Multi-strategy adaptive screenshot scheduling system for video ai algorithm verification

Abstract

The application relates to the technical field of data processing, and discloses a multi-strategy screenshot scheduling system for video AI algorithm verification. In the system, a configuration manager binds a camera, a preset position and a corresponding AI algorithm to form a triple; a timing strategy module generates a screenshot request based on a timing interval; a quantitative strategy module generates a screenshot request based on a periodic working time period and a quantitative quota of the AI algorithm; an event strategy module generates a screenshot request in the case that the AI algorithm triggers a target event; a priority strategy module generates a screenshot request according to a current effective sample collection progress, an F1 value of a sample detection result, a covered scene of a collected sample, an interval time length from the last screenshot and a priority weight; a fusion decision unit integrates all the screenshot requests to obtain a to-be-processed list; and an execution unit controls the corresponding camera to execute the screenshot request according to the to-be-processed list. The system can improve the efficiency of algorithm verification and save resources.

Description

Technical Field

[0001] This application relates to the field of data processing technology, specifically to a multi-strategy adaptive image capture scheduling system for video AI algorithm verification. Background Technology

[0002] Video AI (Artificial Intelligence) target recognition algorithms require extensive real-world validation before deployment. For example, in security scenarios, multiple cameras deployed in different locations capture images in real time, which are then fed into the AI ​​algorithm model for target detection. The algorithm's effectiveness is then evaluated based on the recognition results. In real-time monitoring scenarios of target events (such as fires, illegal intrusions, and unauthorized operations), the efficient scheduling of multiple cameras for video capture is a key factor determining the efficiency of algorithm validation.

[0003] Traditional AI algorithm verification solutions typically employ a single, timed image capture method, setting fixed intervals for each camera and executing image capture operations according to the corresponding intervals for each camera. However, this approach cannot flexibly adapt to the sample requirements of different stages of algorithm verification, nor can it effectively distinguish between time periods containing targets and those without targets. This results in the system collecting a large number of invalid samples and potentially missing valid targets. For example, in a fire monitoring scenario, the actual time window for smoke and fire is only a few minutes. However, capturing images at fixed intervals may only capture a small number of images within the time window of smoke and fire, or even miss them entirely. If the interval is shortened to ensure that smoke and fire images are captured, a large number of invalid samples will be generated in other long periods without targets, resulting in a waste of storage and computing resources, a low effective sample rate, and consequently, low algorithm verification efficiency.

[0004] Therefore, it is necessary to find a method for scheduling cameras to save resources while improving the efficiency of algorithm verification. Summary of the Invention

[0005] The purpose of this application is to provide a multi-strategy adaptive image capture scheduling system for video AI algorithm verification, so as to save resources while improving the verification efficiency of AI algorithms.

[0006] To achieve the above objectives, the technical solution of this application is as follows: In a first aspect, embodiments of this application provide a multi-strategy image capture scheduling system for video AI algorithm verification, the system comprising: Configuration manager, timing strategy module, quantitative strategy module, event strategy module, priority strategy module, fusion decision unit, execution unit and multiple cameras; The configuration manager is configured to bind the camera, preset position and corresponding artificial intelligence algorithm to form a triplet, and configure corresponding image capture parameters for each triplet; the image capture parameters include: time interval, periodic working time period and quantitative limit of artificial intelligence algorithm, confidence threshold of target event, and priority weight. The timing strategy module is configured to generate image capture requests based on the timing interval of triples; The quantitative strategy module is configured to generate a screenshot request based on the periodic working time period and quantitative amount of the triplet artificial intelligence algorithm. The event strategy module is configured to generate a screenshot request when the target event is triggered by the artificial intelligence algorithm corresponding to the triple; The priority strategy module is configured to generate a capture request based on the effective sample collection progress of the triple in the most recent working time period, the F1 value of the sample detection result, the coverage scene of the collected samples, the interval since the last capture, and the priority weight; the effective sample is the sample whose confidence of the target event recognition result of the artificial intelligence algorithm is greater than the confidence threshold. The fusion decision unit is configured to integrate all screenshot requests to obtain a list of pending requests; The execution unit is configured to control the corresponding camera to execute a capture request based on the list of tasks to be processed.

[0007] The multi-strategy image capture scheduling system for video AI algorithm verification provided in this application adopts a multi-strategy joint image capture scheduling approach. Using a triplet formed by binding a camera, preset positions, and an AI algorithm as the basic unit, four strategy modules adaptively generate corresponding image capture requests for each triplet based on preset time intervals, quantitative sample requirements for algorithm analysis, algorithm recognition results for specific events, and the sample verification status of the system operation. These requests are then unified and integrated by a fusion decision unit to obtain a list of pending requests that can be executed sequentially. Finally, the execution unit controls the corresponding cameras to execute parallel image capture requests according to the list of pending requests.

[0008] Compared to traditional single-fixed-period image capture schemes, this application monitors the AI ​​algorithm's recognition results through an event strategy module to ensure timely capture of valid samples containing the target when a target event is triggered, avoiding the omission of valid samples containing the target. A quantitative strategy module executes quantitative image acquisition, enabling flexible adaptation to the sample quantity requirements corresponding to different stages of algorithm verification. A priority strategy module dynamically controls the allocation of image capture resources to triplets with poor verification results and lagging progress, thereby improving overall resource utilization efficiency. Based on this, the system's reliance on high-frequency timed image capture is significantly reduced. Compared to traditional schemes, this application can set longer time intervals (i.e., using low-frequency timed image capture) to meet the need to cover blank periods, thus significantly reducing the generation of invalid samples. This improves algorithm verification efficiency while saving system storage and computing resources. Attached Figure Description

[0009] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is one of the schematic diagrams of a multi-strategy image capture and scheduling system for video AI algorithm verification proposed in an embodiment of this application; Figure 2 This is a schematic diagram of the workflow of the event strategy module in one embodiment of this application; Figure 3 This is the second schematic diagram of a multi-strategy image capture and scheduling system for video AI algorithm verification proposed in one embodiment of this application. Detailed Implementation

[0011] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0012] It should be understood that the phrase "one embodiment" or "an embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0013] In the various embodiments of this application, it should be understood that the sequence number of each process described below does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0014] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects as detailed in this application.

