AI-based video analysis and management system

Abstract

The application relates to the technical field of video analysis, in particular to an AI-based video analysis and management system, which comprises a channel evaluation module, a trajectory extraction module, a cross-source fusion module, a content screening module and a scheduling generation module.In the application, key frame optimization is realized by combining frame activity index evaluation in the object recognition process, trajectory continuity breakpoint repair is realized based on displacement sequence and direction change trend in the trajectory extraction process, trajectory matching and fusion are realized by region and direction cross mapping in cross-channel analysis, effective channels are identified based on trajectory retention features in content screening, and a structured recognition directory is constructed by combining label mapping in the scheduling process, so that the target recognition and trajectory splicing between different channels have a full-process dynamic completion and consistency checking mechanism, the cross-video source target recognition precision and stability can be effectively improved, and more efficient content screening and structured recognition can be realized in a multi-source video environment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of video analysis, in particular to an AI-based video analysis and management system. BACKGROUND

[0002] The technical field of video analysis belongs to the research category of the cross-application of artificial intelligence and computer vision, and the core includes the time sequence interpretation of image frames, target detection, tracking, behavior recognition, scene classification, event extraction and other contents, and has wide application in smart cities, security monitoring, traffic management, industrial manufacturing and other fields. Among them, the video analysis and management system refers to the software platform constructed for the collection, storage, analysis and scheduling of large-scale video data, mainly for effective classification and identification of video streams from monitoring devices, and realizing dynamic scheduling and on-demand management. Usually, through preset rules and manually set models, object recognition and behavior classification in video frames are carried out, and the analysis results are transmitted to the database for recording and management. Through polling, screening and distribution of video data, manual backtracking and retrieval operations are supported.

[0003] The existing video analysis and management technology mainly relies on polling and screening of monitoring videos, object detection and behavior classification through static rule setting. When facing a large number of video sources or a scene with frequent object movement, due to the inability to accurately identify the activity level of key frames, the resource scheduling efficiency is low. At the same time, in the case of object track interruption or cross-channel switching, there is a lack of track splicing capability, which is easy to cause tracking interruption and recognition misjudgment. In addition, the static recognition model is difficult to maintain recognition continuity in a variable scene, thereby limiting the application range of the video analysis system in a multi-source dynamic environment, reducing the accuracy of content screening and scheduling directionality, and affecting the overall operation efficiency and result reliability of the system. SUMMARY

[0004] The purpose of the present application is to solve the shortcomings in the prior art, and an AI-based video analysis and management system is proposed.

[0005] In order to achieve the above-mentioned purpose, the technical scheme adopted by the present application is as follows: the AI-based video analysis and management system comprises:

[0006] The channel evaluation module obtains the object number, thread occupation value and channel polling time period in the continuous frame samples in the video channel, calculates the frame activity index, selects the video channel number with the highest frame activity index into the main processing sequence, and generates the main channel recognition result;

[0007] The track extraction module extracts the frame sequence and displacement coordinate sequence of the same object number in the continuous frames based on the selected video channel number in the main channel recognition result, records the number of frame sequence continuity interruption frames, judges the direction stability, and generates the track continuity record;

[0008] The cross-source fusion module compares the region coordinates and direction paths corresponding to the object numbers in the same frame segment in other video channels at the same time point according to the frame number of all trajectory interruptions in the trajectory continuity record, performs matching and merging operations, and completes trajectory chain supplementation according to the frame index to generate a trajectory chain supplementation integration result;

[0009] The content screening module analyzes the frame time proportion and object appearance frequency change trend of the fusion trajectory in the original video according to the number of continuous frames and the number of fusion nodes of the merged trajectory sequence in the trajectory chain supplementation integration result, identifies the video channel numbers with consistent information, and generates channel effective identification data.

[0010] The scheduling generation module completes classification mapping according to the channel numbers and label information according to all channel numbers in the channel effective identification data, constructs a structure identification directory, and generates an AI video analysis management result.

[0011] As a further scheme of the application, the main channel identification result includes target active frame number, preferred video channel number, channel frame density index, and frame screening priority; the trajectory continuity record includes frame segment integrity label, direction change stability value, path change curve, and trajectory breakpoint identification; the trajectory chain supplementation integration result includes matching trajectory sequence, intersection region label, direction consistency label result, and frame index chain supplementation path; the channel effective identification data includes object fusion ratio, trajectory retention rate, frequency change node, and identification channel index; and the video analysis management result includes channel classification label, structure directory identification, identification data index, and video analysis data.

[0012] As a further scheme of the application, the channel evaluation module includes:

[0013] The object number extraction submodule obtains continuous frame samples in a video channel, analyzes object feature numbers in each frame image data, establishes an object number sequence in time sequence, counts the number of occurrences of each object number in continuous frames, and generates an object number repetition frequency matrix.

[0014] The thread occupation analysis submodule reads thread occupation values in a channel polling time period based on the object number repetition frequency matrix, compares thread resource allocation at different time points for different object numbers, extracts an object number response rate corresponding to a unit thread time, and obtains a thread object response ratio table.

[0015] The frame activity calculation submodule calculates the coupling degree between thread resource use structure and object repetition intensity according to the thread object response ratio table and the average object density of each frame in the polling time period, obtains an activity index of each frame, selects the video channel number corresponding to the frame with the largest activity index, and generates a main channel identification result.

[0016] As a further scheme of the present application, the trajectory extraction module comprises:

[0017] The frame sequence extraction submodule extracts the image frame index and displacement coordinate data corresponding to the same object number in the continuous frame image based on the video channel number screened in the main channel identification result, constructs a number-frame number correspondence mapping table, and records the coordinate change trajectory of each number, to generate an object trajectory coordinate sequence set;

[0018] The frame interruption marking submodule detects non-continuous items in the number frame index sequence according to the object trajectory coordinate sequence set, identifies a number segment with an arbitrary two-frame index difference exceeding 1, records the start and end frames, and inserts an empty frame mark at the breakpoint to obtain a trajectory interruption frame index table;

[0019] The direction stability judgment submodule calculates the direction change angle and displacement vector angle difference value according to the coordinate change direction angle and displacement trend value in the frames before and after the interruption segment according to the trajectory interruption frame index table, judges whether the difference value is less than the set direction change threshold, and generates a direction continuity mark for each interruption segment according to the judgment result, to establish a trajectory continuity record.

