether there is a dynamic flickering bright spot target. If there is, an adjustment factor is adjusted based on the flickering mode features to correct the initial risk assessment value, and the subsequent dynamic tracking frame rate is set according to the corrected value.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of visual target analysis, in particular to a dynamic bright spot target statistics and analysis method based on visual tracking. BACKGROUND

[0002] In various scenes requiring visual monitoring, the capture and analysis of dynamic bright spot targets are common tasks. In the prior art, the detection of bright spot targets often uses a preset threshold. Such a threshold can work when the ambient light is stable, but when the light appears to fluctuate, especially in scenes such as day and night alternation and cloud cover, the preset threshold will lose its adaptability. A threshold that is too high relative to the actual light level will filter out bright spots that should be detected, while a threshold that is too low will misjudge stray light in the background as a bright spot, resulting in a large amount of invalid information.

[0003] The frame rate selection in the tracking process also has limitations. Most solutions use a fixed frame rate. When the number of bright spots in the monitoring range increases sharply or the movement speed increases, a fixed frame rate cannot record enough target details in a unit of time, resulting in broken movement trajectories. In scenes where bright spots are sparse and movement is slow, a fixed frame rate will continue to consume excessive computing resources, resulting in unnecessary consumption.

[0004] For the identification of abnormal bright spots, existing methods often rely on a single feature, such as determining whether the brightness exceeds the normal range. However, in actual scenarios, some bright spots have high brightness but are on a normal movement path, and some bright spots have irregular movement but do not reach the threshold. Relying on a single feature will lead to deviations in abnormality judgment. In addition, some bright spots have periodic flickering, and this dynamic characteristic has not been considered in the evaluation of abnormality, resulting in a gap between the final analysis results and the actual situation. SUMMARY

[0005] The present application aims to provide a dynamic bright spot target statistics and analysis method based on visual tracking to solve the problems raised in the background.

[0006] To achieve the above-mentioned purpose, the present application provides a dynamic bright spot target statistics and analysis method based on visual tracking, which comprises:

[0007] Collecting a reference video sequence and constructing a background model database; obtaining a real-time video stream dataset, determining a dynamic tracking frame rate according to the data size of the real-time video stream dataset, detecting all visual bright spot targets in the real-time video stream and calculating the average brightness value of the visual bright spot targets; setting the average brightness value as a detection reference threshold; analyzing and processing the visual bright spot targets according to the detection reference threshold and the background model database, identifying and marking potential abnormal bright spot targets;

[0008] When there is a potential abnormal bright spot target, spatial position information and motion trajectory features of each potential abnormal bright spot target are extracted; whether the potential abnormal bright spot target is a real abnormal bright spot target is judged based on the spatial position information and the motion trajectory features;

[0009] When it is determined that the abnormal bright spot target is real, an initial risk assessment value is calculated according to a frequency number of appearance of the abnormal bright spot target; a shape contour feature and a size change feature of the real abnormal bright spot target are extracted; a similarity grouping analysis is performed on all real abnormal bright spot targets by using a feature matching algorithm;

[0010] Whether a dynamic flickering bright spot target exists is judged according to a result of the similarity grouping analysis; when it is determined that the dynamic flickering bright spot target exists, an adjustment factor is determined based on a flickering mode feature of the dynamic flickering bright spot target to modify the initial risk assessment value; and a subsequent dynamic tracking frame rate setting is adjusted based on the modified risk assessment value.

[0011] Preferably, when the dynamic tracking frame rate is determined according to the data amount size of the real-time video stream data set, the following operations are performed: the data amount size is compared with a preset low data amount threshold and a preset high data amount threshold respectively; a corresponding dynamic tracking frame rate configuration is output according to a comparison result; wherein the preset low data amount threshold is lower than the preset high data amount threshold; when the data amount size does not exceed the preset low data amount threshold, the dynamic tracking frame rate is set to a low frame rate mode; when the data amount size exceeds the preset low data amount threshold but does not exceed the preset high data amount threshold, the dynamic tracking frame rate is set to a medium frame rate mode; when the data amount size exceeds the preset high data amount threshold, the dynamic tracking frame rate is set to a high frame rate mode; the frame rate value corresponding to the low frame rate mode is lower than that corresponding to the medium frame rate mode, and the frame rate value corresponding to the medium frame rate mode is lower than that corresponding to the high frame rate mode.

[0012] Preferably, when the visual bright spot target is analyzed and processed according to the detection reference threshold and the background model database, the following operations are performed: the brightness value of all visual bright spot targets in the real-time video stream is compared with the detection reference threshold; and the shape feature of each visual bright spot target is matched and checked with the standard shape feature in the background model database; potential abnormal bright spot targets are identified and marked according to a result of the comparison and the matching and checking; when the brightness value of the visual bright spot target exceeds a set proportion range of the detection reference threshold, the visual bright spot target is determined to be a potential abnormal bright spot target and is marked; when the shape feature of the visual bright spot target is not found in the matching record in the background model database, the visual bright spot target is determined to be a potential abnormal bright spot target and is marked.

[0013] Preferably, in judging whether the potential abnormal bright spot target is a real abnormal bright spot target based on the spatial position information and the motion trajectory feature, the following operation is performed: when the spatial position information of the potential abnormal bright spot target shows that its motion trajectory deviates from the preset standard path, it is determined that the potential abnormal bright spot target is a real abnormal bright spot target; when the motion trajectory feature of the potential abnormal bright spot target indicates that its motion speed change exceeds the normal change range in the background model database, it is determined that the potential abnormal bright spot target is a real abnormal bright spot target.