[0015] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0016] Traditional time-interval AI algorithm verification schemes have the following drawbacks: (1) It is impossible to distinguish between high-value target periods and non-target periods, often missing the window of high-value targets (such as fire, illegal intrusion, etc.) and generating a large amount of redundant data during long periods of non-target, resulting in a waste of storage and computing resources; (2) The sample requirements differ at different stages of the algorithm analysis process. For example, a large number of samples are needed in the early stage to quickly evaluate the baseline effect of the algorithm, in the middle stage, samples of false alarm scenarios need to be collected to evaluate the optimization of the algorithm, and in the later stage, only a small number of samples are needed for algorithm result verification. However, the traditional solution has a fixed image capture frequency, which cannot dynamically adapt to the sample requirements required at different stages of AI algorithm analysis.

[0017] To address the aforementioned issues, this application employs a multi-strategy adaptive camera scheduling approach, combining timing, quantitative, event-based, and priority strategies to dynamically match sample requirements at different stages of algorithm verification, promptly capture high-value target windows, reduce redundant data generation, and improve algorithm verification efficiency. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0018] Figure 1 This is one of the schematic diagrams of a multi-strategy image capture and scheduling system for video AI algorithm verification proposed in an embodiment of this application. For example... Figure 1 As shown, the system includes: Configuration manager 101, timing strategy module 102, quantitative strategy module 103, event strategy module 104, priority strategy module 105, fusion decision unit 106, execution unit 107 and multiple cameras; The configuration manager 101 is configured to bind the camera, preset position and corresponding artificial intelligence algorithm to form a triplet, and configure corresponding image capture parameters for each triplet; the image capture parameters include: time interval, periodic working time period and quantitative limit of artificial intelligence algorithm, confidence threshold of target event, and priority weight. The timing strategy module 102 is configured to generate capture requests based on the timing interval of triples; The quantitative strategy module 103 is configured to generate a screenshot request based on the periodic working time period and quantitative quota of the triplet artificial intelligence algorithm. The event strategy module 104 is configured to generate a screenshot request when the target event is triggered by the artificial intelligence algorithm corresponding to the triple; The priority strategy module 105 is configured to generate a capture request based on the effective sample collection progress of the triplet in the most recent working time period, the F1 value of the sample detection result, the coverage scene of the collected samples, the interval since the last capture, and the priority weight; the effective sample is the sample whose confidence of the target event recognition result of the artificial intelligence algorithm is greater than the confidence threshold. The fusion decision unit 106 is configured to integrate all screenshot requests to obtain a list of pending requests; The execution unit 107 is configured to control the corresponding camera to execute a capture request based on the list of pending tasks.

[0019] In this embodiment, multiple cameras deployed in different locations are managed through a video platform interface. The image capture scheduling system interacts with the video platform interface to capture video frames from the real-time video streams of the cameras. Specifically, the image capture scheduling system includes a configuration manager, a fusion decision unit, an execution unit, and four strategy modules for generating image capture requests.

[0020] The configuration manager is used for parameter configuration, storage, and dynamic updates of all policy modules. It uses a triplet formed by binding "camera, preset position, and AI algorithm" as its basic unit, maintaining a corresponding independent parameter set for each triplet. This includes the following capture parameters: timing interval, working time period and quantitative limit of the AI ​​algorithm, confidence threshold of the target event, priority weight, etc. The camera's preset position includes multiple preset parameters, such as horizontal angle, pitch angle, focal length, and aperture. In this embodiment, each camera has at least one preset position.

[0021] Each strategy module independently generates screenshot requests according to its corresponding request strategy. Specifically, the timed strategy module continuously generates screenshot requests corresponding to the triplet at set intervals. The quantitative strategy module generates screenshot requests for the triplet within the corresponding working time period, based on the AI ​​algorithm's working time period and the required quantitative quota for that stage. The event strategy module generates screenshot requests based on the algorithm's recognition results of the samples collected by the triplet. When the algorithm's recognition result triggers a target event, this strategy module continuously generates screenshot requests for that triplet for a continuous period after the current moment to avoid missing valid samples containing the target. A target event refers to an event where the AI ​​algorithm-based target detection model detects a specific target. For example, a target detection model for a fire triggers a corresponding target event when it detects smoke or flames; a target detection model for illegal intrusion triggers a corresponding target event when it detects an unrecorded face. The priority strategy module obtains the effective sample collection progress, model AI algorithm F1 score, scenes covered by the collected samples, time interval since the last image capture, and a set of priority weights corresponding to each triple based on the image capture verification progress of each triple. It then sorts all triples by priority based on this information and generates image capture requests for each triple according to the request frequency corresponding to the priority.

[0022] The fusion decision unit aggregates the screenshot requests generated by the four decision modules, resolves conflicts and sorts them, and finally integrates them to obtain a list of pending tasks. This decision unit prioritizes and deduplicates the screenshot requests from all policy sources to obtain an executable list of pending tasks.

[0023] The execution unit reads multiple capture requests in the order of the pending list and controls the corresponding cameras to perform parallel capture operations. When executing capture requests, the execution unit considers the maximum number of concurrent captures the system can support and the minimum capture interval for a single camera, delaying and waiting for requests in the pending list that cannot be executed in parallel. For each capture request, the execution unit obtains the information of the corresponding triple (i.e., camera, preset position, and AI algorithm), then calls the video platform interface to adjust the camera angle to the corresponding preset position for real-time sample acquisition. After obtaining the video frame image, based on the information of the AI ​​algorithm in the triple, the image is finally sent to the corresponding object detection model for recognition.

[0024] In this embodiment, using triples bound to cameras, preset positions, and algorithms as units, multiple complementary strategies with different focuses are employed to achieve adaptive image capture scheduling. A timing strategy ensures basic coverage throughout the entire time period; a quantitative strategy matches the sample requirements of different stages of AI algorithm verification, reducing redundant data generation; event-driven triggering of high-value target capture avoids missing target event windows; and a priority strategy optimizes the allocation of capture request resources for each triple, tilting capture resources towards triples with poor verification results and lagging progress to improve overall resource utilization efficiency. Finally, a fusion decision unit performs global conflict resolution and priority ranking, ensuring real-time response to critical capture requests (e.g., fire video frame acquisition with high immediacy requirements) while further eliminating redundant capture requests, saving system storage and computing resources, and improving algorithm verification efficiency.