[0020] As a further scheme of the present application, the cross-source fusion module comprises:

[0021] The interruption frame extraction submodule obtains all trajectory numbers with non-continuous labels in the trajectory continuity record, extracts the start and end frame numbers corresponding to each number, and finds other channel frame numbers consistent with the interruption frame timestamp according to the channel mapping relationship, to establish a cross-channel time point mapping and generate an interruption frame channel synchronization table;

[0022] The cross matching discrimination submodule judges whether the area overlap area ratio between objects in two channels is greater than the area overlap threshold and whether the direction vector angle difference is less than the direction consistency threshold according to the region coordinates and motion direction vectors of the object numbers in the corresponding frame segment according to the interruption frame channel synchronization table, generates a Boolean mark after discrimination, and jointly judges the two with a cross threshold to obtain a cross-channel cross matching mark set;

[0023] The trajectory supplement chain integration submodule performs trajectory splicing at the corresponding frame index position based on the cross-channel cross matching mark set for the object numbers satisfying the matching judgment condition, merges and completes the missing frame segment data according to the number, updates the number-frame index mapping relationship after merging the numbers, and establishes a trajectory supplement chain integration result.

[0024] As a further scheme of the present application, the other channel frame sequence number consistent with the interrupt frame timestamp is searched according to the channel mapping relationship, specifically, in the other channel, a frame sequence number with an absolute value of difference from the interrupt frame timestamp less than a preset synchronization time window is searched.

[0025] As a further scheme of the present application, the area overlap area proportion between the two channel objects is judged, specifically, a ratio of an intersection area of the two object areas to a union area of the two object areas is calculated.

[0026] As a further scheme of the present application, the trajectory splicing is performed, specifically, the object trajectory data in the matched other channel, including the region coordinates and the motion direction vector, is copied and filled into the interrupt frame segment of the current trajectory.

[0027] As a further scheme of the present application, the content screening module comprises:

[0028] The trajectory frame proportion calculation submodule obtains all the merged trajectory sequences in the trajectory chain integration result, calculates the ratio of the continuous frame number to the total frame number of the video channel for each trajectory, simultaneously, the number of spliced trajectory nodes in the merging process is counted, the proportion in the whole trajectory is normalized to generate a fusion trajectory frame proportion index;

[0029] The frequency change discrimination submodule extracts the number of occurrences of each object in the merged trajectory in a unit time in a time sequence according to the fusion trajectory frame proportion index, differentiates the change amount and calculates the change slope, and marks as consistent in frequency under the condition that the slope is less than a frequency stability threshold, to obtain a set of object frequency change trend scalars;

[0030] The channel consistency comparison submodule compares the overlap section proportion of the trajectory segment length of the fusion trajectory in the channel and the original trajectory segment length of the main channel according to the set of object frequency change trend scalars, confirms as a matched channel under the condition that the overlap proportion is greater than a trajectory consistency threshold and the difference value of the frequency trend scalar is less than a stability threshold, and obtains channel effective recognition data.

[0031] As a further scheme of the present application, the content screening module comprises:

[0032] The trajectory frame proportion calculation submodule obtains all the merged trajectory sequences in the trajectory chain integration result, calculates the ratio of the continuous frame number to the total frame number of the video channel for each trajectory, simultaneously, the number of spliced trajectory nodes in the merging process is counted, the proportion in the whole trajectory is normalized to generate a fusion trajectory frame proportion index;

[0033] The frequency change discrimination sub-module extracts the number of occurrences of each object in the merged trajectory in a unit time in a time sequence according to the fusion trajectory frame proportion index, differentiates the change amount and calculates the change slope, and marks as consistent frequency under the condition that the slope is less than the frequency stability threshold, to obtain a set of object frequency change trend scalars;

[0034] The channel consistency comparison sub-module compares the overlapping section proportion of the length of the trajectory section in the channel where the fusion trajectory is located and the length of the original trajectory section in the main channel according to the set of object frequency change trend scalars, confirms as a matching channel under the condition that the overlapping proportion is greater than the trajectory consistency threshold and the difference value of the frequency trend scalar is less than the stability threshold, and obtains channel effective identification data.

[0035] As a further scheme of the application, the scheduling generation module comprises:

[0036] The channel classification mapping sub-module obtains all channel numbers in the channel effective identification data, extracts the object number, trajectory identifier and time period information corresponding to each channel number, performs channel attribution clustering according to the trajectory number and time period distribution law, constructs an attribution set according to the channel number, and generates a channel clustering mapping structure;

[0037] The label attribute integration sub-module retrieves the historical label information of the object number in each channel attribution set based on the channel clustering mapping structure, extracts the three-dimensional fields of class, behavior and region in the label field, and aggregates and constructs an attribute array according to the channel number to obtain a set of channel label aggregation vectors;

[0038] The structure directory generation sub-module assigns a corresponding structure path name field to each channel number according to the set of channel label aggregation vectors, combines the channel clustering label and time period identifier in the path structure, constructs a multi-channel structured directory tree and records it as an index item, and establishes an AI video analysis management result.

[0039] Compared with the prior art, the application has the following advantages and positive effects:

[0040] In the application, key frame optimization is realized by combining frame activity index evaluation in the object recognition process, trajectory continuity breakpoint repair is realized based on displacement sequence and direction change trend in the trajectory extraction process, trajectory matching fusion is realized through region and direction cross mapping in cross-channel analysis, effective channels are identified based on trajectory retention features in content screening, and a structured recognition directory is constructed by combining label mapping in the scheduling process, so that the target recognition and trajectory splicing between different channels have a full-process dynamic completion and consistency verification mechanism, which can effectively improve the target recognition accuracy and stability in a multi-source video environment, enhance the video content analysis coverage range and scheduling response efficiency, and further realize more efficient content screening and structured recognition in a multi-source video environment. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 System flowchart of the present application;

[0042] Figure 2 Channel evaluation module flowchart of the present application;

[0043] Figure 3 Trajectory extraction module flowchart of the present application;

[0044] Figure 4 Cross-source fusion module flowchart of the present application;

[0045] Figure 5 Content screening module flowchart of the present application;

[0046] Figure 6 Dispatch generation module flowchart of the present application. DETAILED DESCRIPTION

[0047] In order to make the objects, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0048] In the description of the present application, it should be understood that the terms "length", "width", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application. In addition, in the description of the present application, the meaning of "a plurality of" is two or more, unless otherwise explicitly and specifically limited.