[0014] Preferably, in calculating the initial risk assessment value according to the frequency of occurrence of the abnormal bright spot target, the following operation is performed: the total number of occurrences of the real abnormal bright spot target in the real-time video stream is counted; the size deviation amount is calculated by combining the size change feature of each real abnormal bright spot target and the reference size feature in the background model database; and the initial risk assessment value is output based on the total number of occurrences and the size deviation amount; wherein the size deviation amount reflects the relative change degree of the real abnormal bright spot target relative to the reference size feature.

[0015] Preferably, in performing similarity grouping analysis on all real abnormal bright spot targets using a feature matching algorithm, the following operation is performed: the shape contour feature and the size change feature of each real abnormal bright spot target are combined into a feature vector representation; the similarity scores between all feature vectors are calculated using the feature matching algorithm; a similarity score threshold is set; the feature vector groups whose similarity scores exceed the similarity score threshold are identified by the feature matching algorithm; dynamic clustering processing is performed on each feature vector group; it is judged whether there is a dynamic flickering bright spot target according to the results of the dynamic clustering processing; when a feature vector group contains at least two real abnormal bright spot targets and the shape contour features of the real abnormal bright spot targets show periodic changes, it is determined that the real abnormal bright spot targets in the feature vector group are dynamic flickering bright spot targets.

[0016] Preferably, in modifying the initial risk assessment value by an adjustment factor based on the flickering mode feature of the dynamic flickering bright spot target, the following operation is performed: the flickering frequency feature and the flickering intensity feature of the dynamic flickering bright spot target are extracted; the similarity between the flickering frequency feature and the flickering intensity feature and the reference flickering feature in the historical modification record is calculated; the adjustment factor is selected according to the similarity calculation result; when the similarity between the reference flickering feature in the historical modification record and the current flickering frequency feature and the flickering intensity feature exceeds a preset similarity threshold, the corresponding reference adjustment factor in the historical modification record is used as the current adjustment factor; when the similarity of all reference flickering features does not exceed the preset similarity threshold, the adjustment factor is calculated according to the total number of dynamic flickering bright spot targets; the initial risk assessment value is multiplied by the adjustment factor to output a modified risk assessment value.

[0017] Preferably, when calculating the adjustment factor according to the total number of dynamic flickering bright spot targets, the following operation is performed: the size of the adjustment factor is in direct proportion to the total number of dynamic flickering bright spot targets.

[0018] Preferably, when adjusting the subsequent dynamic tracking frame rate setting based on the modified risk assessment value, the following operation is performed: obtaining the modified risk assessment value; calculating a frame rate adjustment coefficient according to the size of the modified risk assessment value; the frame rate adjustment coefficient is in inverse proportion to the modified risk assessment value; multiplying the current dynamic tracking frame rate by the frame rate adjustment coefficient to output an adjusted subsequent dynamic tracking frame rate.

[0019] Preferably, after the adjusted subsequent dynamic tracking frame rate is set, the following operation is performed: applying the adjusted subsequent dynamic tracking frame rate to a newly collected real-time video stream data set; re-detecting visual bright spot targets and updating the background model database; continuously monitoring changes in the flickering pattern characteristics of dynamic flickering bright spot targets; when the flickering pattern characteristics change exceeds a preset change threshold, recalculating the adjustment factor and iteratively modifying the risk assessment value; outputting the updated risk assessment value and dynamic tracking frame rate configuration.

[0020] Compared with the prior art, the beneficial effects of the present application are:

[0021] Collecting the reference video sequence to construct the background model database provides a reference basis for subsequent analysis and processing, and can better adapt to different background environments in different scenes. By determining the dynamic tracking frame rate according to the size of the real-time video stream data, the calculation resources can be reasonably allocated under the premise of ensuring the tracking effect, and unnecessary consumption of resources can be avoided.

[0022] Detecting visual bright spot targets in real-time video stream and calculating the average brightness value as a detection reference threshold makes the threshold setting more suitable for the light conditions of the real-time scene, reducing the false detection or missed detection caused by changes in light. According to the detection reference threshold and the background model database, the visual bright spot targets are analyzed and processed, and the potential abnormal bright spot targets are identified and marked in combination with various information, improving the reliability of the preliminary screening of abnormal targets.

[0023] When there are potential abnormal bright spot targets, the spatial position information and motion trajectory characteristics thereof are extracted to determine whether they are real abnormal bright spot targets, which comprehensively considers the spatial and motion characteristics of the targets, making the identification of real abnormal targets more convincing. According to the frequency of the occurrence of the abnormal bright spot targets, an initial risk assessment value is calculated, and in combination with the extracted morphological contour characteristics and size change characteristics, a feature matching algorithm is used for similarity grouping analysis, which can classify the abnormal targets more carefully.

[0024] According to the similarity grouping analysis result, it is judged whether there is a dynamic flickering bright spot target, when there is, based on the flickering mode characteristics, an adjustment factor is determined to correct the initial risk assessment value, and the subsequent dynamic tracking frame rate is adjusted according to the adjusted value, so that the risk assessment is more in line with the actual situation, and the adjustment of the tracking frame rate is more targeted, and the dynamic changing scene can be better coped with. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 A working principle diagram of the dynamic bright spot target statistical and analysis method based on visual tracking is provided.

[0026] Figure 2 A flowchart for potential abnormal bright spot target identification is provided.

[0027] Figure 3 A flowchart for adjustment factor determination and risk assessment value correction is provided.

[0028] Figure 4 A flowchart for dynamic tracking frame rate adjustment is provided.

[0029] Figure 5 A flowchart for dynamic tracking frame rate iterative updating is provided. DETAILED DESCRIPTION

[0030] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0031] Please refer to Figure 1 The present application provides a dynamic bright spot target statistical and analysis method based on visual tracking, which comprises:

[0032] Collecting a reference video sequence to construct a background model database.

[0033] Obtaining a real-time video stream data set, determining a dynamic tracking frame rate according to the data size of the data set, detecting all visual bright spot targets in the real-time video stream and calculating the average brightness values of these targets.