[0025] In one embodiment of this application, the timing strategy module is specifically configured to perform the following steps: Determine the initial first request frequency based on the timing interval of the target triplet; The historical target detection rate of the target triplet is statistically analyzed over multiple first working time periods, and the average target detection rate is calculated; the first working time period is obtained by dividing the periodic working time period based on the artificial intelligence algorithm. For each first working time period, if the historical target detection rate of the first working time period is greater than the average target detection rate, the first request frequency is increased during the first working time period; and if the historical target detection rate of the first working time period is less than the average target detection rate, the first request frequency is decreased during the first working time period. A screenshot request corresponding to the target triple is generated based on the current first request frequency, and a random delay is added when generating each screenshot request.

[0026] In one embodiment, the timing strategy module obtains the timing interval of the target triplet based on the image capture parameters of the target triplet to determine the first request frequency. This first request frequency is used as the initial frequency within the periodic working time period of the AI ​​algorithm corresponding to the target triplet. Based on a preset time period length, the periodic working time period of the AI ​​algorithm is divided into multiple first working time periods. For example, with a preset time period length of 4 hours, the 24-hour periodic working time period is divided into six first working time periods: 0:00-5:00, 6:00-11:00, 12:00-17:00, and 18:00-23:00. For each first working time period, the total number of samples collected by the target triplet within the first historical time period and the number of samples detecting the target event within that first historical time period are calculated. Optionally, in practical applications, the first historical time period can be customized according to actual needs; this embodiment does not impose any restrictions. The historical target detection rate of target triplet i within the first working time period is calculated as follows: .

[0027] Then, the average historical target detection rate for each first working time period is calculated and determined as the average target detection rate. The historical target detection rate for each first working time period is compared with this average target detection rate. If the historical target detection rate is greater than the average target detection rate, the first working time period is determined as a high-value time period, and the frequency of first requests within this time period is increased, i.e., the image capture interval is shortened. ;in As an acceleration factor, and .

[0028] If the historical target detection rate is lower than the average target detection rate, this first working time period is identified as a low-value time period. Therefore, the frequency of the first request within this time period is reduced, i.e., the image capture interval is lengthened. ;in It is a deceleration factor, and If the historical target detection rate equals the average target detection rate, the current first request frequency will remain unchanged.

[0029] By dynamically adjusting the frequency of first requests in each first time period based on the historical target detection rate during the execution of the timing strategy, the image capture resources are biased towards the period with a higher historical detection rate, thereby improving the effective sample rate and further reducing the number of invalid samples.

[0030] In addition, to avoid multiple cameras triggering image capture operations simultaneously, which could lead to excessive system load and potential failure, a random delay is added to the generation time of each image capture request generated based on the first request frequency. ,in jitter coefficient (e.g.) This allows the image capture requests from different cameras to be distributed over time.

[0031] In one embodiment of this application, the quantitative strategy module is specifically configured to perform the following steps: Based on the periodic working time period and quantitative quota of the artificial intelligence algorithm bound to the target triple, the initial second request frequency of the target triple within the working time period is calculated. During the specified working time period, the collection progress of the target triplet is calculated based on the quantitative limit of the artificial intelligence algorithm of the target triplet and the number of samples collected during the current working time period. Based on the acquisition progress, the total duration of the periodic working period, and the duration already consumed, calculate the acquisition rate deviation between the current acquisition rate and the expected acquisition rate of the target triplet. If the sampling rate deviation is less than a first threshold, the second request frequency is increased; and if the sampling rate deviation is greater than a second threshold, the second request frequency is decreased; where the second threshold is greater than the first threshold. Based on the current second request frequency, a screenshot request for the target triple is generated.

[0032] In one embodiment, the initial second request frequency of the target triplet during the working period is determined based on the duration of the periodic working time of the AI ​​algorithm bound to the target triplet and the quantitative quota of the algorithm in the current verification phase. During this working period, the sample collection progress of the target triplet is monitored in real time. : ; Where N is the number of samples to be collected for the target triplet within the current working time period; N represents the number of samples collected for the target triplet within the current working time period. It's worth noting that when the AI ​​algorithm binds to only one target triplet, N is the same as the AI ​​algorithm's quantitative quota; when the AI ​​algorithm binds to multiple target triplets, the AI ​​algorithm's quantitative quota is allocated to each target triplet to determine the number of samples N to be collected for each target triplet.

[0033] Based on the current sample acquisition progress of the target triplet, combined with the total duration of the periodic working period and the time already consumed, the acquisition rate deviation between the current actual acquisition rate of the target triplet and the expected acquisition rate under ideal conditions is calculated. : .

[0034] The second request frequency is dynamically adjusted based on the current acquisition rate deviation. Specifically, the second request frequency is increased when the current acquisition rate deviation is less than a first threshold, and decreased when the acquisition rate deviation is greater than a second threshold. In practical applications, a first threshold less than 0 (e.g., -0.2) and a second threshold greater than 0 (e.g., 0.2) can be preset according to actual needs. By monitoring the sample acquisition progress corresponding to the triplet in real time, the second request frequency is dynamically adjusted to meet the quantitative quota requirements of the samples. Within each periodic working period, when the number of samples acquired reaches the quantitative quota, the generation of image capture requests stops, and the second request frequency is automatically replenished to the initial value in the next periodic working period.

[0035] As one embodiment of this application, the quantitative strategy module is configured to calculate the initial second request frequency of the target triplet within the working time period based on the periodic working time period and quantitative quota of the artificial intelligence algorithm bound to the target triplet, including: The historical effective sample rate of all triples bound to the artificial intelligence algorithm is statistically analyzed; the historical effective sample rate is determined based on the total number of samples taken by the camera to which the triple belongs for the artificial intelligence algorithm and the number of effective samples. The allocation weights corresponding to each triplet are determined based on the historical effective sample rate of each triplet bound to the artificial intelligence algorithm. Based on the weights assigned to each triple and the quantitative limit of the artificial intelligence algorithm, the sample limit corresponding to each triple is determined. Based on the sample quota of the target triplet and the periodic working time of the artificial intelligence algorithm, the initial second request frequency corresponding to the target triplet is calculated.

[0036] In one embodiment, when an AI algorithm is bound to a triplet of multiple cameras, a suitable quantitative quota is allocated to the camera to which each triplet belongs, based on the historical effective sample rate of each triplet.