[0049] Please refer to Figure 1 The AI-based video analysis and management system comprises:

[0050] The channel evaluation module obtains the object number, thread occupation value and channel polling time period in the continuous frame samples in the video channel, compares the object number repetition times and thread occupation value in proportion, combines the object density average of the channel in the polling time, calculates the frame active index, screens the video channel number with the highest frame active index into the main processing sequence, and generates the main channel identification result;

[0051] The trajectory extraction module extracts a frame sequence and a displacement coordinate sequence of the same object number in continuous frames based on the video channel number screened in the main channel identification result, records the number of frames with interrupted continuity of the frame sequence, judges the direction stability by combining the object direction angle change between the interrupted frames and the displacement path change trend between the frames, and generates a trajectory continuity record;

[0052] The cross-source fusion module compares the region coordinates and direction paths corresponding to the object numbers in the frame segment at the same time point in other video channels in sequence according to the interrupted frame numbers in the trajectory continuity record, performs a matching and merging operation on the basis that the region overlap mark and the direction consistency mark meet the cross threshold standard, completes the trajectory chain supplement according to the frame index, and generates a trajectory chain supplement integration result;

[0053] The content screening module analyzes the frame time proportion and object appearance frequency change trend of the fusion trajectory in the original video according to the continuous frame number and the fusion node number of the merged trajectory sequence in the trajectory chain supplement integration result, and performs sectional comparison with the original trajectory length of the object in the main channel, identifies the video channel number with consistent information, and generates channel effective identification data;

[0054] The scheduling generation module completes classification mapping according to the channel number and the label information according to all the channel numbers in the channel effective identification data, constructs a structure identification directory, and generates an AI video analysis management result.

[0055] The main channel identification result includes target active frame number, preferred video channel number, channel frame density index, frame screening priority, the trajectory continuity record includes frame segment integrity label, direction change stability value, path change curve, trajectory breakpoint identification, the trajectory chain supplement integration result includes matching trajectory sequence, cross region mark, direction consistency mark result, frame index chain supplement path, the channel effective identification data includes object fusion ratio, trajectory retention rate, frequency change node, identification channel index, and the video analysis management result includes channel classification label, structure directory identification, identification data index, and video analysis data.

[0056] Please refer to Figure 2 , the channel evaluation module comprises:

[0057] The object number extraction submodule obtains continuous frame samples in a video channel, analyzes the object feature numbers in each frame image data, establishes an object number sequence in time sequence, counts the number of times each object number appears in continuous frames, and generates an object number repetition frequency matrix;

[0058] To obtain the continuous frame samples in the video channel, the device interface needs to be accessed to access a specific monitoring system, for example, set the video frame sampling rate to 30 fps, and continuously extract 30 frame samples within 1 second. After extraction, the object feature number in the frame image data needs to be parsed, which can be assigned according to the target detection and recognition result. For example, two objects numbered 101 and 102 are detected in frame 1, and objects numbered 101 and 103 are detected in frame 2. The number sequence is [101, 102]→[101, 103]. Then the object number sequence in time sequence is established, and the number of occurrences of each number is counted frame by frame. The number frequency can be calculated by dividing the number of occurrences of the number in the continuous frame by the total number of frames. For example, number 101 appears 25 times in 30 frames, so its repetition frequency is 25 / 30=0.833. After counting the repetition frequencies of all object numbers, an object number repetition frequency matrix is constructed. The rows of the matrix represent different numbers, the columns represent frame numbers, and the elements are the number of occurrences of the corresponding number in the corresponding frame or its transformed form. Based on the preliminary analysis of the matrix, a number recording window needs to be set in the system, for example, set the window size to 10 frames, record the cumulative number of occurrences of each number in the window to form a window frequency statistical array, and then sort and exclude numbers with a frequency less than a certain threshold. Set the threshold to 3 times. If the number of occurrences of a number in 10 frames is less than 3, it is removed. The threshold can be set based on the number of times the general background objects in the scene occasionally appear. In actual testing, for example, in scene A, the average number of object number repetition frequencies is 0.76, and the standard deviation is 0.12. The threshold can be set to , that is, the number appears at least 6 times in 10 frames to be considered as a valid number. Such setting helps to exclude random noise interference, and finally obtains the object number repetition frequency matrix.

[0059] The thread occupation analysis submodule reads the thread occupation value in the channel polling time period based on the object number repetition frequency matrix, compares the thread resource allocation of different object numbers at different time points, extracts the object number response rate corresponding to the unit thread time, and obtains the thread object response ratio table.

[0060] Based on the object number repetition frequency matrix, the thread occupation value in the channel polling time period is extracted. The thread occupation value can be recorded by the multi-thread task scheduling module. The total amount of thread resources in a 1-second polling period is 100 units. Different amounts of resources are consumed by object processing tasks in each frame scheduling record. For example, number 101 occupies 10 units of resources in frame 1, and number 102 occupies 5 units of resources. The record thread occupation vector is [101:10, 102:5]. Then the resource occupation is accumulated according to the number dimension to form a total occupation table. At the same time, the resource allocation of each number is standardized according to the thread usage time length, and the number response rate in the unit thread time is calculated. Assuming that number 101 occupies 50 units of thread resources in 5 frames, and the average execution time of its task is 2 seconds, then its thread response rate is This indicates an average resource requirement of 25 units per second. This type of calculation is performed on all numbers to form a thread response rate table. To avoid outliers affecting the calculation, a response rate upper limit threshold needs to be set, which is set to 80% of the system's maximum schedulable thread resources. That is, if the maximum number of threads is 100 units, the response rate upper limit is set to 80. If the actual response rate of a number exceeds this value, it is considered a scheduling anomaly and needs to be marked as an anomaly number. After being removed, it will participate in subsequent activity calculations. Finally, a thread usage efficiency index table for each number within a time period is formed. The table structure is a two-dimensional array, with rows corresponding to frames, columns corresponding to numbers, and elements representing thread response rates. Then, the thread response rates at the number level are compared by combining the number frequency matrix to generate a thread object response ratio table.

[0061] The frame activity calculation submodule calculates the coupling degree between thread resource usage structure and object repetition intensity based on the thread object response ratio table and the average density of each frame object within the polling period, using the following formula:

[0062] ;

[0063] Calculate and obtain the activity index for each frame. The video channel number corresponding to the frame with the highest activity index is selected to generate the main channel recognition result; among which, Representing the Frame activity index, Representing the The first frame The number of times an object is repeated. Representing the The first frame The thread utilization rate of each object. Representing the The normalized object density of a frame (defined as) , The number of objects in this frame. (Maximum number of objects across all frames) Representative object Average repetition intensity within the polling period The total number of objects per frame;

[0064] The thread object response ratio table is invoked. Based on the average object density of each frame within the polling time period, the coupling degree between the thread resource usage structure and the object repetition intensity is calculated. First, the frame object density is defined as the number of detected object numbers in a unit frame. For example, if frame 1 detects numbers 101, 102, and 103, the density is 3. The density of all frames is recorded to form an array. To unify data scale, normalization processing is required, including normalized density. ,in The maximum value in the array, for example , ,but This processing causes the density of different frames to fluctuate between 0 and 1, and then for each frame... , based on the intra-frame number Use an index to extract the number of times it repeats in the frequency matrix. Thread response rate in the response ratio table and combined with average repetition intensity Calculate the frame activity index If there are 3 objects numbered 101, 102, and 103 in a certain frame, their corresponding parameter values ​​are shown in Table 1.