[0034] Setting the calculated average brightness values as a detection reference threshold.

[0035] According to the detection reference threshold and the constructed background model database, the visual bright spot targets in the real-time video stream are analyzed and processed, and potential abnormal bright spot targets are identified and marked.

[0036] When there is a marked potential abnormal bright spot target, the current spatial position information and motion trajectory features of each target are extracted.

[0037] Based on the extracted spatial position information and motion trajectory features, it is determined whether the potential abnormal bright spot target is a real abnormal bright spot target.

[0038] When it is determined that the target is a real abnormal bright spot target, the number of occurrences of the target in the real-time video stream is counted, and based on this, an initial risk assessment value is calculated. At the same time, the shape contour features and size change features of the real abnormal bright spot target are extracted. A feature matching algorithm is used to perform similarity grouping analysis of all identified real abnormal bright spot targets based on the shape contour features and size change features.

[0039] Based on the results of the similarity grouping analysis, it is determined whether there is a dynamic flickering bright spot target. If it is determined that there is a dynamic flickering bright spot target, an adjustment factor is determined based on the flickering pattern features of these targets, and the initial risk assessment value is modified using the adjustment factor. Finally, based on the modified risk assessment value, the dynamic tracking frame rate setting applied in subsequent processing is adjusted.

[0040] Embodiment 1: refer to Figure 2 In the implementation of determining the dynamic tracking frame rate based on the data size of the real-time video stream data set, the preset low data size threshold and the preset high data size threshold are predefined in the system configuration, and the value of the preset low data size threshold is strictly less than the value of the preset high data size threshold. The system processor obtains the overall data size of the current real-time video stream data set, which generally reflects the comprehensive characteristics such as video stream duration, resolution or total pixel quantity. The data size is compared with the preset low data size threshold and the preset high data size threshold in sequence, and the comparison operation generates a corresponding comparison result code. Based on the generated comparison result code, the system automatically outputs the corresponding dynamic tracking frame rate configuration instruction. If the data size value is equal to or less than the preset low data size threshold, the frame rate configuration instruction specifies that the system enters the low frame rate mode, and the frame rate value corresponding to the low frame rate mode is generally in a lower fixed range, for example, 5 to 10 frames of image per second. When the data size value is greater than the preset low data size threshold but less than or equal to the preset high data size threshold, the frame rate configuration instruction specifies that the medium frame rate mode is enabled, and the frame rate value corresponding to the medium frame rate mode is higher than the low frame rate mode value, in an intermediate range, for example, 15 to 25 frames of image per second. In the case where the data size value exceeds the preset high data size threshold, the frame rate configuration instruction specifies that the high frame rate mode is switched to, and the frame rate value corresponding to the high frame rate mode is the highest, for example, 30 frames or higher of image data per second. After the system completes the mode setting, the video processing unit acquires and processes the image sequence according to the corresponding frame rate value of the selected low frame rate mode, medium frame rate mode or high frame rate mode.

[0041] In the process of analyzing and processing visual bright spot targets according to the detection reference threshold and the background model database, the system processor performs pixel-by-pixel scanning on each frame of image of the real-time video stream, identifies and locates all visual bright spot targets, and assigns a unique identifier to each detected target. For each identified visual bright spot target, the system extracts its average brightness value or brightness distribution statistical value. The brightness statistical value of the target is compared and analyzed numerically with the detection reference threshold calculated previously to generate a brightness deviation report. At the same time, the system extracts the key morphological features of the target, including but not limited to the approximate polygon of the target's contour shape, the aspect ratio, the area ratio, the contour complexity index, and the Fourier descriptor, etc. The morphological feature vector is sent to the background model database for matching query, and the query operation traverses all pre-stored standard morphological feature records in the database. The matching check process calculates the similarity measure value between the query vector and the stored vector in the database.

[0042] The system performs the identification and labeling logic of potential abnormal bright spot targets by combining the report generated by the brightness comparison analysis and the matching result returned by the morphological feature database query. The proportion range set by the brightness analysis module is defined by the system configuration parameters, for example, the range interval of the upper and lower floating percentage of twenty percent of the brightness threshold. When the brightness statistical value of the visual bright spot target is higher or lower than the fixed proportion range of the detection reference threshold, regardless of the morphological feature matching check result, the system decision logic module outputs an instruction to label the target as a potential abnormal bright spot target and attach a specific abnormal identification code to its unique identifier. The morphological matching check module independently runs its decision logic, and if the database query of the morphological feature vector of a target fails to return any valid matching record, i.e. the similarity score of all stored features in the database with the target feature is lower than the preset matching tolerance lower limit, regardless of the brightness comparison result of the target, the system decision logic module also outputs an instruction to label it as a potential abnormal bright spot target and assign the same type of identification code. After the target labeling is completed, the subsequent processing module only focuses on and operates the target instances with the abnormal identification code. The original data of all targets, the brightness analysis report, the morphological matching result, and the labeling status are recorded to the system log.

[0043] The entire data processing flow adopts a multi-thread architecture, the luminance contrast analysis module and the morphological feature matching check module operate in parallel, and the system determination logic module receives the output information of the two modules in real time. The luminance contrast analysis module calculates the absolute deviation and the relative percentage deviation of the luminance value of each target from the detection reference threshold. The morphological feature matching check module calculates the Euclidean distance between the target contour feature and the database record, uses the nearest neighbor search algorithm in the multi-dimensional feature space, and returns the closest records and their similarity scores. The system determination logic module determines the marking operation threshold according to the configuration strategy, and when the output of any module triggers an abnormal condition, the determination logic immediately updates the target state. The database connection management subroutine is responsible for maintaining the connection pool of the background model database, optimizing the query response time. In the low frame rate mode, the morphological matching check may use simplified feature vectors and relax the matching tolerance to improve processing efficiency. In the high frame rate mode, full-dimensional feature vectors are used for accurate matching, and strict tolerance threshold determination is performed. The marked target data together with its spatial position coordinates are highlighted in the visualization interface and trigger the start process of the subsequent processing unit, and the entire analysis processing result is simultaneously used as the input basic data for judging the reality of the potential target in the next stage. The system releases the memory resources for the normal targets that do not trigger abnormal conditions after performing routine data recording. The abnormal marking state is valid until the target leaves the monitoring area or is finally classified as a real abnormal bright point target or excluded from risk.

