A real-time shooting quality early warning method and system based on shooting behavior characteristics

Abstract

The application relates to the technical field of computer vision and image processing, and discloses a real-time shooting quality early warning method and system based on shooting behavior characteristics. The method comprises the following steps: collecting multi-source behavior data, identifying an activity category, taking a preview timestamp as a time reference, performing time alignment and data cleaning, and obtaining standard shooting behavior data; inputting the standard shooting behavior data into a multi-dimensional quality risk model; obtaining an activity template, obtaining a comprehensive risk score through weighted summation, and mapping the comprehensive risk score into a corresponding early warning section; when the number of continuous frames in a preset time window reaches a preset frame threshold, and the comprehensive risk scores of each frame picture are all mapped into the same early warning section, triggering early warning information; calculating a behavior change amount and identifying an active correction action; and correlating and storing the early warning information, the active correction action and a quality measurement index as a photographer behavior portrait. The application realizes real-time detection of quality risks in a shooting process, feeds back the quality risks to photographers, and evaluates the ability of the photographers.

Description

Technical Field

[0001] This invention relates to the field of computer vision and image processing technology, and in particular to a real-time shooting quality warning method and system based on shooting behavior characteristics. Background Technology

[0002] In the existing architecture of live photo systems, the core focus of shooting, uploading, cloud processing, and filtering / display is on the transmission efficiency and automation capabilities after photo generation. For example, how to quickly synchronize photos from the camera to the cloud, and how to intelligently filter, differentiate, and enhance photos stored in the cloud. The quality of a photo is fixed the moment the shutter is pressed; quality control measures are implemented after imaging. Subsequent steps can only identify and remove defects in the existing product, without substantially changing the final image quality. However, the quality determinants during the shooting process are concentrated in the photographer's actions regarding framing, focusing, composition, exposure, and shutter release, as well as their real-time judgment of current lighting conditions and subject status. Existing technology cannot provide photographers with systematic warnings during shooting. Photographers cannot quantitatively perceive a series of quality defects, such as whether the camera's composition is significantly tilted, whether the focus is accurately locked onto the subject, or whether there is a serious risk of backlighting under current lighting conditions.

[0003] Current methods rely solely on the photographer's individual professional experience to make real-time adjustments and judgments on-site. However, under the dual pressures of a fast shooting pace and physical exertion, it's difficult to guarantee the accuracy, stability, and consistency of experience-based judgments, easily leading to unusable shots. When multiple camera positions are deployed at the same event site, and multiple photographers work collaboratively, the differences in their technical skills, camera angles, and on-site judgments are significant. The problem of delayed quality control is further amplified in this spatial dimension. The monitoring perspective for shooting remains at the level of isolated photos uploaded from a single device in a single session. Event organizers cannot grasp the distribution of shooting quality and abnormal trends from each camera position in real time during the event. There is a lack of real-time visualization of camera position status and quality, making it impossible to quantitatively assess the individual technical capabilities of photographers. Furthermore, it's impossible to differentiate the dimensional focus of different shooting activities, making it difficult to adapt event scheduling resources in advance and implement targeted interventions. Summary of the Invention

[0004] The purpose of this invention is to provide a solution that can detect potential quality risks in real time during the framing and shooting process of a photographer. By combining a multi-dimensional assessment, graded early warning and corrective feedback closed-loop mechanism, the quality control node is moved from the final film stage to the shooting behavior stage. At the same time, the corrective behavior of the photographer is quantitatively recorded to form a continuously evolving behavioral profile, thereby eliminating the inherent defects of the existing post-review mechanism in terms of real-time performance, scalability and behavioral quantification.

[0005] In a first aspect, embodiments of the present invention provide a real-time shooting quality warning method based on shooting behavior characteristics, comprising: Multi-source shooting behavior data is collected and activity categories are identified. Using the preview frame timestamp as the time reference, time alignment and data cleaning are performed to obtain standard shooting behavior data. The multi-source shooting behavior data is a multi-dimensional time-series dataset containing terminal status, framing environment and operation characteristics. Standard shooting behavior data is input into a multidimensional quality risk model to obtain normalized risk scores for each quality dimension; the multidimensional quality risk model is a set of models for normalized risk assessment of the image presentation dimension and shooting execution dimension of the shot; Obtain the activity template corresponding to the activity category, which stores the weight allocation parameters for each quality dimension, and perform a weighted summation on the normalized risk scores to obtain the comprehensive risk score; Based on a preset warning threshold, the comprehensive risk score is mapped to a corresponding warning zone. The warning zones include a non-warning zone, a first-level warning zone, a second-level warning zone, and a third-level warning zone, and the severity of the risk corresponding to each level of warning zone increases sequentially. When the number of consecutive frames within a preset time window reaches a preset frame threshold, and the comprehensive risk score of each frame is mapped to the same warning segment, a warning message is triggered. The difference between standard shooting behavior data in each shooting behavior dimension within a preset number of frames before and after the warning trigger time is calculated as the behavior change amount; if the behavior change amount exceeds the preset change threshold and automatic device adjustment is excluded, it is identified as an active correction action, and the quality metric index of the active correction action is calculated. Warning information, proactive corrective actions, and quality metrics are associated and stored as photographer behavior profiles. After each event, incremental updates are performed based on the newly added photographer behavior profiles for that event, and batch recalculation is performed on the stored data of the photographer behavior profiles when a preset period arrives.