[0065] Table 1 Intra-frame object activity parameters

[0066]

[0067] Table 1 lists the values ​​of the intra-frame numbering parameter, assuming the normalized density of the frame is . The intermediate values ​​of each item are calculated as follows:

[0068] For number 101: ;

[0069] For number 102: ;

[0070] For number 103: ;

[0071] Taking the absolute values, they are 1.13, 0.7, and 0 respectively, with a sum of 1.83. The average is then calculated as follows: .

[0072] This value indicates that the activity index of this frame is 0.61. It is compared with the calculated values ​​of other frames, and the frame number corresponding to the highest activity is selected as the main processing target to finally obtain the main channel recognition result.

[0073] The formula aims to integrate a frame-level activity evaluation metric by calculating resource activity offsets for each object. The formula first considers each frame... All object numbers Perform iterative calculations, with the starting point being the term. This multiplication operation will number the objects. In frame Number of repetitions in Its thread occupancy response rate These parameters are combined to form a joint parameter that reflects the unit response intensity of the object within the current frame, taking into account both its frequency of occurrence and its thread resource usage intensity. This value is then divided by the normalized object density of the current frame. i.e. The division operation is used to correct the structural difference of the number of objects in different frames, prevent the activity from being amplified or compressed in the object dense frame because of the higher overall resource allocation, and has the meaning of normalization. Next, the density adjusted value is subtracted from the average repetition intensity of the object in the polling period to measure whether the behavior of the object in the current frame deviates from its overall time series performance, that is, to evaluate whether the instantaneous behavior is abnormal. The greater the deviation, the stronger the activity of the object in the current frame. After the subtraction operation, the absolute value is taken, which aims to eliminate the influence of the deviation direction and only keep the offset amplitude to ensure that the deviation of all objects is considered, whether it is enhanced or weakened. Finally, the sum of all object numbers is summed and divided by the total number of objects to form the average offset amplitude of the number of objects in the current frame, that is, the activity index of the frame The index can be used for cross-frame comparison by selecting the maximum corresponding frame number for identification processing. Therefore, the whole set of operation logic realizes the fusion of the activity intensity of the object level by multiplication, the density normalization of the frame level by division, the comparison of the time series performance by subtraction, the elimination of the influence of the deviation direction by absolute value, and the extraction of the overall offset features of the frame by summation and average. Thus, a frame-level activity index system with strict structure, clear hierarchy and quantifiable is formed.

[0074] Please refer to Figure 3 The trajectory extraction module includes:

[0075] The frame sequence extraction submodule extracts the image frame index and displacement coordinate data corresponding to the same object number in the continuous frame image based on the video channel number filtered in the main channel identification result, constructs a number and frame number correspondence mapping table, and records the coordinate change trajectory of each number to generate an object trajectory coordinate sequence set;

[0076] The video channel number filtered in the main channel recognition result is obtained, and the continuous frame image information corresponding to the number is extracted. Each frame image is composed of object numbers and two-dimensional spatial coordinates output by the object detection module. There can be multiple object numbers in each frame. Only the object with the same number as the previous frame is extracted as the matching result. Then a mapping structure of the number and the frame sequence number is established, and the frame sequence number of each number in the video is one-to-one corresponding to its coordinate value. For example, the number 201 appears in the 3rd, 4th and 5th frames, and its corresponding coordinates are (120, 150), (124, 154) and (127, 159). The generated frame sequence is [3, 4, 5], and the coordinate sequence is [(120, 150), (124, 154), (127, 159)]. The extraction action is grouped and processed according to the number. If there are multiple object numbers in the same frame, their coordinates and frame indexes are extracted respectively. In this process, the original detection data is filtered, and only the number recognition results with a confidence higher than 0.7 are retained. The confidence value is taken from the score field in the object detection output. The threshold is set because when the confidence is lower than 0.7, the recognition accuracy is lower than 85% in the system evaluation, which cannot guarantee the spatial continuity. Based on this, data cleaning is performed to remove noise objects, and the frame index sequence and two-dimensional coordinate trajectory corresponding to each object number are formed. The object number index is aggregated into an array structure, and a data set that can be used for subsequent continuity analysis is constructed, and an object trajectory coordinate sequence set is generated, as shown in Table 2:

[0077] Table 2 Object number frame sequence and coordinate list

[0078]

[0079] As shown in Table 2, object numbers 201, 305 and 402 are continuously detected in their corresponding frame segments, and coordinate sequences and confidence data are generated to construct the trajectory structure.

[0080] The frame interruption marking submodule detects non-continuous items in the number frame index sequence according to the object trajectory coordinate sequence set, identifies any two frame index difference values exceeding 1, records the start and end frames, and inserts an empty frame mark at the breakpoint to obtain a trajectory interruption frame index table.

[0081] According to the object trajectory coordinate sequence set, the continuity in the frame index sequence corresponding to each object number is detected, each number corresponding frame sequence is traversed, and the paragraph with adjacent frame sequence number difference not being 1 is judged. If the frame sequence corresponding to the number 301 is [8, 9, 10, 13, 14, 15], there is a frame interruption between the 10th frame and the 13th frame, because 13-10=3, the difference is greater than 1, which is an interrupted frame segment, the starting frame is recorded as 10, the ending frame is recorded as 13, the number is recorded as 301. The operation is realized through a frame sequence difference array, that is, the difference between the current frame number and the next frame number is calculated. If the difference is greater than 1, it is a breakpoint. The operation is performed on all numbers, all interrupted paragraphs are recorded, and then an empty frame identifier is inserted into the corresponding trajectory data. The placeholder data structure uses the form (x: NaN, y: NaN), which ensures that the trajectory sequence length is aligned with the time axis. The empty frame identifier is only used to indicate that the target is temporarily missing, not a detection failure, to avoid structural misplacement due to missing data in subsequent analysis. At the same time, the number, breakpoint position, frame number range and other meta information of each interrupted segment are recorded to form a structured annotation file, and the trajectory interrupted frame index table is obtained, as shown in Table 3:

[0082] Table 3: Trajectory interrupted frame segment identifier table

[0083]

[0084] As shown in Table 3, through the interrupted frame analysis operation, the system identifies the interrupted segments of numbers 301 and 402, which are further used for direction judgment logic.