[0044] In the embodiment 2, when judging whether the potential abnormal bright point target is a real abnormal bright point target based on the extracted spatial position information and motion trajectory features, the system calls the spatial trajectory analysis module. The input of the module is a list of all visual bright point targets marked as potential abnormalities in the previous step and their associated data. The data packet of each potential abnormal bright point target includes its unique identifier, timestamp sequence, and spatial coordinate set corresponding to each time point. The spatial coordinates are represented in a two-dimensional or three-dimensional coordinate system established in the monitoring scene, and the origin and scale of the coordinate system are determined by system calibration parameters. The module preprocesses the target motion trajectory data, removes coordinate noise points, and performs smoothing interpolation processing on the coordinate sequence to generate a continuous trajectory curve.

[0045] The preset standard path information is stored in a system path configuration library. The path configuration library contains a plurality of predefined reference motion trajectory templates, each template trajectory is stored in the form of a sequence of control point coordinates and an interpolation function, and is associated with a specific target category identifier. The system performs a point-to-point alignment comparison between the actual motion trajectory point sequence of each potential abnormal bright point target and the corresponding category template trajectory in the path configuration library. The alignment process uses a time registration algorithm and a spatial transformation matrix to adjust the consistency of the space-time coordinate system of the target trajectory and the template trajectory. The system calculates the cumulative value of the perpendicular distance of the target trajectory point sequence from the template trajectory, and if the cumulative value exceeds the preset distance deviation threshold and lasts for more than a preset number of consecutive frames, it is determined that the target motion trajectory deviates from the standard path. The preset distance deviation threshold is dynamically set according to the target size and scene scale, and the deviation determination result is output by the trajectory analysis submodule.

[0046] The motion trajectory feature analysis subsystem is activated. This subsystem processes the raw displacement data and the time interval between consecutive frames of each potential abnormal target, and calculates its average speed, instantaneous speed extreme value, and speed variation standard deviation, etc. The background model database has special partition storage of normal motion speed variation range data of different target categories, including the reasonable interval of speed average, the maximum tolerance standard deviation of speed fluctuation, and the typical acceleration characteristic curve of a specific road section. The system compares and analyzes the motion speed feature vector calculated by the target with the normal variation range of the same target stored in the database. The comparison operation uses a combination of boundary check and trend similarity evaluation: if the target speed average exceeds the preset normal value interval, the instantaneous speed extreme value exceeds the maximum speed variation range allowed for this category, or the speed variation standard deviation is higher than the preset upper limit of fluctuation tolerance, the motion speed feature exceeds the normal variation range.

[0047] The system anomaly determination logic controller receives independent output reports from the spatial trajectory deviation analysis submodule and the motion speed feature analysis submodule. If the determination state report output by any submodule triggers an abnormal condition (i.e. meets the trajectory deviation from the standard path or the speed variation exceeds the normal range), the controller updates the state marker of the corresponding target identifier to a real abnormal bright point target, and generates an abnormal type code (the type code distinguishes between trajectory deviation type or speed abnormality type). The determination result is written into the global target state table, and at the same time triggers the data storage module to record the complete spatial position information, motion trajectory feature vector, determination basis data and time stamp log. The target state update operation notifies the subsequent processing flow to enable the operation unit for the real abnormal bright point target. The potential abnormal marker of the target that does not trigger any abnormal condition is removed and restored to the normal target tracking state.

[0048] In the calculation of the initial risk assessment value according to the frequency of the real abnormal bright spot target, the statistical calculation module queries the global target state table in the current video stream processing period, and filters all entries with the state marker of real abnormal bright spot target. The system traverses the appearance records of these target entries in the continuous video frame sequence, counts the number of appearances of each target, and then calculates the total number of appearances of all real abnormal bright spot targets in the current statistical time window. The span of the time window is set by the system configuration parameter, which usually covers the time length of all valid frames of the current processing video segment. The total number is used as a basic input quantity for risk assessment.

[0049] The size deviation amount calculation engine is executed in parallel. The engine calls the list of entities currently identified as real abnormal bright spot targets, extracts the size change feature of the feature database record of each target in the list. The target size change feature usually includes target area pixel area or axial length sequence information. The background model database has a reference size feature partition, which stores the size baseline data and historical statistics of different categories of targets in the standard state. The system reads the category attribute identifier of the target, and retrieves the corresponding reference size feature data set (which may include statistical average, mode, fitting curve equation, etc.) from the reference size partition. The calculation engine performs matching operations on the size measurement value or sequence statistical value of the target at the current time and the reference value, to generate a size deviation amount indicator. The deviation amount indicator can be designed as: the absolute difference of the current measurement value relative to the reference average; the percentage change rate of the difference relative to the reference average; or a distance measure based on time series (such as dynamic time warping distance). The deviation amount calculation result is recorded in the memory structure, associated with the corresponding target identifier.