[0006] Secondly, embodiments of the present invention also provide a real-time shooting quality early warning system based on shooting behavior characteristics, characterized in that it includes: The shooting behavior acquisition module is used to collect multi-source shooting behavior data and identify activity categories. It uses the preview frame timestamp as the time base to perform time alignment and data cleaning to obtain standard shooting behavior data. The multi-source shooting behavior data is a multi-dimensional time-series dataset that includes terminal status, framing environment and operation characteristics. The edge-side risk identification module is used to input standard shooting behavior data into a multi-dimensional quality risk model to obtain normalized risk scores for each quality dimension; the multi-dimensional quality risk model is a set of models for normalized risk assessment of the image presentation dimension and shooting execution dimension of the shooting scene. The continuous frame warning triggering module is used to obtain an activity template corresponding to the activity category and storing weight allocation parameters for each quality dimension; to perform a weighted summation of the normalized risk scores to obtain a comprehensive risk score; to map the comprehensive risk score to a corresponding warning segment based on a preset warning threshold, wherein the warning segment includes a non-warning segment, a first-level warning segment, a second-level warning segment, and a third-level warning segment, and the risk severity corresponding to each level of warning segment increases sequentially; when the number of consecutive frames within a preset time window reaches a preset frame number threshold, and the comprehensive risk score of each frame is mapped to the same warning segment, a warning message is triggered. The behavior profile update module is used to calculate the difference in standard shooting behavior data across each shooting behavior dimension within a preset number of frames before and after the warning trigger time as the behavior change amount; if the behavior change amount exceeds a preset change threshold and automatic device adjustment is excluded, it is identified as an active correction action, and the quality metric of the active correction action is calculated; the warning information, active correction action, and quality metric are associated and stored as a photographer behavior profile; after each event, an incremental update is performed based on the photographer behavior profile newly added for this event, and a batch recalculation is performed on the stored data of the photographer behavior profile when a preset period arrives.

[0007] The technical solution of this invention collects multi-source behavioral data and identifies activity categories. Using the preview timestamp as a time reference, it performs time alignment and data cleaning to obtain standard shooting behavior data, which can eliminate noise, outliers, and time asynchrony problems introduced during data collection. Then, the standard shooting behavior data is input into a multi-dimensional quality risk model, which transforms quality risk identification from "post-event manual detection" to "automatic calculation during the process," enabling quality risks to be quantitatively perceived while shooting is in progress. It also obtains activity templates, and after weighted summation, obtains a comprehensive risk score, achieving matching between evaluation standards and activity delivery requirements, making quality assessment scene-adaptive. Furthermore, the continuous frame confirmation mechanism effectively avoids false alarms caused by instantaneous interference.

[0008] The technical solution of this invention calculates the difference in standard shooting behavior data across various shooting behavior dimensions within a preset number of frames before and after the warning trigger time as the behavior change quantity. Values ​​exceeding a preset change threshold are identified as proactive correction actions. The system automatically detects whether the photographer has performed a correction action, distinguishes between proactive intervention and passive response by excluding automatic device adjustments, and then converts the proactive correction action into measurable data, thereby achieving closed-loop tracking of photographer behavior feedback. This invention also associates and stores warning information, proactive correction actions, and quality metrics as a structured photographer behavior profile. Through a dual-track maintenance strategy combining real-time incremental updates and periodic batch recalculation, the photographer behavior profile is updated immediately after each event to reflect the latest performance, while periodic weighted recalculation corrects accumulated errors, tilting the photographer's ability trend assessment towards recent performance, thus achieving quantifiable tracking of the photographer's capabilities. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating a real-time shooting quality early warning method based on shooting behavior characteristics according to the present invention.

[0010] Figure 2 This is a schematic diagram of the structure of a real-time shooting quality early warning system based on shooting behavior characteristics according to the present invention. Detailed Implementation

[0011] The present invention will be described in detail below through embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0012] In this invention, the terms "in one possible embodiment," "exemplary," or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "in one possible embodiment," "exemplary," or "for example" in this invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "in one possible embodiment," "exemplary," or "for example" is intended to present the relevant concepts in a specific manner.