[0085] The direction stability judgment submodule calculates the difference between the direction change angle and the displacement vector angle according to the coordinate change direction angle and the displacement trend value in the frames before and after the interrupted segment according to the trajectory interrupted frame index table, judges whether the difference is less than the set direction change threshold, and generates a direction continuity mark for each interrupted segment according to the judgment result, and establishes a trajectory continuity record;

[0086] The trajectory interrupted frame index table is called, the start and end frame numbers and object numbers of each interrupted segment are read, for each interrupted segment, the coordinate values of the number in the frame before and after the interruption are extracted and the direction angle is calculated. The direction angle is calculated by using arctangent operation to obtain the angle value of the displacement vector. For example, the frame before the interruption is the 10th frame, the coordinates of the number 501 are (100, 200), the frame after the interruption is the 13th frame, the coordinates are (120, 220), the displacement vector is (20, 20), and the direction angle is The same operation is performed on all interruption segments, and then the included angle change between the front and rear frame direction angles is calculated. If the number 601 is 80 degrees and 94 degrees before and after the frame 15, the included angle change is 14 degrees. It is judged whether it is less than the direction change threshold of 15 degrees. The threshold is derived from the system stability evaluation test. When the direction angle change exceeds 15 degrees, the target motion direction consistency is significantly disturbed, and it is not suitable to determine as a continuous trajectory. Therefore, 15 degrees is set as the upper limit of the direction continuity judgment. If the change angle is less than the threshold, a "continuous" label is marked in the record. Otherwise, it is marked as "non-continuous". The judgment result is added to the original interruption frame index table with a Boolean identifier. The trajectory continuity record is established, as shown in Table 4:

[0087] Table 4 interruption segment direction angle change and continuity judgment table

[0088]

[0089] As shown in Table 4, only the interruption segment of number 702 has a direction angle change exceeding the threshold and is marked as "non-continuous". The remaining segments are judged as continuous trajectory segments.

[0090] Please refer to Figure 4 , the cross-source fusion module comprises:

[0091] The interruption frame extraction submodule obtains all trajectory numbers with non-continuous labels in the trajectory continuity record, extracts the start frame and end frame sequence numbers corresponding to each number, and finds other channel frame sequence numbers consistent with the interruption frame timestamp according to the channel mapping relationship. Cross-channel time point mapping is established, and an interruption frame channel synchronization table is generated;

[0092] All trajectory numbers marked as "non-continuous" in the trajectory continuity record are obtained, the starting frame and ending frame indexes of the corresponding interruption section are extracted in sequence, and the video channel identifier to which each trajectory belongs is matched according to the frame number. The global timestamp where the frame is located is extracted, and then a one-to-one correspondence between the timestamp and the channel frame index is established. On this basis, the frames of other video channels that are consistent with the time point of the main channel interruption frame are located, the frame section of these channels at the corresponding moment is obtained through index comparison, and then a cross-channel time synchronization structure mapping array is established. The array item structure is [main channel number, interruption frame timestamp, auxiliary channel number, corresponding frame index]. For example, if the interruption frame of main channel 1 is frame 58, and the timestamp is 12.35 seconds, and if the timestamp of frame 62 of channel 3 is also 12.35 seconds, then a group of synchronized frame sections is recorded. This structure can be synchronized by judging whether the absolute value of the timestamp difference is less than the synchronization tolerance threshold. The system sets the synchronization tolerance to 0.03 seconds. That is, if the time difference is less than this value, it is considered to be the same moment frame. For example, the difference between 12.35 seconds and 12.36 seconds is 0.01 seconds, which is less than the threshold value, and meets the synchronization requirement. The matching pair is recorded. After traversing all the timestamps of the interruption trajectories, the corresponding frame index mapping table is established in all video channels. The channel number, frame number, and timestamp triplets of the matching pair are output in a structured manner, and the interruption frame channel synchronization table is generated, as shown in Table 5:

[0093] Table 5: Interruption frame channel synchronization table

[0094]

[0095] As shown in Table 5, interruption frames 58, 102, and 145 in channels 3, 2, and 4 respectively find synchronized frames with consistent timestamps, and a synchronization mapping relationship is established.

[0096] The cross-matching discrimination submodule judges whether the area overlap area ratio between the objects in two channels is greater than the area overlap threshold value and whether the direction vector angle difference is less than the direction consistency threshold value according to the area coordinates and motion direction vectors of the objects in the corresponding frame section based on the interruption frame channel synchronization table. Boolean identifiers are generated respectively, and then the two are jointly judged with the cross threshold value to obtain a cross-channel cross-matching marking set.

[0097] The interrupt frame channel synchronization table is called, each set of synchronization frame segments is traversed in turn, and the coordinate area and motion direction vector information of all object numbers in the synchronization frame are extracted. First, the two-dimensional bounding box coordinates of the object number are extracted as an index, and the area overlap is calculated. The intersection area and union area of the bounding box of the object number in the main channel interrupt frame are operated, and the IoU value is obtained as the area overlap index. If the number A is in the main channel 58th frame area (100, 100)-(150, 150), and the number B is in the auxiliary channel 62nd frame area (110, 110)-(155, 160), the IoU is 0.64, which is greater than the area overlap threshold 0.6, and it is judged that the area is matched. The area threshold is set according to the experimental statistical results of the channel camera view angle difference, and the minimum effective overlap value is set to 0.6 to avoid false matching. Then, the motion direction vector of the number is judged. The coordinate difference of the nearest two frames is taken as the direction vector, and the included angle between the two direction vectors is calculated. For example, vector A is (5, 4), and vector B is (6, 3). The included angle calculation formula is , the calculated angle is 17.3 degrees, which is less than the direction consistency threshold 20 degrees, and it is determined that the directions are consistent. The direction angle threshold is set based on the evaluation of the multi-channel direction fluctuation range in the system tracking experiment of the stable moving object. The maximum tolerance angle is set to 20 degrees. If both conditions are met, the area overlap flag is 1, and the direction consistency flag is 1. Finally, the logical AND result of the two is taken as the cross matching result Boolean flag. If both are 1, it is matched, otherwise it is not matched. The matching state of all object numbers in all synchronization frame segments is recorded in a structured manner, and the cross-channel cross matching flag set is obtained.