[0037] First, count the number of samples collected by the camera to which the target triple belongs within the second historical time period (e.g., 1 year) for the AI ​​algorithm, and the number of samples that detected the target event, and calculate the historical effective sample rate of the target triple i: ; Calculate the assigned weights for target triple i based on the historical effective sample rates of each triple: ; Calculate the corresponding sample quota based on the assigned weights of the target triple i and the quantitative quota N of the AI ​​algorithm. as follows: .

[0038] By weighting the allocation of sample demand based on the historical effective sample rate of each target triplet, the system can dynamically perceive the differences in the AI ​​algorithm verification progress of each location and achieve adaptive resource allocation according to the location's needs. A larger proportion of sample quotas is allocated to locations with higher historical effective sample rates. Compared with the method of allocating sample quotas equally to all locations, this not only improves the efficiency of algorithm verification but also reduces the number of invalid samples collected at locations with lower effective sample rates, further saving system resources.

[0039] In one embodiment of this application, the event strategy module is specifically configured to perform the following steps: The confidence level of the latest recognition result of the artificial intelligence algorithm bound to the target triplet is obtained in real time, and the confidence level is compared with the confidence level threshold. If the confidence level of the latest identification result of the artificial intelligence algorithm is greater than or equal to the confidence threshold, a target event is determined to be triggered. Upon triggering the target event, a first duration timer is started, and a capture request for the target triple is generated according to the third request frequency. If the target event is triggered again during the first duration of the timer, the first duration of the timer will be reset. After the first duration of timing ends, a second duration of timing begins, and during the second duration of timing, the generation of screenshot requests for the target triplet is stopped.

[0040] In one embodiment, the event strategy module switches the state of the target triplet based on the confidence level of the latest identification result obtained after the target triplet sample is processed by the target detection model, so as to achieve high-frequency collection of high-value target samples.

[0041] Figure 2 This is a schematic diagram illustrating the workflow of the event strategy module in one embodiment of this application. For example... Figure 2 As shown, the event strategy module monitors the AI ​​algorithm detection results of samples collected by each triplet in real time and compares the confidence level of the latest recognition result of each triplet with the corresponding confidence threshold of the AI ​​algorithm. If the confidence level of the latest sample's recognition result is greater than or equal to the confidence threshold, the target triplet is determined to trigger a target event, and the sample triggering the target event is identified as a high-value valid sample. For example, in a fire monitoring scenario, if the confidence level of the target detection model's recognition result for the latest sample is greater than the confidence threshold, the latest sample is determined to trigger a target event (fire occurrence).

[0042] Upon triggering the target event, the event strategy module switches the target triplet to an encrypted state and initiates a first-duration timer. During this timer, it generates screenshot requests for the target triplet at a third request frequency. Furthermore, during the first-duration timer, it continuously monitors the real-time sample recognition results of the triplet. If the target triplet triggers the target event again, the current timer is reset, and the first-duration timer restarts to extend the duration of generating screenshot requests.

[0043] After the first timeout period ends, the event strategy module switches the target triplet to a cooldown state and starts a second timeout period. During this second timeout period, it stops monitoring target event triggers for the sample detection results of the target triplet and does not generate any screenshot requests for the target triplet. The cooldown state prevents abnormal scenarios (such as continuous false alarms) from causing infinite encryption and exhausting system resources. Furthermore, if the number of consecutive timeouts reaches a preset threshold, the event strategy module stops generating screenshot requests for the target triplet and immediately switches to a cooldown state. After the second timeout period ends, the cooldown state of the target triplet ends, and monitoring target event triggers for the sample detection results of the target triplet resumes.

[0044] Optionally, the second duration can be set to twice the first duration. It is worth noting that the first duration, second duration, and confidence threshold of each triple in each system are independent of each other and do not interfere with each other. That is, when any triple is in a cooling-off state, it does not affect the event strategy module's generation of screenshot requests for other triples.

[0045] By employing the above method, an encryption state is set to ensure rapid and frequent sample collection when high-value targets (i.e., the confidence level of sample detection results is greater than or equal to the confidence threshold) appear; a cooling-off state is set to prevent abnormal situations (such as continuous false alarms) from causing unlimited encryption to exhaust system resources. This approach can both respond quickly and intensively to the appearance of valid targets, and avoid resource exhaustion through the limitations of the cooling-off state and the number of collections threshold.

[0046] In one embodiment, a sliding window event triggering mechanism can be used to count the number of times the target triplet in the most recent event window triggers the target event. When the number of times the sample triggers the target event reaches a preset fixed threshold, the target triplet is switched to an encrypted state, and the number of times the sample in the window triggers the target event is continuously monitored. When the number of times the sample in the window triggers the target event falls below the fixed threshold, the target triplet is switched to a cooldown state.

[0047] In one embodiment of this application, the priority strategy module is specifically configured to perform the following steps: Based on the first interval, determine the initial fourth request frequency; Based on the effective sample collection progress, F1 score of sample detection results, coverage of collected samples, interval since the last capture, and priority weight of each triplet in the most recent working time period, the comprehensive priority score of the triplet is calculated using at least one of the following methods: calculated based on the first interval time, or calculated when the triplet triggers a key event; the key event includes at least one of the following: obtaining the latest recognition result of the artificial intelligence algorithm, the number of samples collected within the periodic working time period of the artificial intelligence algorithm reaching a stage threshold, or a change in the priority weight of the triplet; Sort all triplet composite priority scores from highest to lowest, and determine the range of scores that the target triplet's target composite priority score belongs to among all priority scores. If the target overall priority score falls within a first score range, increase the fourth request frequency of the target triplet; if the target overall priority score falls within a second score range, maintain the current fourth request frequency; and if the target overall priority score falls within a third score range, decrease the fourth request frequency of the target triplet; the overall priority score in the first score range is higher than the overall priority score in the second score range, and the overall priority score in the second score range is higher than the overall priority score in the third score range.

[0048] In one embodiment, the priority strategy module integrates four dimensions—verification progress, verification effect, sample diversity, and timeliness—to comprehensively prioritize each triple, thereby enabling intelligent and dynamic allocation of image capture resources.

[0049] Specifically, the priority strategy module obtains the effective sample collection progress, sample detection result F1 value, coverage scene of collected samples, interval time since the last image capture, and priority weight of each triplet within the most recent working time period (i.e., the periodic working time period of the AI ​​algorithm) according to a preset first interval, and calculates the corresponding single priority score, as follows: Data collection progress priority score: ; Validation effect priority score: ; Diversity Priority Score: The number of scene conditions covered by the target triplet's samples refers to the number of different scene conditions actually collected by the target triplet during the AI ​​algorithm verification process, including different time periods (morning / noon / evening / night) and different weather conditions (sunny / cloudy / rainy / foggy), etc.