[0013] In professional photo live streaming events, the quality control of the shooting relies on a manual review mechanism in the final stage. After the photos are uploaded to the server, each final photo is screened, unqualified photos are removed, and the photos are returned for reshoots.

[0014] The manual review mechanism suffers from a serious lag in discovering quality issues. By the time quality problems such as the subject being too small, the horizontal tilt, backlighting imbalance, or focus shift are discovered, the crucial moment of the photo has already passed. The photographer cannot revisit the shooting event that has already occurred, nor can they remedy the missed crucial moment, resulting in irreversible delivery losses.

[0015] An event with a scale of thousands of participants may generate tens of thousands of original photos. The human resource cost of manual review is extremely high. It is difficult to form a real-time feedback loop on the shooting quality of each photographer during the event, and it is also difficult to ensure the consistency of quality control.

[0016] Meanwhile, in existing technologies, the error patterns, common error types, and correction efficiency of photographers are not recorded in a structured manner, making it impossible to conduct data-driven capability assessments and differentiated job assignments for photography teams. This results in a lack of quantifiable growth curves for the delivery level of photography teams.

[0017] Therefore, how to detect quality risks in real time during the shooting process and provide feedback to the photographer, and how to evaluate the photographer's individual capabilities, are technical problems that urgently need to be solved in this field.

[0018] To address this, this invention presents a mechanism that enables photographers to instantly detect potential quality risks during framing and shooting. By combining a multi-dimensional assessment, tiered early warning, and corrective feedback closed-loop mechanism, the quality control point is moved from the final image stage to the shooting behavior stage. Simultaneously, the corrective behaviors of photographers are quantitatively recorded to form a continuously evolving behavioral profile, thereby eliminating the inherent defects of existing post-screening mechanisms in terms of real-time performance, scalability, and behavioral quantification.

[0019] Firstly, embodiments of the present invention provide a real-time shooting quality warning method based on shooting behavior characteristics, such as... Figure 1 As shown, it includes the following steps: S110 collects multi-source shooting behavior data and identifies activity categories. Using the preview frame timestamp as the time reference, it performs time alignment and data cleaning to obtain standard shooting behavior data.

[0020] For example, this embodiment uses an annual meeting celebration as an example to illustrate the overall process. The event organizer creates an "Annual Meeting Celebration" type event in the background in advance and configures a check-in area, stage area, audience seating, group photo area, exhibition booth and six outdoor camera positions.

[0021] Specifically, after the photographer powers on the camera, the acquisition program running on the shooting terminal immediately activates the six sensor channels. When the photographer half-presses the shutter button to frame the shot, the terminal reads the real-time preview frame and focus frame position from the viewfinder preview channel; reads the exposure parameters from the camera application interface; and reads the current latitude and longitude coordinates from the GPS module.

[0022] The inertial measurement unit (IMU) is a sensor module that integrates a gyroscope and an accelerometer, used to measure the terminal's attitude angle and angular velocity changes in space; the GPS module is a receiver of the Global Positioning System, used to obtain the latitude and longitude coordinates of the terminal's location.

[0023] Simultaneously, a scene classifier running locally on the camera terminal categorizes the activity list to which the current preview frame belongs. This scene classifier is a pre-trained lightweight convolutional neural network. The input is a low-resolution thumbnail of the preview frame. It extracts local features of the image through convolutional layers, reduces the dimensionality of the features through pooling layers, and finally outputs the classification probability through a fully connected layer, i.e., the probability distribution of the six camera positions. When the preview frame is identified as the "stage area," combined with the event type "celebration" selected by the organizer when the event was created, the event category label "celebration-stage" is obtained.

[0024] Next, since the collected data streams have different timestamps and frequencies, the system uses the timestamp sequence of the preview frame as the baseline time axis and aligns the other data streams according to the nearest neighbor principle: For a preview frame with timestamp t=0.033 seconds, the pose data is taken from the pose sample closest to that time, and the exposure parameters are taken from the exposure reading closest to that time. At the same time, abnormal data is removed to obtain standard shooting behavior data. Each record includes "timestamp, preview frame reference, focus frame coordinates, four exposure parameters, three pose angles, lens direction, position coordinates and scene label".

[0025] S120 inputs standard shooting behavior data into a multidimensional quality risk model to obtain normalized risk scores for each quality dimension.

[0026] Specifically, standard shooting behavior data is input into the local multidimensional quality risk model of the shooting terminal. The multidimensional quality risk model is a set of models that normalize the risk assessment of the image presentation dimension and shooting execution dimension of the shooting image. The sub-models built based on the image presentation dimension include the composition risk model, exposure risk model, focus risk model, person status risk model and occlusion risk model. The sub-models built based on the shooting execution dimension include the key link coverage model and the equipment anomaly risk model.