[0050] The initial risk assessment value generator integrates the two data sources mentioned above: the total number of appearances of real abnormal bright spot targets data array, and the size deviation amount calculation result data set of all real abnormal targets. The system uses a weight algorithm model to process the data set. The total number of appearances is linearly mapped to the impact score A according to the system preset weight factor; the size deviation amount set of all targets is first normalized, and then the statistical average or weighted average is calculated, which is mapped to the impact score B through the preset conversion factor and nonlinear correction function. The score A and the score B are then output through a combination rule model (such as weighted sum or maximum value function) to output the original value of the initial risk assessment value. The value is scaled and transformed to a preset unified evaluation scale (e.g. 0 to 10 range), forming a comparable initial risk assessment value result. The finally generated initial risk assessment value is written into the system risk assessment state table, and the key parameter snapshots involved in the calculation are recorded, including the time window definition, the number of targets, and the size deviation amount distribution statistical value. The value is used as a basic assessment quantity in the subsequent adjustment process.

[0051] Example 3: see Figure 3In the implementation process of similarity grouping analysis of all real abnormal bright spot targets by using the feature matching algorithm, the system initializes the grouping analysis engine. The engine loads all instances of the current real abnormal bright spot targets determined and their associated feature data. The data record of each target instance contains its unique identifier, time stamp sequence, morphological contour feature array and size change feature matrix. The morphological contour feature array stores the contour polygon vertex coordinate set or contour Fourier descriptor sequence of the target in consecutive frames; the size change feature matrix records the length of the major axis, area value and length-width ratio change curve of the target at each time point. The grouping analysis engine constructs a composite feature vector for each target instance: aligning the morphological contour feature array and the size change feature matrix by time dimension, merging them into a unified multi-dimensional feature representation vector by a feature fusion algorithm . The vector dimension is determined by the number of contour feature points and the number of size parameters, and a fixed-length digital sequence is generated by using equal weight splicing.

[0052] The system enables the configured feature matching algorithm core (selecting the improved cosine similarity algorithm combined with the dynamic time warping algorithm), which calculates the pairwise similarity scores of the feature vectors between all real abnormal bright spot targets. The similarity score calculation process includes the following processing: standardizing each vector to eliminate dimensional differences; calculating the directional similarity in the vector space; and evaluating the morphological consistency of the feature sequence over time. The algorithm automatically generates an N x N similarity score matrix SM, where N is the total number of current targets. Set the system's preset similarity score threshold (taking a value range of 0.6-0.8 configurable), the algorithm scans the entire SM matrix, identifies all element pairs that satisfy , and identifies the target identifiers with high similarity into the same candidate group.

[0053] The clustering execution module receives the candidate group relationship graph and applies an incremental hierarchical clustering algorithm for processing. The algorithm establishes a similarity tree structure: taking the candidate groups as the initial clustering units, iteratively merging adjacent units with similarity exceeding to form hierarchical clustering clusters. During the clustering process, the distance calculation strategy is dynamically adjusted, and higher sensitivity weights are given to the contour components in the feature vectors . After completing the clustering process, a number of final feature vector groups are output, each contains a feature vector set of at least two target instances. The system analysis engine performs periodic pattern detection on each : extracts the morphological contour feature time sequence of all targets in the group, performs Fourier spectrum analysis and autocorrelation function calculation. When a significant spectral peak is detected and the contour changes in three consecutive periods conform to the preset periodic fluctuation template, it is determined that the The corresponding target set belongs to the dynamic flashing bright spot targets, and its periodic parameter indicators are recorded.

[0054] For the identified set of dynamically flickering bright spot targets, the system activates the flicker feature extraction unit. This unit calculates key flicker pattern features—flicker frequency features—from its original brightness data sequence for each target instance. (Number of complete brightness cycle changes per unit time) and flicker intensity characteristics (Percentage of the relative difference between the highest and lowest brightness within a period). The feature extraction process employs a sliding window peak detection algorithm, with the window size adaptively adjusting to the target's appearance duration. The acquired... and The constituent feature pairs (F,I) are submitted to the historical correction record database for similarity matching queries.

[0055] The historical correction record database is designed with a multi-level index structure: LSH (Local Sensitive Hashing) is used to bucket the historical reference flicker features. The similarity calculation process is defined as follows:

[0056]

[0057] in, Indicates the current flashing frequency The difference from the historical reference frequency, Indicates the intensity of the flash. The difference between the strength and the reference strength; This is the frequency difference weighting coefficient (typical value 0.7). This is the intensity difference weighting coefficient (typical value 0.3); A smoothing factor is used to prevent division by zero errors (fixed value 0.01). The system calculates the current (F,I) feature and compares it with all historical reference features. , )of Value, set a preset similarity threshold Does it exist in the search history? Matching items: If a matching record exists, extract the reference adjustment factor associated with that record. Used as the current adjustment factor; if no matching record is found, the adjustment factor calculation procedure is executed: ( This represents the total number of dynamically flashing bright targets. This is the scaling factor (adjustable from 0.05 to 0.2). Get the adjustment factor. Then, the risk correction module performs the following operation: reads the initial risk assessment value. Calculate the correction value The corrected result is output to the system risk assessment register. The entire process generates an audit log recording the adjustment factor calculation path and parameter trajectory.

[0058] The generation of the similarity matrix SM in the feature matching process employs a multi-thread block computation technique, which automatically activates GPU acceleration when the target number N > 100. The merging threshold of the clustering algorithm is adaptively adjusted according to the size of the feature vector group: a loose merging strategy is adopted for large-scale clusters, and a strict similarity verification is implemented for small-scale clusters. The detection of the flicker frequency employs an anti-interference design, which filters out false periodic signals caused by environmental noise through multi-resolution wavelet analysis. The historical correction record database implements a lazy update mechanism, which automatically archives new flicker pattern features as new reference records every 24 hours. The value of the adjustment factor is limited within the system preset safe boundary [0.5, 5.0] to avoid extreme correction results. After the risk correction operation is executed, the original and the corresponding relationship of the correction parameter are highlighted in the visualization interface for the operator to review. The system sets an abnormal handling program to monitor the feature matching time overhead, and automatically downgrades to a simplified matching mode to ensure real-time performance in case of timeout.