[0050] Freshness priority score: .

[0051] In this embodiment, all triples in the system use the same set of priority weights, which are composed of multiple individual weights, including individual weights for sample collection progress, validation effectiveness, sample diversity, and sample freshness. In practical applications, the analytic hierarchy process (AHP) can be used to help determine the individual weights corresponding to sample collection progress, validation effectiveness, sample diversity, and sample freshness.

[0052] Calculate the overall priority score based on the individual priority scores and their corresponding weights. as follows: ; Wherein, W1 is the single-item weight of data collection progress; W2 is the single-item weight of validation effect; W3 is the single-item weight of sample diversity; W4 is the single-item weight of sample freshness; and, .

[0053] After calculating the overall priority score for each triplet, all triplets are sorted from highest to lowest based on their overall priority score. For any target triplet, the score range of its corresponding overall priority score in the sorted results is determined, and the frequency of generating screenshot requests for that target triplet is adjusted according to this score range. Specifically, the top 20% of overall priority scores in the sorted results are assigned to the first score range, scores between 20% and 80% are assigned to the second score range, and the bottom 20% are assigned to the third score range. For triplets with overall priority scores in the first score range, the current fourth request frequency is increased by a specified multiple (e.g., 2 times); for triplets in the second score range, the current fourth request frequency remains unchanged; and for triplets in the third score range, the current fourth request frequency is decreased by a specified multiple (e.g., 0.5 times). The priority strategy module generates corresponding screenshot requests based on the fourth request frequency of each triplet.

[0054] In this way, the system can ensure that the image capture resources are biased towards triples with lagging sample collection progress, poor verification results, incomplete scene coverage, and long periods without image capture, thereby improving the overall resource utilization efficiency.

[0055] In addition, the priority strategy module will immediately calculate the overall priority score in the event of the following key events to adjust resource allocation in a timely manner, specifically including: (1) Obtain the latest recognition results of the artificial intelligence algorithm. After obtaining the new recognition results, the priority decision module recalculates the comprehensive priority score corresponding to the triple based on the newly obtained F1 value; (2) When the number of samples collected during the periodic working period of the artificial intelligence algorithm reaches the stage threshold (e.g., 50%, 100%), the comprehensive priority score corresponding to the triplet is recalculated. (3) If the priority weight of the triple changes, the overall priority score corresponding to the triple is recalculated. In practical applications, users can manually modify the priority weights according to actual needs during system operation. For example, if users tend to consider the sample validation effect more than the scene coverage, they can manually increase the single weight W2 of validation effect and decrease the single weight W3 of sample diversity.

[0056] As one embodiment of this application, the fusion decision unit is specifically configured to perform the following steps: Based on the second interval, obtain the screenshot requests generated by each strategy module; For image capture requests from the same camera, perform the following steps: Multiple screenshot requests for the same preset position are merged, and the screenshot request with the highest priority is selected and retained according to the strategy priority from high to low. The strategy priority from high to low is: event strategy, comprehensive priority score belonging to the first score range, comprehensive priority score belonging to the second score range, quantitative strategy, timed strategy, and comprehensive priority score belonging to the third score range. For multiple screenshot requests with different preset positions, sort them from high to low according to the comprehensive priority score of the triplet to which the screenshot request belongs; If the number of image capture requests from the camera exceeds a single camera threshold, the image capture requests exceeding the single camera threshold are removed.

[0057] In one embodiment, the fusion decision unit aggregates the image capture requests generated by each strategy module at a fixed second interval, and divides all image capture requests by camera. For all image capture requests from each camera, intra-camera integration is performed in the following manner: (1) When there are multiple screenshot requests for the same preset position, only the screenshot request with the highest policy priority is retained. The policy priorities from high to low are: event policy, comprehensive priority score belongs to the first score range, comprehensive priority score belongs to the second score range, quantitative policy, timed policy, and comprehensive priority score belongs to the third score range; (2) Multiple screenshot requests from different preset positions are sorted from high to low according to their overall priority scores; (3) Count the number of image capture requests after sorting and compare it with the single camera threshold. If the number is greater than the single camera threshold, truncate and remove image capture requests that exceed the single camera threshold, and keep only the top-ranked image capture requests.

[0058] In one embodiment of this application, the fusion decision unit is further configured to perform the following steps: The screenshot requests are sorted from high to low according to the priority of the strategy described above; For multiple screenshot requests with the same strategy priority, sort them from high to low according to the overall priority score; For multiple screenshot requests with the same overall priority score, they are sorted according to the order of request time.

[0059] In the above embodiments, after resolving conflicts and prioritizing image capture requests from the same camera, it is also necessary to perform global request integration on all remaining image capture requests in the current system. The specific process is as follows: (1) Sort all screenshot requests in descending order of strategy priority; (2) For multiple screenshot requests with the same strategy priority, sort them from high to low according to the comprehensive priority score of the triple to which each screenshot request belongs; (3) When the strategy priority and comprehensive priority scores are the same, the screenshot requests are sorted according to the order of the request time, and the screenshot requests with earlier request times are executed first.

[0060] After integrating the image capture requests within a single camera and globally, an executable list of pending tasks is generated based on the current sorting results and sent to the execution unit.

[0061] In one embodiment of this application, the execution unit is specifically configured to perform the following steps: Obtain the current service status of the camera corresponding to the image capture request; the service status is any of the following: idle, real-time viewing, video playback, patrol, manual control, or fault. If the service status is idle or video playback, the image capture request is executed directly; When the business status is real-time viewing, the image capture request is continuously retried according to the policy priority corresponding to the image capture request until the number of retries reaches the first retry threshold or the retry duration reaches the second retry threshold. When the business status is cruise, the image capture request is forcibly executed after the camera is adjusted to the closest stopping position to the preset position corresponding to the image capture request. When the business status is manual control, the image capture request will be continuously retried until the business status ends. If the service status is faulty, skip the screenshot request; Before executing the image capture request, record a snapshot of the current state of the camera; the snapshot includes: preset position, angle, and lens focal length; After executing the image capture request, control the camera to restore the state corresponding to the state snapshot.