[0027] Each sub-model independently outputs a normalized risk score with a value range of [0,1]. The closer the score is to 1, the higher the risk severity of that quality dimension.

[0028] S130: Obtain the activity template corresponding to the activity category, which stores the weight allocation parameters for each quality dimension; perform a weighted summation of the normalized risk scores to obtain the comprehensive risk score.

[0029] Specifically, since the current activity category is "Celebration," a "Celebration" activity template is read from the template repository and loaded into the runtime configuration area. This template records the weight coefficients of the quality dimension corresponding to each sub-model and the comprehensive risk score. The calculation formula is as follows:

[0030] in, i This is the index for the quality dimension, with a value of ; For the first i The weight coefficients for each quality dimension, and satisfying and Normalization constraints; For the first i Normalized risk scores for each quality dimension.

[0031] S140, based on the preset early warning threshold, maps the comprehensive risk score to the corresponding early warning zone.

[0032] For example, in this embodiment of the invention, four warning zones are divided based on a preset warning threshold, as shown below: 1. When At that time, the currently captured frame is determined to be a non-warning segment; 2. When At that time, the currently captured frame is identified as a Level 1 warning zone; 3. When At that time, the currently captured frame is identified as a level-two warning zone; 4. When At that time, the current frame of the captured image is determined to be a level three warning zone.

[0033] S150: When the number of consecutive frames within a preset time window reaches a preset frame threshold and the comprehensive risk score of each frame is mapped to the same warning segment, a warning message is triggered.

[0034] In this embodiment of the invention, the warning is not triggered immediately because the comprehensive risk score of a single frame falls into the warning zone. Instead, the risk status of multiple consecutive frames is cumulatively confirmed by a sliding counter.

[0035] Specifically, the initial value of the sliding counter is zero. Whenever a new preview frame is acquired, the comprehensive risk score of the frame is mapped to the corresponding warning segment, and the warning segment mapped to the current frame is compared with the warning segment mapped to the previous frame: if the two belong to the same warning segment, the sliding counter increments by 1; if the two belong to different warning segments, the sliding counter is reset and starts counting again.

[0036] For example, in this embodiment of the invention, the preset time window and preset frame number threshold are adaptively adjusted according to the activity category and risk dimension. When the cumulative value of the sliding counter reaches the preset frame number threshold, and the time span elapsed to reach the threshold does not exceed the preset time window length, the continuous frame confirmation condition is determined to be met. Based on the warning segment into which the comprehensive risk score falls, a differentiated graded warning response is executed. Each warning segment corresponds to a different degree of risk severity and response strategy, specifically: 1. When the overall risk score falls into the first-level warning zone, it is determined that the current shooting image has quality issues such as slight tilt or slight misalignment that can be corrected in post-processing. At this time, auxiliary lines or edge prompt icons are superimposed on the viewfinder. The auxiliary lines can be golden ratio reference lines or horizontal reference lines to visually guide the photographer to make real-time fine adjustments to the composition in the current shooting image. The prompt is silent and non-intrusive.

[0037] 2. When the comprehensive risk score falls into the level 2 warning zone, it is determined that the current shooting scene has problems such as the subject being too small, backlighting, or faces being obscured, which affect the quality of the final image and need to be resolved on-site. At this time, a reminder is sent to the photographer via short text message or color box.

[0038] Short text messages such as "The subject is too small, we suggest increasing the focal length" and "Backlighting, please adjust the shooting angle" are displayed; color boxes are displayed as highlighted blocks overlaid on the edge of the viewfinder or around the subject area, using color coding to indicate the type of risk, such as red for exposure problems and yellow for composition problems.

[0039] 3. When the comprehensive risk score falls into the level three warning zone, it is determined that the current image has serious quality problems that are difficult to remedy in post-production, such as severe out-of-focus issues, large areas of overexposure or underexposure, closed eyes of key figures, and missing key elements. At this time, a strong reminder response is triggered: the vibration motor provides a long vibration or vibration sequence prompt, a warning voice is played, and a warning notification is sent to the remote monitoring terminal at the event site in a timely manner to coordinate with the photographer to reshoot or retake the shot.

[0040] Specifically, the remote monitoring terminal is equipped with a visual interface for an activity monitoring quality dashboard, which can display the data collection status and quality assessment results of each camera position in real time via a web interface or native application interface. The dashboard displays the monitoring data of all online camera positions in the form of tables combined with heat maps, and includes at least the following five information columns: camera position identifier column, online status column, warning zone column, link coverage column, and suggested operation column. For example, it could be "Camera position 1, online, non-warning zone, sufficient coverage, recommended to maintain".