[0098] The trajectory chain integration submodule is based on the cross-channel cross matching flag set. For the object number that meets the matching condition, the trajectory splicing is performed at the corresponding frame index position, the missing frame segment data is completed by merging according to the number, the number after merging is updated, and the number and frame index mapping relationship is regenerated to establish the trajectory chain integration result.

[0099] Based on the cross-channel cross-matching tag set, trajectory splicing is performed on all object number combinations that meet the matching conditions. First, the start and end frames of the interrupted trajectory in the main channel are located and marked as trajectory number T1. The continuous trajectory segments of the matched object number in the auxiliary channel are extracted. Assuming the segment number is T2, the start frame is F1, and the end frame is F2, all frame numbers and coordinate information in the T2 trajectory segment are inserted between the interrupted segments of the T1 trajectory. The frame numbers are arranged in timestamp order and the trajectory sequence is reconstructed to complete the trajectory chain completion operation. After trajectory completion, the original T1 and T2 trajectory numbers are merged. The main channel number priority principle is set, the T1 number is retained, the T2 number is canceled and the trajectory-number mapping table is updated. This mapping structure updates all identifier records containing the T2 number to T1 to ensure the consistency of trajectory numbers in subsequent tracking. This operation process is executed one by one according to the interrupted trajectory number. Finally, a set of updated trajectory list structures after trajectory chain completion is generated. Each trajectory in the structure contains fields such as start frame, end frame, coordinate sequence, merged record and trajectory status label, and establishes the trajectory chain integration result.

[0100] Please see Figure 5 The content filtering module includes:

[0101] The trajectory frame proportion calculation submodule obtains all merged trajectory sequences in the trajectory chain integration result, calculates the ratio of the number of continuous frames of each trajectory to the total number of frames of the video channel to which it belongs, and counts the number of trajectory nodes spliced ​​during the merging process. The proportion of these nodes in the entire trajectory is normalized to generate the fused trajectory frame proportion index.

[0102] For each merged trajectory in the trajectory chain integration result, extract its corresponding video channel number, trajectory start and end frame numbers, and total number of frames in the channel, and calculate the trajectory duration in frames. Total number of frames per channel The ratio of the percentage of frames obtained If a trajectory starts with frame 120, ends with frame 165, and has a total of 1000 frames, then Then, the number of fusion nodes included in the trajectory was counted. The fusion node is defined as a newly added frame index node during the cross-source matching and merging process, and its number is recorded and normalized to the proportion of trajectory length. If a trajectory has a length of 46 frames and the number of fused nodes is 9, then The normalized fusion degree reflects the proportion of the fused trajectory to the overall trajectory. The system sets the anomaly threshold for the proportion of fused nodes to be 0.5, that is, when... When the trajectory is considered to be over-fused, it is excluded from the screening process. After traversing all trajectories, a structured record containing the trajectory number, duration frame number, frame proportion, fusion node number, and normalized fusion ratio is constructed. The trajectory-based indicators for frequency trend judgment are summarized, and the fusion trajectory frame proportion indicators are generated, as shown in Table 6:

[0103] Table 6 Fusion trajectory frame proportion indicator table

[0104]

[0105] As shown in Table 6, the T01 and T03 trajectory fusion proportions are low, and can be included in effective trajectory analysis. T02 is excluded due to the node fusion ratio exceeding the threshold.

[0106] The frequency change discrimination submodule extracts the number of occurrences of each object in the merged trajectory in a unit of time in the time sequence according to the fusion trajectory frame proportion indicator. The change value is differentiated and the change slope is calculated. If the slope is less than the frequency stability threshold, it is marked as consistent frequency, and the object frequency change trend scalar set is obtained.

[0107] The fusion trajectory frame proportion indicator is called to analyze the frequency change of the object number of each available trajectory. The number of occurrences of the object in each second in the frame segment corresponding to the trajectory is extracted, and the frequency time sequence is constructed , where each is the frame number of the occurrence of the trajectory object in the second, such as object A appearing 23, 25, and 27 frames in the first, second, and third seconds, respectively. The frequency sequence is . The difference value is calculated for the sequence, and the linear slope is calculated. After processing using the least squares method, the slope is obtained. The growth rate per unit second is normalized to , where . If , then . The system sets the frequency consistency judgment threshold to 0.05. If , it means that the frequency of the object in the merged trajectory fluctuates little, and it is identified as a frequency stable object. If , it is a frequency changing object and is not included in the next consistency comparison. The frequency change slope threshold of 0.05 is derived from the 95% confidence interval result obtained in the multi-scene trajectory fluctuation analysis. After calculating the frequency change rate of the object in all trajectories, a structure is recorded, which contains the trajectory number, object number, average frequency, frequency slope, and frequency stability label. The object frequency change trend scalar set is obtained.

[0108] The channel consistency comparison submodule compares the proportion of the overlapping section of the length of the track segment of the fusion track in the channel and the length of the original track segment of the main channel according to the object frequency change trend scalar set, and confirms the matching channel under the condition that the overlapping proportion is greater than the track consistency threshold and the frequency trend scalar difference is less than the stability threshold, and obtains channel effective identification data;

[0109] According to the object frequency change trend scalar set, the track number with a frequency stability label of "stable" is read, the original track segment length of the same object number in the fusion channel and the main channel is extracted, and the section overlap ratio is calculated. The overlap ratio , wherein is the number of overlapping frames of two track segments. For example, if the fusion track frame segment is [100, 140], the original track of the main channel is [120, 155], and the intersection is [120, 140] with 21 frames, , the system sets the track consistency threshold to 0.7. If the overlap ratio , the track timing consistency requirement is met, and the frequency slope of the object in the main channel and the frequency slope of the fusion track are extracted. If the difference between them is , it is considered that the object behaves consistently in the main and auxiliary channels, and the frequency stability condition is met. If both conditions are met, the corresponding channel number is added to the effective identification channel list. The judgment operation is performed in parallel for all objects and tracks. Finally, the effective channel numbers are integrated, the channel number, track number, object number, overlap ratio, frequency difference, and channel consistency determination result are structured and recorded, and channel effective identification data is obtained.

[0110] Please refer to Figure 6 , the scheduling generation module includes:

[0111] The channel classification mapping submodule obtains all channel numbers in the channel effective identification data, extracts the object number, track identifier and time period information corresponding to each channel number, performs channel attribution clustering according to the track quantity and period distribution law, and constructs an attribution set according to the channel number to generate a channel clustering mapping structure.