[0062] In practical applications, video platforms typically run their own video services in real time. Traditional AI algorithm verification schemes often fail to consider the camera's current operational status when performing image capture operations. Forcibly switching the pan-tilt-zoom (PTZ) position after generating the capture request often interrupts the camera's ongoing services, impacting user experience. In one embodiment, by acquiring the target camera's real-time operational status before executing the capture request and employing a pre-defined service compatibility handling method, the system can complete the image capture operation required by the AI ​​algorithm without interfering with the camera's normal operation.

[0063] Specifically, before executing a capture request from the pending list, the execution unit first obtains the current service status of the camera that the capture request needs to control, and then performs operations such as capturing, retrying, or skipping based on the service status. In this embodiment, the service status of the video platform and the corresponding processing methods are as follows: (1) Idle state or video playback state: This indicates that the current camera is not running any services (video playback is achieved by reading historical stored data and does not occupy the real-time video stream), and you can directly execute the image capture request; (2) Real-time viewing status: This indicates that the camera is currently running a real-time video stream viewing service. Therefore, it needs to be retried and wait for the user to finish viewing in real time before executing the capture request. The monitoring unit continuously retries the current capture request and records the number of retryes and the retry duration. If the number of retryes reaches the first retry threshold (e.g., 10 times) or the retry duration reaches the second retry threshold (e.g., 30 seconds), the retry is deemed to have failed, the retry is stopped, and the capture request is skipped. Optionally, for capture requests with high policy priority (including event policies and comprehensive priority scores belonging to the first score range), the capture request can be forcibly executed if the retry duration reaches the third threshold (e.g., 60 seconds), and the real-time viewing service is restored after execution.

[0064] (3) Cruise Status: This indicates that the camera is currently in automatic cruise mode. In this case, the execution unit acquires all stopping positions during automatic cruise, and when the camera pan-tilt angle is adjusted to the stopping position closest to the preset position of the current capture request, it forcibly inserts and executes the capture request. After executing the capture request, the cruise service is resumed.

[0065] (4) Manual control status: This indicates that the camera is currently under manual control by the user and the image capture request cannot be executed immediately. Therefore, an unlimited number of delayed retries will be performed until the user ends manual control, at which point the image capture request will be executed.

[0066] (5) Fault status: This indicates that the camera is currently malfunctioning and cannot execute the image capture request. Therefore, the execution unit skips the image capture request directly.

[0067] Before executing a capture request, the execution unit records a complete snapshot of the camera's state (preset position, angle, focal length, etc.) and restores it after execution to ensure that the original business operations of the video platform are not affected. Optionally, if the capture request is skipped (due to retry failure or camera malfunction), the execution unit also records the corresponding request skipping log and camera state information for user viewing.

[0068] In one embodiment, a time-slice reservation mechanism can be employed to divide the target camera's operating time into multiple time slices. This allows the video platform's original services and the system's image capture request processing to alternately occupy different time slices, thereby ensuring the timeliness of image capture request processing without affecting the video platform's original services. For example, one second can be divided into 20 time slices, with the video platform's original services and the system's image capture request processing each occupying 10 time slices alternately.

[0069] In one embodiment, video stream replication technology is used in advance to create corresponding virtual video streams based on the real-time video streams of each camera in the video platform. This separates the original business of the video platform from the image capture request, and samples are collected based on the virtual video stream, thereby ensuring that the parallel operation of the image capture request and the original business does not interfere with each other.

[0070] As one embodiment of this application, the system further includes a strategy adjustment unit, comprising: an evaluation module, a monitoring module, and an adjustment module; The evaluation module is configured to calculate evaluation metrics for all triples, including at least one of the following: calculating the capture efficiency based on the number of valid samples and the total number of samples collected for all triples; calculating the target capture rate based on the number of triggered target events corresponding to all triples and the total number of actual target events; calculating the strategy contribution of each strategy module based on the number of valid samples and the total number of samples captured by the timing strategy module, the quantitative strategy module, the event strategy module, and the priority strategy module, respectively; and calculating the resource balance based on the number of samples collected by each camera. The monitoring module is configured to compare the evaluation index with the corresponding optimization threshold, and if the evaluation index is lower than the optimization threshold, generate optimization suggestions for at least one of the following capture parameters: time interval, confidence threshold of target event, and priority weight. The adjustment module is configured to generate corresponding adjustment instructions for the screenshot parameters based on the optimization suggestions, and send them to the configuration manager; The configuration manager is also configured to update the corresponding screenshot parameters according to the adjustment instructions.

[0071] Figure 3 This is a second schematic diagram of a multi-strategy image capture and scheduling system for video AI algorithm verification proposed in an embodiment of this application. For example... Figure 3 As shown, in this embodiment, the system also deploys a strategy adjustment unit 108, including an evaluation module, a monitoring module, and an adjustment module. The evaluation module analyzes the system's image capture scheduling execution effect by calculating multiple evaluation indicators; the monitoring module generates corresponding image capture parameter optimization suggestions based on the image capture scheduling execution effect; the adjustment module generates corresponding adjustment instructions based on the image capture parameter optimization suggestions and sends them to the configuration manager; the configuration manager modifies the image capture parameters according to the adjustment instructions, and the changes take effect immediately. This achieves adaptive parameter adjustment and optimization of the system's image capture scheduling, improving overall resource utilization and AI algorithm verification efficiency.

[0072] The evaluation module calculates the following five evaluation metrics at a fixed period (e.g., daily) to assess the overall performance of the system's image capture and scheduling execution: (1) Capture efficiency = number of samples containing valid targets / total number of samples collected by the system; (2) Target capture rate = number of actual triggered target events / total number of actual target events; (3) The contribution of strategy i = the number of samples containing effective targets of strategy i / the total number of samples based on strategy i; where strategy i ∈ {timed strategy, quantitative strategy, event strategy, priority strategy}; (4) Business interruption rate = number of business interruptions caused by executing screenshot requests / total number of screenshots (the optimization target for this indicator is 0%). (5) Resource balance = 1 - standard deviation of the number of samples collected by each camera / average number of samples collected by all cameras.

[0073] The monitoring module acquires the evaluation metrics calculated by the evaluation module in real time, compares each metric with its corresponding optimization threshold, and generates optimization suggestions based on at least one of the following capture parameters when a metric falls below the optimization threshold: timing interval, confidence threshold of the target event, and priority weight. For example, it might suggest extending the timing interval and increasing the confidence threshold of the target event to reduce the number of samples collected and save system resources.