[0041] When one of the cameras triggers a Level 3 warning consecutively or there is insufficient coverage of key areas, the system will automatically trigger a decision and push the scheduling decision to the front end, remind the person in charge, and provide targeted guidance to the photographer.

[0042] S160, calculate the difference in standard shooting behavior data in each shooting behavior dimension within a preset number of frames before and after the warning trigger time as the behavior change amount; if the behavior change amount exceeds the preset change threshold and automatic device adjustment is excluded, it is identified as an active correction action, and the quality metric index of the active correction action is calculated.

[0043] Specifically, after the warning information is triggered, a corrective action detection process is automatically initiated to determine whether the photographer has taken proactive adjustment measures in response to the quality risks indicated by the warning.

[0044] In this embodiment of the invention, the preset frame number is exemplarily set to three frames. Specifically, the difference between the standard shooting behavior data of the three frames before and the three frames after the warning trigger time is calculated to obtain the behavioral change amount of each shooting behavior dimension. This shooting behavior dimension includes at least the exposure parameter dimension, terminal posture dimension, focus parameter dimension, and camera position dimension. Next, the behavioral change amount of each dimension is compared with the preset change threshold for the corresponding dimension, and the presence of a correction action is identified based on the comparison results. Specifically, this includes the following five judgment methods: 1. Parameter adjustment detection: The exposure parameters include at least shutter speed, aperture value, ISO and exposure compensation value. If the change in any of the above exposure parameters exceeds the preset exposure change threshold, for example, if the exposure compensation change value exceeds 0.5EV, or the shutter speed exceeds one stop, it is determined that the photographer has performed exposure parameter adjustment correction.

[0045] 2. Composition Adjustment Detection: The terminal posture dimensions include at least pitch and roll angles. If the change in pitch or roll angle exceeds a preset posture change threshold, for example, if the change exceeds 1.5°, it is determined that the photographer has performed a composition angle adjustment correction. Furthermore, the offset of the subject detection frame center point relative to the center of the image or the golden ratio reference point is compared. If the offset decreases by more than a preset percentage of 20%, it is determined that the photographer has performed a subject recomposition.

[0046] 3. Focus adjustment detection: Compare the coordinates of the focus frame in the viewfinder before and after the warning. If the movement distance of the focus frame exceeds the preset focus movement threshold, such as the focus frame moving more than 5% of the screen width, it is determined that the photographer has performed refocus correction.

[0047] 4. Camera position adjustment detection: The camera position data is used to determine whether the shooting position has changed significantly. If the camera position moves more than the preset camera position movement threshold, for example, the camera position movement threshold can be set to 2 meters. If the overall risk score of subsequent frames decreases after the camera position is moved, it is determined that the photographer has performed a camera position adjustment correction.

[0048] 5. Reshoot Action Detection: When a Level 3 warning is triggered and the reason for the warning is the absence of a key step, monitor whether a qualified photo is added to that step within a preset effective period after the warning is issued. The preset effective period can be set to 15 minutes. If a qualified photo is added within this period, it is determined that the photographer has performed a reshoot correction.

[0049] S170 associates and stores warning information, proactive corrective actions, and quality metrics as photographer behavior profiles; after each event, it performs incremental updates based on the newly added photographer behavior profiles for that event, and performs batch recalculation of the stored data of photographer behavior profiles when the preset period arrives.

[0050] Specifically, early warning information, proactive corrective actions, and quality metrics are stored together in a structured document format, and photographer identification is written into the photographer's behavioral profile storage as the primary key.

[0051] After each event, an incremental update is performed based on the newly added photographer behavioral profiles from that event: 1. Record the early warning statistics, correction statistics, and capability indicator increments of this event in the photographer's historical behavioral profile record.

[0052] 2. Every Sunday at 12:00 AM in the preset cycle, the trend values ​​of each capability dimension are recalculated based on the cumulative data of the most recent twelve months with time decay weights, and the recalculation results are written back to the profile storage.

[0053] The technical solution of this invention collects multi-source behavioral data and identifies activity categories. Using the preview timestamp as a time reference, it performs time alignment and data cleaning to obtain standard shooting behavior data, which can eliminate noise, outliers, and time asynchrony problems introduced during data collection. Then, the standard shooting behavior data is input into a multi-dimensional quality risk model, which transforms quality risk identification from "post-event manual detection" to "automatic calculation during the process," enabling quality risks to be quantitatively perceived while shooting is in progress. It also obtains activity templates, and after weighted summation, obtains a comprehensive risk score, achieving matching between evaluation standards and activity delivery requirements, making quality assessment scene-adaptive. Furthermore, the continuous frame confirmation mechanism effectively avoids false alarms caused by instantaneous interference.