[0059] Example 4: Refer to Figure 4 When the system confirms that it cannot find a matching reference in the historical correction record according to the rules of Example 3, it automatically triggers the adjustment factor calculation program based on the total number of dynamic flickering bright target. Assume that the system detects 12 dynamic flickering bright target instances in the current monitoring period. The target number counter sends the integer value 12 to the adjustment factor generator. The generator presets a linear proportionality coefficient k = 0.15 (this parameter is stored in the system configuration file and can be dynamically adjusted). The calculation logic is executed: adjustment factor . The calculation process is completed in an independent arithmetic unit, and the result is rounded to two decimal places. The system synchronously records the calculation log, which includes the timestamp, input value, and parameter source.

[0060] After obtaining the corrected risk assessment value (the value comes from the risk correction module output in Example 3), the system activates the frame rate adjustment control module. The preset basic constant C = 220 (reflecting the maximum processing capacity of the system). The frame rate adjustment coefficient calculation unit performs the operation: divide the constant C by the corrected risk value to obtain the adjustment coefficient . This calculation is performed in a floating-point operation unit, and the result is rounded to two significant digits. The system detects that the current dynamic tracking frame rate is in the high frame rate mode, and its configuration value is 200 frames / second. The frame rate calculation engine executes: frames / second. Given that the maximum supported by the physical device is 200 frames / second, the system automatically activates the upper threshold protection mechanism, locking the output frame rate at 200 frames / second.

[0061] The following table shows the processing response logic of the system in six typical scenarios:

[0062] Dynamic flicker target number Risk correction value Current frame rate mode Calculated frame rate value Final execution frame rate Constraint handling type 5 6.20 Low frame rate (10) 35.48×10≈355 120 Hardware upper limit clipping 8 7.50 Medium frame rate (20) 29.33×20≈587 200 Multi-target optimization degradation 12 8.60 High frame rate (30) 25.58×30≈767 200 Physical device limiting 3 5.80 Low frame rate (10) 37.93×10≈379 120 Energy efficiency management constraint 15 9.40 High frame rate (30) 23.40×30≈702 180 Thermal protection frequency reduction 1 4.20 Medium frame rate (20) 52.38×20≈1047 100 Communication bandwidth limitation

[0063] The process of converting the calculated frame rate value into the actual execution frame rate undergoes multiple constraint processing: the frame rate mapping controller receives the original calculated value, which is sequentially filtered through the constraint filters of the physical layer, environmental layer, and policy layer. The physical layer filter checks the technical specifications of the camera sensor chip and discards values that exceed the maximum sampling rate of the light sensing element. The environmental layer filter calls real-time temperature monitoring data and automatically activates the frequency reduction algorithm when the case temperature exceeds 60°C, reducing the frame rate value in proportion to the temperature overrun. The policy layer filter applies the preset resource allocation rules: when the network transmission bandwidth occupancy rate exceeds 85%, the bandwidth protection mechanism is started, and the frame rate is executed according to the formula Dynamic compression.

[0064] The system implements a smooth transition technique when performing frame rate switching operations. Taking the above 12 target scenarios as an example: switching from 30 frames / second to 200 frames / second is implemented in three stages. The first stage (0-500 milliseconds) gradually increases the acquisition rate by 30%; the second stage (500-1000 milliseconds) enables the frame buffer pool preloading mechanism to fill the data gap; and the third stage (after 1000 milliseconds) stabilizes at 200 frames / second. The data integrity monitoring module continues to operate during the switching process, adding a timestamp check code to each millisecond of video stream data packet, and any check failure triggers reacquisition of data in that period. The target tracking algorithm uses motion trajectory prediction compensation technology during frame rate changes, estimates the potential displacement of the target during the frame interval change through Kalman filtering algorithm, and maintains the spatiotemporal continuity of target tracking.

[0065] The newly added processing resource allocation subsystem dynamically deploys computing resources according to the final execution frame rate. When the frame rate is increased to 200 frames / second, the system automatically calls the 16 computing cores of the GPU accelerated processing node and allocates dedicated memory channels to transfer high frame rate video streams. The video decoder switches to a lightweight mode and closes unnecessary color space conversion modules. The disk write module starts the circular buffer management strategy, retaining only the last 20 seconds of full frame rate raw data to avoid overloading the storage system. The system monitoring interface displays real-time frame rate adjustment parameter dashboards, including dynamic flashing target quantity bar charts, risk value change curves, and calculation / execution frame rate control tables, among other visual elements. Operators can manually set the k coefficient experimental interval on the dashboard, and the system automatically records the differences in running effects under different parameters. The abnormal state response program generates a system optimization suggestion report when the deviation between the calculated frame rate value and the execution frame rate continuously exceeds 50%, prompting the upgrade of hardware devices or the adjustment of monitoring scene parameters.

[0066] Example 5:Figure 5 After obtaining the adjusted dynamic tracking frame rate parameter generated according to the foregoing flow, the system operation execution module automatically configures the video acquisition hardware and processing pipeline. The parameter is written into the video stream controller register, which overwrites the original frame rate setting value. The controller synchronously adjusts the frequency of the image sensor clock signal and the bandwidth allocation of the data transmission channel according to the new parameter. The newly acquired real-time video stream data set starts to be input into the system buffer at the updated frame rate specification, and the data set is regarded as an independent input source in subsequent processing.

[0067] The video stream processing core completely executes the target detection full flow for the new data set: the pixel scanning unit re-recognizes all visual bright spot targets using the same brightness threshold parameter configuration; the spatial positioning engine allocates a three-dimensional coordinate system position identifier to each detected target; and the background model database enters a controlled update state, and its update operation fuses historical model data and statistical characteristics of the new input frame sequence. The update process adopts a hierarchical progressive manner: first, a sliding average calculation of the pixel point gray value is performed on the static background area; second, morphological repair and texture reconstruction are performed on the motion target residual area; and finally, an adaptive background learning algorithm is started for the high-frequency change area to calculate the pixel stability index, and the pixel points below the preset stability threshold are included in the background candidate set. The database version management subsystem automatically creates model snapshots before and after the update and establishes a difference log.