[0074] For example, optimization suggestions for timing strategies: When the capture success rate is less than 5%, it is recommended to extend the timing interval; when the target capture rate is less than 80%, it is recommended to shorten the timing interval. Optimization suggestions for event strategies: When the proportion of triggered target events that actually contain the target (i.e., the true trigger rate) is less than 30%, it is recommended to increase the confidence threshold of the target event to reduce false triggers; when the true trigger rate is greater than 90% and the target capture rate is less than 80%, it is recommended to decrease the confidence threshold. Optimization suggestions for priority strategies: Sort each strategy by contribution from highest to lowest, and based on the contribution ranking results, it is recommended to adjust the priority of each strategy.

[0075] The adjustment module generates corresponding adjustment instructions based on optimization suggestions for the capture parameters and sends them to the configuration manager. Upon receiving the adjustment instructions, the configuration manager modifies the capture parameters accordingly and the changes take effect immediately. All adjustment records are stored as log data, supporting manual review and retrospection. Furthermore, in the event of abnormal situations (such as a sudden 50% drop in capture effectiveness within a short period, continuous capture failures, or a business interference rate > 0), the monitoring module generates alarm information to facilitate timely troubleshooting and resolution of faults and interference.

[0076] It is worth noting that when the priority weights change, the priority strategy module immediately recalculates the overall priority score of each triplet based on the updated priority weights.

[0077] In this embodiment, a quantitative evaluation mechanism for the system's image capture scheduling effect is established through a strategy adjustment unit. Based on the evaluation results, the strategy parameters are adaptively adjusted, realizing the adaptive dynamic allocation of image capture resources and the adaptive dynamic adjustment of image capture strategy parameters. This achieves dynamic and continuous optimization of the overall system resource utilization and AI algorithm verification efficiency, saving labor costs while lowering the technical threshold, making it suitable for non-professionals.

[0078] To verify the effectiveness of this solution, this embodiment deploys and verifies the target AI algorithm in a real-world project. The experiment deploys 10 cameras and uses 8 target AI algorithms for parallel verification. The cameras used in this experiment are uniformly managed by a video platform that supports standard protocols such as GB / T28181 and ONVIF, and streaming media playback supports RTSP, RTMP, and HLS protocols.

[0079] After running this system for 4 weeks, compared with the traditional time-interval AI algorithm verification scheme, the image capture efficiency of this system is 22.6%, which is much higher than the traditional scheme's 8.2%. In addition, the target capture rate of this system is stable at over 95%, and the total sample collection volume is reduced by 30% compared with the traditional scheme.

[0080] In fire monitoring scenarios, the timing strategy module of this system contributes about 70% of the basic image captures; the event strategy module contributes about 15% of the high-value image captures (with the highest image capture effectiveness rate reaching 38.2%); and the priority strategy module contributes about 15% of the supplementary image captures.

[0081] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0082] For the sake of simplicity, the method embodiments are described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and components involved are not necessarily essential to this application.

[0083] Those skilled in the art will understand that embodiments of this application can be provided as methods, apparatus, or computer program products. Therefore, embodiments of this application can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, embodiments of this application can take the form of computer program products implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0084] This application describes embodiments with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (systems), and computer program products according to embodiments of this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0085] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0086] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0087] Although preferred embodiments of the embodiments of this application have been described, those skilled in the art, once they understand the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, this application is to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of this application.

[0088] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0089] The above provides a detailed description of the multi-strategy adaptive image capture and scheduling system for video AI algorithm verification provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A multi-strategy image capture and scheduling system for video AI algorithm verification, characterized in that, include: Configuration manager, timing strategy module, quantitative strategy module, event strategy module, priority strategy module, fusion decision unit, execution unit and multiple cameras; The configuration manager is configured to bind the camera, preset position and corresponding artificial intelligence algorithm to form a triplet, and configure corresponding image capture parameters for each triplet; the image capture parameters include: time interval, periodic working time period and quantitative limit of artificial intelligence algorithm, confidence threshold of target event, and priority weight. The timing strategy module is configured to generate image capture requests based on the timing interval of triples; The quantitative strategy module is configured to generate a screenshot request based on the periodic working time period and quantitative amount of the triplet artificial intelligence algorithm. The event strategy module is configured to generate a screenshot request when the target event is triggered by the artificial intelligence algorithm corresponding to the triple; The priority strategy module is configured to generate a capture request based on the effective sample collection progress of the triple in the most recent working time period, the F1 value of the sample detection result, the coverage scene of the collected samples, the interval since the last capture, and the priority weight; the effective sample is the sample whose confidence of the target event recognition result of the artificial intelligence algorithm is greater than the confidence threshold. The fusion decision unit is configured to integrate all screenshot requests to obtain a list of pending requests; The execution unit is configured to control the corresponding camera to execute a capture request based on the list of tasks to be processed.

2. The multi-strategy image capture and scheduling system for video AI algorithm verification according to claim 1, characterized in that, The timing strategy module is specifically configured to execute the following steps: Determine the initial first request frequency based on the timing interval of the target triplet; The historical target detection rate of the target triplet is statistically analyzed over multiple first working time periods, and the average target detection rate is calculated; the first working time period is obtained by dividing the periodic working time period based on the artificial intelligence algorithm. For each first working time period, if the historical target detection rate of the first working time period is greater than the average target detection rate, the first request frequency is increased during the first working time period; and if the historical target detection rate of the first working time period is less than the average target detection rate, the first request frequency is decreased during the first working time period. A screenshot request corresponding to the target triple is generated based on the current first request frequency, and a random delay is added when generating each screenshot request.

3. The multi-strategy image capture and scheduling system for video AI algorithm verification according to claim 1, characterized in that, The quantitative strategy module is specifically configured to perform the following steps: Based on the periodic working time period and quantitative quota of the artificial intelligence algorithm bound to the target triple, the initial second request frequency of the target triple within the working time period is calculated. During the specified working time period, the collection progress of the target triplet is calculated based on the quantitative limit of the artificial intelligence algorithm of the target triplet and the number of samples collected during the current working time period. Based on the acquisition progress, the total duration of the periodic working period, and the duration already consumed, calculate the acquisition rate deviation between the current acquisition rate and the expected acquisition rate of the target triplet. If the sampling rate deviation is less than a first threshold, the second request frequency is increased; and if the sampling rate deviation is greater than a second threshold, the second request frequency is decreased. The second threshold is greater than the first threshold; Based on the current second request frequency, a screenshot request for the target triple is generated.