[0054] The technical solution of this invention calculates the difference in standard shooting behavior data across various shooting behavior dimensions within a preset number of frames before and after the warning trigger time as the behavior change quantity. Values ​​exceeding a preset change threshold are identified as proactive correction actions. The system automatically detects whether the photographer has performed a correction action, distinguishes between proactive intervention and passive response by excluding automatic device adjustments, and then converts the proactive correction action into measurable data, thereby achieving closed-loop tracking of photographer behavior feedback. This invention also associates and stores warning information, proactive correction actions, and quality metrics as a structured photographer behavior profile. Through a dual-track maintenance strategy combining real-time incremental updates and periodic batch recalculation, the photographer behavior profile is updated immediately after each event to reflect the latest performance, while periodic weighted recalculation corrects accumulated errors, tilting the photographer's ability trend assessment towards recent performance, thus achieving quantifiable tracking of the photographer's capabilities.

[0055] exist Figure 1 Based on the illustrated embodiment, as one implementation of this invention, multi-source shooting behavior data during the shooting process is simultaneously collected through multiple sensor channels of the terminal device, and the activity category is identified. The multi-source shooting behavior data includes preview frames, focus frame positions, exposure parameters, terminal posture data, lens direction, and camera position.

[0056] Specifically, the image data of the preview frame is read in real time from the viewfinder preview channel at a first sampling frequency. The preview frame is a pre-processed image stream output by the camera image signal processor, which has not yet been encoded or compressed and retains complete YUV color information.

[0057] The focus frame position is obtained based on the callback triggering method of the autofocus system (AF). The autofocus system refers to the functional module inside the camera that automatically completes focus recognition and lens driving through phase detection or contrast detection principles. After the autofocus system completes a focus scan and locks the focus, it actively calls back and reports the current coordinate position of the focus frame in the viewfinder. The coordinate position is in the form of a rectangle, including the coordinates of the top left corner and width and height information. To avoid time deviation between the focus frame position and the preview frame image, the sampling period of the focus frame position is configured to be less than a preset time threshold.

[0058] Exposure parameters, consisting of shutter speed, aperture, ISO, and exposure compensation, are read from the camera application interface at a second sampling frequency. Specifically, shutter speed determines the length of time the sensor receives light, aperture determines the size of the light-gathering aperture, ISO determines the sensor's response speed to light, and exposure compensation reflects the photographer's adjustment of the main light direction based on the automatic metering results.

[0059] The terminal attitude data, consisting of pitch, roll, and yaw angles, is read from the inertial measurement unit at a third sampling frequency. Specifically, the pitch angle describes the terminal's pitch rotation angle around the horizontal axis, the roll angle describes the terminal's left and right roll angle around the vertical axis, and the yaw angle describes the terminal's horizontal orientation angle around the vertical axis. The camera orientation switching event is also read using an event-triggered method; the acquisition program captures this event and records the current camera orientation.

[0060] The camera position is read at a fourth sampling frequency. The third sampling frequency is higher than the first sampling frequency, and the first sampling frequency is higher than the second and fourth sampling frequencies. When the terminal's computing power is lower than a preset computing power threshold, the first sampling frequency is adaptively reduced to a preset minimum frame rate.

[0061] For example, the above frequency relationships are configured differently according to the physical change characteristics of various types of data: posture data reflects the photographer's hand shake, and the third sampling frequency can be set to a value between 50Hz and 100Hz; the preview frame change frequency matches the persistence of human vision, and the first sampling frequency is set to 10Hz to 30Hz; exposure parameters are generally actively adjusted by the photographer, and the change interval is calculated in seconds, so the second sampling frequency can be set to 1Hz; the GPS position remains basically unchanged at the same camera position, so the fourth sampling frequency can be set to a low-frequency sampling of 1Hz.

[0062] Based on the scene classifier built into the acquisition device, the current framing environment is classified in real time and labeled with the corresponding activity category. For each quality dimension under the current activity category, the number of samples collected by the sensor is recorded; when the number of samples is lower than the preset sample size threshold, a framing supplement prompt is triggered to guide the photographer to increase the framing time of this type of scene to ensure the representativeness of the samples for subsequent portrait statistics.

[0063] Next, the coefficient of variation is calculated for the sensor data corresponding to a preset number of consecutive frames. When the coefficient of variation exceeds the preset variation threshold, the current acquisition environment is determined to be unstable. For example, if a photographer is quickly chasing a fast-moving subject and the posture data changes in distance within a short period of time, the risk judgment weight within the time window corresponding to the preset number of consecutive frames is reduced by a preset amount. This can prevent the instantaneous risk false alarm caused by violent shaking from being upgraded to a formal warning.