[0068] At the same time, the system flicker analysis thread focuses on the entity set that has been classified as a dynamic flickering bright spot target. Each target entity is bound to an independent monitoring thread, which maintains its flicker feature time sequence buffer. The flicker frequency feature is tracked using real-time spectrum analysis: a 1024-point fast Fourier transform is performed every second to record the amplitude change trajectory of the dominant frequency component; and the flicker intensity feature is calculated by moving the time window to calculate the peak-to-valley ratio, and the window length dynamically stretches according to the target duration. The feature change tracking unit generates a change rate report every second, including: frequency drift amount (the difference between the current frequency and the reference frequency), intensity fluctuation coefficient σ (the intensity standard deviation within the window). The system preset change threshold envelope is defined by a four-dimensional parameter space: frequency drift tolerance upper limit , intensity fluctuation coefficient warning value , continuous exceeding duration threshold , spatial consistency verification proportion . The feature analysis result is written into the shared memory exchange area for the judgment module to read.

[0069] When any dynamic flickering bright spot target meets the following combination of conditions, the re-computation flow is triggered: condition one, the frequency drift amount continuously exceeds for three consecutive sampling periods; and condition two, the intensity fluctuation coefficient mutates more than and maintained for more than two seconds; condition three, more than proportion of members simultaneously mutate the feature. The decision logic controller activates the feature re-extraction instruction, which interrupts the current processing pipeline with the highest priority. The system intercepts the target's recent 5-second high-frame-rate raw brightness data, recalculates its flicker frequency feature and flicker intensity feature. The new feature dataset is marked as a derivative calculation version, with a timestamp and spatial position verification code attached.

[0070] The historical correction record database enters the forced query mode: the query condition is expanded to include spatiotemporal environmental parameters, including geographic location code, weather identifier, and device working condition label. The similarity calculation uses an upgraded version of the metric algorithm, with an added weight factor for the continuity of feature mutations. If there is no matching record, the adjustment factor calculator uses the latest dynamic flicker target total number performs linear calculation, which has been updated according to the re-detection results. The risk correction module obtains the current real-time risk assessment intermediate value , which comes from the rolling calculation results of the last ten minutes. The adjustment factor and are multiplied to output . The frame rate adjustment coefficient generator responds immediately to changes in risk values, completing coefficient calculation and target frame rate conversion within 200 milliseconds.

[0071] The execution of the frame rate dynamic switching process uses a three-step buffering mechanism: the first stage configures 70% of the target frame rate as the temporary acquisition rate; the second stage opens the data pipeline preloading, filling the frame buffer to 80% capacity; the final stage switches to full target frame rate operation. During the entire switching period, the target tracking algorithm maintains a dual-track processing state: the existing track continues to be predicted by Kalman filtering, and new acquisition data simultaneously generates alternative tracks. The track fusion controller performs track matching and merging after switching is complete, and triggers spatial interpolation compensation for track segments with merging errors exceeding pixel tolerance. The system console displays real-time updates of the parameter matrix, including the number of re-calculation event triggers, the current feature mutation index, and the evolution curve of the iterative risk assessment value. The configuration management subsystem automatically generates configuration change audit reports every five minutes, recording the complete state transition path and the effectiveness of the constraint conditions. All re-calculation operations are marked as system optimization events, and the event logs are synchronously transmitted to the remote analysis platform.

[0072] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting; it is not intended to exclude myriad other embodiments of the present application that other inventors can develop based on the same general inventive concepts embodied by the described embodiments. That is, although the present application is described in terms of particular embodiments and implementations, it is to be understood that the terminology used is for the purpose of descriptive clarity and that it is intended to be limited only by the words recited in the appended claims. It is to be understood that the terms such as first and second, etc., merely are used to differentiate one from another without necessarily implying or requiring any actual relationship or order between them. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0073] While the embodiments of the application have been shown and described herein, it is to be understood that the application is not limited to these embodiments. Rather, many modifications, changes, substitutions, and alterations can be made thereto without departing from the spirit and scope of the present application as defined by the appended claims and their equivalents.