4. The multi-strategy image capture and scheduling system for video AI algorithm verification according to claim 3, characterized in that, The quantitative strategy module is configured to calculate the initial second request frequency of the target triple within the working period based on the periodic working time and quantitative quota of the artificial intelligence algorithm bound to the target triple, including: The historical effective sample rate of all triples bound to the artificial intelligence algorithm is statistically analyzed; the historical effective sample rate is determined based on the total number of samples taken by the camera to which the triple belongs for the artificial intelligence algorithm and the number of effective samples. The allocation weights for each triplet are determined based on the historical effective sample rates of each triplet bound to the aforementioned artificial intelligence algorithm. Based on the weights assigned to each triple and the quantitative limit of the artificial intelligence algorithm, the sample limit corresponding to each triple is determined. Based on the sample quota of the target triplet and the periodic working time of the artificial intelligence algorithm, the initial second request frequency corresponding to the target triplet is calculated.

5. The multi-strategy image capture and scheduling system for video AI algorithm verification according to claim 1, characterized in that, The event strategy module is specifically configured to execute the following steps: The confidence level of the latest recognition result of the artificial intelligence algorithm bound to the target triplet is obtained in real time, and the confidence level is compared with the confidence level threshold. If the confidence level of the latest identification result of the artificial intelligence algorithm is greater than or equal to the confidence threshold, a target event is determined to be triggered. Upon triggering the target event, a first duration timer is started, and a capture request for the target triple is generated according to the third request frequency. If the target event is triggered again during the first duration of the timer, the first duration of the timer will be reset. After the first duration of timing ends, a second duration of timing begins, and during the second duration of timing, the generation of screenshot requests for the target triplet is stopped.

6. The multi-strategy image capture and scheduling system for video AI algorithm verification according to claim 1, characterized in that, The priority strategy module is specifically configured to perform the following steps: Based on the first interval, determine the initial fourth request frequency; Based on the effective sample collection progress, F1 score of sample detection results, coverage of collected samples, interval since the last capture, and priority weight of each triplet in the most recent working time period, the comprehensive priority score of the triplet is calculated using at least one of the following methods: calculated based on the first interval time, or calculated when the triplet triggers a key event; the key event includes at least one of the following: obtaining the latest recognition result of the artificial intelligence algorithm, the number of samples collected within the periodic working time period of the artificial intelligence algorithm reaching a stage threshold, or a change in the priority weight of the triplet; Sort all triplet composite priority scores from highest to lowest, and determine the range of scores that the target triplet's target composite priority score belongs to among all priority scores. If the target overall priority score falls within a first score range, increase the fourth request frequency of the target triplet; if the target overall priority score falls within a second score range, maintain the current fourth request frequency; and if the target overall priority score falls within a third score range, decrease the fourth request frequency of the target triplet; the overall priority score in the first score range is higher than the overall priority score in the second score range, and the overall priority score in the second score range is higher than the overall priority score in the third score range.

7. The multi-strategy image capture and scheduling system for video AI algorithm verification according to claim 6, characterized in that, The fusion decision unit is specifically configured to perform the following steps: Based on the second interval, obtain the screenshot requests generated by each strategy module; For image capture requests from the same camera, perform the following steps: Multiple screenshot requests for the same preset position are merged, and the screenshot request with the highest priority is selected and retained according to the strategy priority from high to low. The strategy priority from high to low is: event strategy, comprehensive priority score belonging to the first score range, comprehensive priority score belonging to the second score range, quantitative strategy, timed strategy, and comprehensive priority score belonging to the third score range. For multiple screenshot requests with different preset positions, sort them from high to low according to the comprehensive priority score of the triplet to which the screenshot request belongs; If the number of image capture requests from the camera exceeds a single camera threshold, the image capture requests exceeding the single camera threshold are removed.

8. The multi-strategy image capture and scheduling system for video AI algorithm verification according to claim 7, characterized in that, The fusion decision unit is also configured to perform the following steps: The screenshot requests are sorted from high to low according to the priority of the strategy described above; For multiple screenshot requests with the same strategy priority, sort them from high to low according to the overall priority score; For multiple screenshot requests with the same overall priority score, they are sorted according to the order of request time.

9. The multi-strategy image capture and scheduling system for video AI algorithm verification according to claim 1, characterized in that, The execution unit is specifically configured to perform the following steps: Obtain the current service status of the camera corresponding to the image capture request; the service status is any of the following: idle, real-time viewing, video playback, patrol, manual control, or fault. If the service status is idle or video playback, the image capture request is executed directly; When the business status is real-time viewing, the image capture request is continuously retried according to the policy priority corresponding to the image capture request until the number of retries reaches the first retry threshold or the retry duration reaches the second retry threshold. When the business status is cruise, the image capture request is forcibly executed after the camera is adjusted to the closest stopping position to the preset position corresponding to the image capture request. When the business status is manual control, the image capture request will be continuously retried until the business status ends. If the service status is faulty, skip the screenshot request; Before executing the image capture request, record a snapshot of the current state of the camera; the snapshot includes: preset position, angle, and lens focal length; After executing the image capture request, control the camera to restore the state corresponding to the state snapshot.

10. The multi-strategy image capture and scheduling system for video AI algorithm verification according to claim 9, characterized in that, The system also includes a strategy adjustment unit, comprising: an evaluation module, a monitoring module, and an adjustment module; The evaluation module is configured to calculate evaluation metrics for all triples, including at least one of the following: calculating the capture efficiency based on the number of valid samples and the total number of samples collected for all triples; calculating the target capture rate based on the number of triggered target events corresponding to all triples and the total number of actual target events; calculating the strategy contribution of each strategy module based on the number of valid samples and the total number of samples captured by the timing strategy module, the quantitative strategy module, the event strategy module, and the priority strategy module, respectively; and calculating the resource balance based on the number of samples collected by each camera. The monitoring module is configured to compare the evaluation index with the corresponding optimization threshold, and if the evaluation index is lower than the optimization threshold, generate optimization suggestions for at least one of the following capture parameters: time interval, confidence threshold of target event, and priority weight. The adjustment module is configured to generate corresponding adjustment instructions for the screenshot parameters based on the optimization suggestions, and send them to the configuration manager; The configuration manager is also configured to update the corresponding screenshot parameters according to the adjustment instructions.

Images