[0064] The coefficient of variation is the ratio of the standard deviation to the absolute value of the mean, used to measure the relative fluctuation of data within a time window.

[0065] This invention achieves an optimal balance between acquisition accuracy and computing power consumption by setting the sampling frequency of each sensing channel differently according to its physical change characteristics; the adaptive frequency reduction mechanism ensures that this solution is deployable on shooting terminals with different performance levels; and by monitoring the sample size and dynamically reducing the weight of the coefficient of variation, it overcomes the inherent defects of traditional single fixed frequency acquisition solutions in terms of insufficient coverage of scene diversity and rampant false alarms due to motion blur.

[0066] exist Figure 1 Based on the illustrated embodiment, as one implementation of this invention, time alignment and data cleaning are performed using the preview frame timestamp as the time base, specifically including the following steps: Using the timestamp sequence of the preview frames as the reference timeline, the nearest neighbor interpolation method is used to align the focus frame position, exposure parameters, terminal attitude data, lens direction, and camera position to the reference timestamp, with the alignment tolerance not exceeding the preset time deviation range.

[0067] It should be noted that because the preview frame timestamp comes from the acquisition time marker output by the camera image signal processor, it has the highest time reference accuracy in the sensing channel system. Furthermore, the risk assessment of all subsequent image presentation dimensions uses the preview frame as the main input, which can ensure that the image data is aligned with other auxiliary data without deviation.

[0068] The data cleaning process includes the following steps: 1. Apply a moving average filter to the inertial measurement unit angular velocity sequence in the terminal attitude data. Slide the filter across the angular velocity sequence gradually with a preset window length. Use the arithmetic mean of all sampled values ​​within each window as the filtered output for the current window. The mathematical expression is:

[0069] in, k For the target index of timestamp, for k Filtered output at time t, L For window length, For the original angular velocity sequence in j The sampled value at time.

[0070] 2. Apply Kalman filtering to the position of the aircraft. Use the plane coordinates and moving speed components of the aircraft position as state variables. Construct a state transition matrix based on the uniform motion model. Perform prediction at each sampling time to obtain the prior estimate. Then, use the new sampled values ​​to weight and correct the prior estimate to obtain the posterior estimate.

[0071] Specifically, assuming state variables ,in The coordinates of the machine position plane are... For the velocity component, the state transition matrix As shown below:

[0072] in, The time interval between two adjacent GPS samples, in this embodiment of the invention It takes 1 second.

[0073] Next, the prior estimate of the current time step is derived from the state of the previous time step. :

[0074] in, for Posterior estimation at time step.

[0075] The prior estimate is weighted and corrected using the current GPS sample values ​​to obtain the posterior estimate. :

[0076] in, Here is the weighting coefficient matrix for the Kalman gain, with values ​​ranging from 1 to 2. ; These are GPS sample values; H The observation matrix is ​​the matrix that maps the state vector to the measurement space.

[0077] 3. Perform two anomaly checks on the aligned data for each frame: First, if a frame reads a sensitivity value that exceeds the sensitivity range allowed by the terminal hardware specifications... If a sensor is deemed to be abnormal, the frame will be removed from the standard behavior data. Second, if the absolute value of the angular velocity output by the inertial measurement unit exceeds a preset angular velocity threshold, it is determined that the sensor is abnormal and the frame is discarded. In this embodiment, the angular velocity threshold is set to... , which is the angular velocity of the human hand holding the camera.

[0078] 4. If, within a preset time window of 2 seconds, the duration of missing valid sampling data for any sensor dimension exceeds a preset missing duration threshold, the data will be classified as uncertain based on the comprehensive risk score corresponding to that time window.

[0079] This invention addresses the issue of time asynchrony between images and sensor data caused by misaligned timestamps across multiple sensor channels by establishing a unified reference timeline using preview frame timestamps. It also improves the signal-to-noise ratio of the data for subsequent risk models by applying targeted signal processing to high-frequency attitude jitter and low-frequency GPS drift using sliding mean filtering and Kalman filtering, respectively. Furthermore, the fallback mechanism of anomaly removal and integrity verification avoids misjudgment risks in scenarios with single-point anomalies and missing data, ensuring that the quality of standard shooting behavior data reaches the stability level required for subsequent multi-dimensional evaluation.

[0080] exist Figure 1 Based on the illustrated embodiment, as one implementation of this invention, obtaining an activity template corresponding to the activity category and storing weight allocation parameters for each quality dimension specifically includes: Different activity templates are pre-configured for different activity categories and stored in the template repository. Each activity template contains a set of weight allocation parameters and a set of stage configurations.