[0112] When all channel numbers in the valid recognition data of the acquisition channel are obtained, the channel number field in the fusion result table is first read, and the number set is formed by extracting the number of each record with a valid recognition of 1 and then extracting the object number and track identification associated with the number. The track identification is generally represented by continuous frame index, such as "TID_083". In the track sequence corresponding to the recognition object "A01", the frame segment range of its appearance is counted in time sequence to obtain the corresponding start frame and end frame, and the frame segment is mapped to the corresponding time period information combined with the video frame rate. The time period accuracy unit is set to seconds. For a video with a frame rate of 25 fps, the frame segment [225, 325] can be mapped to the time period [9.00 s, 13.00 s]. Within the channel number range of all object tracks, it is necessary to determine whether the number of tracks under a certain channel number meets the minimum clustering threshold. The minimum clustering threshold is defined as , and the value is set to 4 according to the experiment, that is, each channel must have at least 4 valid tracks, otherwise it is not included in the clustering. The track number is calculated by counting the object numbers under each number without repetition. The track period density judgment process is to construct a distribution curve with time as the horizontal axis and track number as the vertical axis, and to evaluate the time period aggregation of each channel number. If the track is concentrated in a certain time period with a long time span, the channel is identified as "time period concentrated type". If the track is evenly distributed on the time axis, it is "full time period diffusion type". Taking the actual channel number "C002" as an example, it contains 6 object tracks, the time span covers 10s to 50s, and the track is mainly concentrated in the interval of 20s to 30s, which can be marked as "time period concentrated type". After obtaining the above channel clustering label, a channel clustering label dictionary is constructed, and each record corresponds to the fields of channel number, object number array, track intensive type label, and track number. Finally, the dictionary items are arranged according to the channel number to obtain the channel clustering mapping structure.

[0113] Table 7 Channel number example distribution table

[0114]

[0115] As shown in Table 7, the track intensive type is related to the time period distribution, and the track number and the minimum clustering threshold jointly affect the screening of the attribution set. The channel clustering mapping structure records the track attribution characteristics under each channel number.

[0116] The label attribute integration sub-module retrieves the historical label information of the object number in each channel attribution set based on the channel clustering mapping structure, extracts the class, behavior, and region fields in the label field, and aggregates and constructs the attribute array according to the channel number to obtain the channel label aggregation vector set.

[0117] When the channel clustering mapping structure is called, the channel number and object number array are extracted from each clustering mapping, and the corresponding record in the label library is found for each object number. The label library field is set as object number, category label, behavior label, and area label. The label information of object "B011" is as follows: category = personnel, behavior = walking, and area = doorway area. The label dimension is divided into three categories. Each category needs to be standardized. The category field has five categories, namely "personnel", "vehicle", "non-static object", "equipment", and "pet". The behavior field has eight categories, and the area field has ten categories. If the label information of an object number is not found during the calling process, it is considered as a non-labeled item, and its label is set as "undefined". The label code uses an integer array to represent, for example, the category label "personnel" is coded as 1, the behavior label "walking" is coded as 3, and the area label "doorway area" is coded as 7. The corresponding label vector of the object is [1, 3, 7]. The label vectors of object numbers "B011", "B012", and "B016" under channel number "C002" are [1, 3, 7], [2, 4, 9], and [1, 5, 2] respectively. The label vectors under each channel number are aggregated. The aggregation method is to take the average of each dimension. The calculation process is as follows: category dimension: (1+2+1) / 3=1.33, behavior dimension: (3+4+5) / 3=4.00, and area dimension: (7+9+2) / 3=6.00. The aggregation result is rounded to two decimal places and recorded as an attribute vector. The operation is repeated for all channel numbers, and the final channel label aggregation vector set is obtained. The vector set structure is that each channel number corresponds to an attribute array record.

[0118] The structure directory generation submodule assigns a corresponding structure path name field to each channel number based on the channel label aggregation vector set, combines the channel clustering label and time period identifier in the path structure, constructs a multi-channel structured directory tree and records it as an index item, and establishes an AI video analysis management result.

[0119] According to the channel label aggregation vector set, a structure path field is constructed for each channel number, and the field construction rule is "channel number_category label_behavior label_region label_time period identification", wherein the category, behavior and region labels are obtained by taking the integer of the aggregation vector, such as taking the integer of the vector [1.33, 4.00, 6.00] as [1, 4, 6], and the identification meanings are "personnel_ standing_passage area" respectively, the channel number is set as "C002", and the time period identification is generated by the earliest start frame and the latest end frame in the channel clustering mapping structure, for example, the frame segment [225, 825] corresponds to the time [9.00s, 33.00s], and the path structure is "C002_personnel_ standing_passage area_9-33s", all the path structures will form a structured directory tree, the root node of the directory tree is the video channel number, the first level child node is the channel number, the second level node is the path structure name, the leaf node records the label vector and the channel clustering label dictionary corresponding item number as index mapping, and finally the structure tree is written into a JSON format index file, imported into a system retrieval module for query and call, and an AI video analysis management result is generated.

[0120] The above is only a preferred embodiment of the present application, and does not limit the present application in other forms. Any skilled person in the art can use the disclosed technical content to make changes or modifications to equivalent embodiments applied to other fields, but any simple modification, equivalent change and modification made according to the technical essence of the present application to the above embodiments without departing from the technical solution content of the present application still belongs to the protection scope of the present application technical solution.