Claims

1. A method for dynamic bright spot target statistics and analysis based on visual tracking, comprising the following steps: collecting a reference video sequence and constructing a background model database; obtaining a real-time video stream dataset, determining a dynamic tracking frame rate according to the data size of the real-time video stream dataset, detecting all visual bright spot targets in the real-time video stream and calculating the average brightness value of the visual bright spot targets; setting the average brightness value as a detection reference threshold; analyzing and processing the visual bright spot targets according to the detection reference threshold and the background model database, identifying and marking potential abnormal bright spot targets; when there are potential abnormal bright spot targets, extracting the spatial position information and motion trajectory features of each potential abnormal bright spot target; judging whether the potential abnormal bright spot target is a real abnormal bright spot target based on the spatial position information and motion trajectory features; when it is determined to be a real abnormal bright spot target, calculating an initial risk assessment value according to the frequency of occurrence of the abnormal bright spot target; extracting the shape contour features and size change features of the real abnormal bright spot target; performing similarity grouping analysis on all real abnormal bright spot targets using a feature matching algorithm; judging whether there is a dynamic flickering bright spot target according to the results of the similarity grouping analysis; when it is determined that there is a dynamic flickering bright spot target, determining an adjustment factor based on the flickering mode features of the dynamic flickering bright spot target to modify the initial risk assessment value; and adjusting the subsequent dynamic tracking frame rate setting based on the modified risk assessment value; when determining the dynamic tracking frame rate according to the data size of the real-time video stream dataset, performing the following operations: comparing the data size with a preset low data size threshold and a preset high data size threshold, respectively; outputting the corresponding dynamic tracking frame rate configuration according to the comparison results; wherein the preset low data size threshold is lower than the preset high data size threshold; when the data size does not exceed the preset low data size threshold, setting the dynamic tracking frame rate to a low frame rate mode; when the data size exceeds the preset low data size threshold but does not exceed the preset high data size threshold, setting the dynamic tracking frame rate to a medium frame rate mode; when the data size exceeds the preset high data size threshold, setting the dynamic tracking frame rate to a high frame rate mode; the frame rate value corresponding to the low frame rate mode is lower than that corresponding to the medium frame rate mode, and the frame rate value corresponding to the medium frame rate mode is lower than that corresponding to the high frame rate mode. 2.The dynamic bright spot target statistics and analysis method based on visual tracking according to claim 1, wherein, when analyzing and processing the visual bright spot targets according to the detection reference threshold and the background model database, the following operations are performed: comparing the brightness value of each visual bright spot target in the real-time video stream with the detection reference threshold; and matching and checking the shape feature of each visual bright spot target with the standard shape feature in the background model database; identifying and marking the potential abnormal bright spot target according to the comparison and matching checking results; determining that the visual bright spot target is a potential abnormal bright spot target and marking it when the brightness value of the visual bright spot target exceeds the set proportion range of the detection reference threshold; determining that the visual bright spot target is a potential abnormal bright spot target and marking it when the shape feature of the visual bright spot target cannot be found in the background model database. 3.The dynamic bright spot target statistics and analysis method based on visual tracking according to claim 2, wherein, when judging whether the potential abnormal bright spot target is a real abnormal bright spot target based on the spatial position information and the motion trajectory feature, the following operations are performed: determining that the potential abnormal bright spot target is a real abnormal bright spot target when the spatial position information of the potential abnormal bright spot target shows that its motion trajectory deviates from the preset standard path; and determining that the potential abnormal bright spot target is a real abnormal bright spot target when the motion trajectory feature of the potential abnormal bright spot target indicates that its motion speed change exceeds the normal change range in the background model database.

4. The method of claim 3, wherein in calculating the initial risk assessment value according to the number of occurrences of the abnormal bright point target, the following operation is performed: counting the total number of occurrences of the real abnormal bright point target in the real-time video stream. calculating the size deviation amount by combining the size change feature of each real abnormal bright spot target with the reference size feature in the background model database; outputting the initial risk assessment value based on the total occurrence number and the size deviation amount; wherein, the size deviation amount reflects the relative change degree of the real abnormal bright spot target relative to the reference size feature. 5.The dynamic bright spot target statistics and analysis method based on visual tracking according to claim 4, wherein, when performing the similarity grouping analysis on all the real abnormal bright spot targets by using the feature matching algorithm, the following operations are performed: combining the shape contour feature and the size change feature of each real abnormal bright spot target into a feature vector representation; calculating the similarity score between all the feature vectors by using the feature matching algorithm; setting a similarity score threshold; identifying the feature vector group whose similarity score exceeds the similarity score threshold by using the feature matching algorithm; performing dynamic clustering processing on each feature vector group; judging whether there is a dynamic flickering bright spot target according to the dynamic clustering processing result; determining that the real abnormal bright spot targets in the feature vector group are dynamic flickering bright spot targets when the feature vector group contains at least two real abnormal bright spot targets and the shape contour features of the real abnormal bright spot targets show periodic changes.

6. The visual tracking based dynamic glint target statistics and analysis method of claim 5, wherein when the initial risk assessment value is modified by the adjustment factor determined based on the flickering pattern features of the dynamic flickering glint target, the following operations are performed: extracting flickering frequency features and flickering intensity features of the dynamic flickering glint target; performing similarity calculation between the flickering frequency features and flickering intensity features and reference flickering features in a historical modification record; selecting an adjustment factor based on the similarity calculation result; when a similarity between a reference flickering feature in the historical modification record and the current flickering frequency features and flickering intensity features exceeds a preset similarity threshold, using a corresponding reference adjustment factor in the historical modification record as the current adjustment factor; when the similarity between all reference flickering features does not exceed the preset similarity threshold, calculating the adjustment factor based on a total number of the dynamic flickering glint targets; and multiplying the initial risk assessment value by the adjustment factor to output a modified risk assessment value.

7. The visual tracking based dynamic glint target statistics and analysis method of claim 6, wherein when the adjustment factor is calculated based on the total number of the dynamic flickering glint targets, the following operations are performed: the size of the adjustment factor is in a positive proportional relationship with the total number of the dynamic flickering glint targets; a frame rate adjustment coefficient is calculated based on the size of the modified risk assessment value; the frame rate adjustment coefficient is in an inverse proportional relationship with the modified risk assessment value; and the current dynamic tracking frame rate is multiplied by the frame rate adjustment coefficient to output an adjusted subsequent dynamic tracking frame rate.

8. The dynamic bright spot target statistics and analysis method based on visual tracking according to claim 7, when adjusting the subsequent dynamic tracking frame rate setting based on the modified risk assessment value, the following operation is performed: obtaining the modified risk assessment value; The visual glint targets are re-detected and the background model database is updated. The flickering pattern features of the dynamic flickering glint targets are continuously monitored. When the flickering pattern features change exceeds a preset change threshold, the adjustment factor is recalculated and the risk assessment value is iteratively modified.

9. The visual tracking based dynamic highlight object statistics and analysis method of claim 8, after the adjusted subsequent dynamic tracking frame rate setting, performing the following operation: applying the adjusted subsequent dynamic tracking frame rate to a newly acquired real-time video stream data set; An updated risk assessment value and dynamic tracking frame rate configuration are output. ​ ​ ​

Images