Claims

1. An AI-based video analysis and management system, characterized by, The system comprises: The channel evaluation module obtains object numbers, thread occupation values and channel polling time periods in continuous frame samples in a video channel, calculates frame activity indexes, screens video channel numbers with the highest frame activity indexes into a main processing sequence, and generates a main channel identification result; The trajectory extraction module extracts frame sequences and displacement coordinate sequences of the same object numbers in continuous frames based on the video channel numbers screened in the main channel identification result, records the numbers of frame sequences with continuity interruption, judges direction stability, and generates trajectory continuity records; The cross-source fusion module compares region coordinates and direction paths corresponding to object numbers in frame segments at the same time point in other video channels according to all trajectory interruption frame numbers in the trajectory continuity records, performs matching and merging operations, and completes trajectory chain supplementation according to frame indexes, to generate a trajectory chain supplementation integration result; The content screening module analyzes frame time proportion and object appearance frequency change trends of fusion trajectories in original videos according to the number of continuous frames and the number of fusion nodes of the merged trajectory sequences in the trajectory chain supplementation integration result, identifies video channel numbers with consistent information, and generates channel effective identification data; The scheduling generation module completes classification mapping according to channel numbers and label information according to all channel numbers in the channel effective identification data, constructs a structure identification directory, and generates an AI video analysis management result. 2.The AI-based video analysis and management system of claim 1, wherein, The main channel identification result comprises target active frame numbers, preferred video channel numbers, channel frame density indicators and frame screening priorities, the trajectory continuity record comprises frame segment integrity labels, direction change stability values, path change curves and trajectory breakpoint identifiers, the trajectory chain supplementation integration result comprises matching trajectory sequences, intersection region markers, direction consistency marker results and frame index chain supplementation paths, the channel effective identification data comprises object fusion ratio, trajectory retention rate, frequency change node and identification channel index, and the video analysis management result comprises channel classification label, structure directory identifier, identification data index and video analysis data. 3.The AI-based video analysis and management system of claim 1, wherein, The channel evaluation module comprises: The object number extraction submodule obtains continuous frame samples in a video channel, analyzes object feature numbers in each frame image data, establishes an object number sequence in time sequence, counts the number of occurrences of each object number in continuous frames, and generates an object number repetition frequency matrix; The thread occupation analysis submodule reads thread occupation values in a channel polling time period based on the object number repetition frequency matrix, compares thread resource allocation of different object numbers at different time points, extracts object number response rates corresponding to unit thread time, and obtains a thread object response ratio table; The frame activity calculation submodule calculates the coupling degree between thread resource use structure and object repetition intensity according to the thread object response ratio table and the average object density of each frame in the polling time period, obtains the activity indexes of each frame, screens the video channel number corresponding to the frame with the largest activity index, and generates a main channel identification result. 4.The AI-based video analysis and management system of claim 1, wherein, The trajectory extraction module comprises: The frame sequence extraction submodule extracts image frame indexes and displacement coordinate data corresponding to the same object number in continuous frame images based on the video channel number screened in the main channel identification result, constructs a number-frame number correspondence mapping table, records the coordinate change trajectory of each number, and generates an object trajectory coordinate sequence set; The frame interruption marking submodule detects non-continuous items in the number frame index sequence according to the object trajectory coordinate sequence set, identifies a number segment with a difference of more than 1 between any two frame indexes, records the start and end frames, and inserts an empty frame mark at the breakpoint to obtain a trajectory interruption frame index table; The direction stability judgment submodule calculates the direction change angle and the displacement vector angle difference value according to the coordinate change direction angle and the displacement trend value in the frames before and after the interruption segment according to the trajectory interruption frame index table, judges whether the difference value is less than a set direction change threshold, and generates a direction continuity mark for each interruption segment according to the judgment result to establish a trajectory continuity record. 5.The AI-based video analysis and management system according to claim 1, wherein, The cross-source fusion module includes: The interruption frame extraction submodule obtains all trajectory numbers with non-continuous labels in the trajectory continuity record, extracts the start frame and end frame number corresponding to each number, and finds other channel frame numbers consistent with the interruption frame timestamp according to the channel mapping relationship to establish a cross-channel time point mapping and generate an interruption frame channel synchronization table; The cross matching discrimination submodule judges whether the area overlap area ratio between two channel objects is greater than an area overlap threshold and whether the direction vector angle difference is less than a direction consistency threshold according to the region coordinates and motion direction vectors of the object number in the corresponding frame segment according to the interruption frame channel synchronization table, respectively generates a Boolean mark of whether the area overlap area ratio is greater than the area overlap threshold and a Boolean mark of whether the direction vector angle difference is less than the direction consistency threshold, and obtains a cross-channel cross matching mark set; The trajectory supplement integration submodule performs trajectory splicing at the corresponding frame index position based on the cross-channel cross matching mark set for the object number satisfying the matching judgment condition, merges and completes the missing frame segment data according to the number, updates the number-frame index mapping relationship after merging the number, and establishes a trajectory supplement integration result. 6.The AI-based video analysis and management system according to claim 5, wherein, The channel frame number consistent with the interruption frame timestamp is found according to the channel mapping relationship, specifically, in the channel, the frame number with an absolute value of the difference from the interruption frame timestamp less than a preset synchronization time window is searched. 7.The AI-based video analysis and management system of claim 5, wherein, The area overlap area ratio between two channel objects is calculated, specifically, the ratio of the intersection area of the two object regions to the union area of the two object regions. 8.The AI-based video analysis and management system of claim 5, wherein, The trajectory splicing is performed, specifically, the object trajectory data in the channel satisfying the matching judgment condition, including the region coordinates and the motion direction vector, is copied and filled into the interruption frame segment of the current trajectory. 9.The AI-based video analysis and management system of claim 1, wherein, The content screening module includes: The trajectory frame proportion calculation submodule obtains all merged trajectory sequences in the trajectory supplement chain integration result, calculates the ratio of the continuous frame number to the total frame number of the video channel to which each trajectory belongs, simultaneously counts the number of trajectory nodes spliced in the merging process, normalizes the proportion in the full trajectory, and generates a fusion trajectory frame proportion index; The frequency change discrimination submodule extracts the number of occurrences of each object in the merged trajectory in a unit of time in the time sequence according to the fusion trajectory frame proportion index, differentiates the change amount and calculates the change slope, and marks the frequency as consistent under the condition that the slope is less than the frequency stability threshold, to obtain a set of object frequency change trend scalars; The channel consistency comparison submodule compares the overlapping section proportion of the length of the trajectory segment in the channel where the fusion trajectory is located and the length of the original trajectory segment in the main channel according to the set of object frequency change trend scalars, confirms the matching channel under the condition that the overlapping proportion is greater than the trajectory consistency threshold and the difference between the frequency trend scalars is less than the frequency stability threshold, and obtains channel effective identification data. 10.The AI-based video analysis and management system of claim 1, wherein, The scheduling generation module includes: The channel classification mapping submodule obtains all channel numbers in the channel effective identification data, extracts the object number, trajectory identifier, and time period information corresponding to each channel number, classifies the channels according to the number of trajectories and the time period distribution law, constructs an attribution set according to the channel number, and generates a channel classification mapping structure; The label attribute integration submodule retrieves the historical label information of the object number in each channel attribution set based on the channel classification mapping structure, extracts the class, behavior, and region three-dimensional fields in the label field, and aggregates and constructs an attribute array according to the channel number, to obtain a channel label aggregation vector set; The structure directory generation submodule assigns a corresponding structure path name field to each channel number according to the channel label aggregation vector set, combines the channel classification label and the time period identifier in the path structure, constructs a multi-channel structured directory tree and records it as an index item, and establishes an AI video analysis management result.

Images