[0081] The weight allocation parameter set consists of weight coefficients corresponding to each quality dimension. Each weight coefficient is greater than zero and the sum of all weight coefficients is equal to 1. The weight coefficients are configured differently according to the emphasis on delivery quality for each activity category.

[0082] For example, different activity types exhibit fundamentally different sensitivities to different quality dimensions. The table below shows the weighting coefficients for each activity type across various risk dimensions:

[0083] The stage configuration set consists of the names of the stages included in the activity, the minimum number of qualified coverage shots for each stage, and the start and end time information of each stage. It is used to provide a benchmark for the coverage risk model judgment of key stages in the shooting execution dimension. It explicitly defines the minimum coverage standard and time boundary for each stage, so that "whether the shooting progress of the current stage is behind schedule" is transformed from a subjective feeling into a quantifiable risk indicator.

[0084] This invention provides a differentiated weighting mechanism to match evaluation strategies with delivery requirements for different activity categories; and provides an explicit benchmark for determining the shooting execution dimension through a process configuration set, making activity progress risks "feeling after the fact" and "measurable in real time".

[0085] exist Figure 1 Based on the illustrated embodiment, as one implementation of this invention, when the number of consecutive frames within a preset time window reaches a preset frame threshold, and the comprehensive risk score of each frame is mapped to the same warning segment, a warning message is triggered, specifically including the following steps: Step a1: Obtain the time window length and frame threshold corresponding to the current activity category.

[0086] The time window length and frame count threshold are determined based on the category of the current activity; the time window length is negatively correlated with the scene dynamics corresponding to the activity category, while the frame count threshold is positively correlated with the scene dynamics.

[0087] Specifically, in highly dynamic scenarios, the risk state itself changes frequently, requiring a rapid response to short-lived issues within a short time window, and setting a large frame rate threshold to avoid instantaneous false alarms caused by motion blur, thus triggering false warnings.

[0088] Step a2: Within the time window, read and count the number of frames whose comprehensive risk scores are continuously mapped to the same warning segment; when the number of consecutive frames reaches the frame number threshold, trigger the warning information for that warning segment.

[0089] For example, the warning information could be "Backlit face: The brightness of the face area is lower than the background brightness for 5 consecutive frames"; "Focus deviation: The focus frame does not fall on the subject area and the subject edge has low sharpness".

[0090] Step a3: If, during the statistical process, the comprehensive risk score of one frame corresponds to a different warning segment than the previous frame, then the current accumulated number of consecutive frames is cleared to zero and the number of consecutive frames is re-accumulated from that frame.

[0091] Only when the risk state remains stable in the same segment over multiple consecutive frames can the current quality problem be identified as a persistent risk rather than a transient disturbance.

[0092] Step a4: When the normalized risk score of any quality dimension exceeds the preset mandatory upgrade threshold, the warning segment corresponding to the current warning information is upgraded by one level in the direction of increasing risk severity, and the upgraded warning segment does not exceed the level three warning segment.

[0093] The purpose of this embodiment is to avoid the technical problem that serious defects in a single dimension may be diluted during the weighted summation process in actual situations, resulting in the comprehensive risk score failing to reach the appropriate warning range. This embodiment ensures that serious single-point problems are not diluted by the comprehensive score by setting a mandatory upgrade threshold for a single dimension.

[0094] The embodiments of the present invention achieve the best balance between early warning sensitivity and stability under different activity types by adaptively adjusting the time window and frame number threshold according to the dynamic degree of the scene; the count reset mechanism during segment transition strictly ensures the validity of continuous evidence and avoids misjudgment caused by cross-segment accumulation.

[0095] exist Figure 1Based on the illustrated embodiment, as one implementation of this invention, the multidimensional quality risk model includes a composition risk model, an exposure risk model, a focus risk model, a subject status risk model, and an occlusion risk model constructed based on the image presentation dimension. Specifically, it includes: First, the composition risk model takes the preview frame, the coordinates of the subject detection box, and the terminal's pose angle as input.

[0096] (1) If the proportion of the subject detection frame area to the total area of ​​the image is lower than the preset proportion threshold, it is judged as a risk of subject missing, indicating that the photographer has failed to effectively highlight the subject. In this embodiment of the invention, the preset proportion threshold is set to 0.08, and the specific formula is as follows:

[0097] in, The proportion of the main detection frame area to the total image area. The width of the main detection frame. The height of the main detection frame, The width of the entire frame. The height of the entire frame.

[0098] Normalized score of subject missing risk for:

[0099] (2) Center point of the main detection frame Normalized Euclidean distance between the image and the preset composition reference point An offset exceeding a preset offset threshold is considered a risk of subject offset; in this embodiment of the invention, the preset offset threshold is set to 0.35, and the golden ratio point is used as the composition reference point. Euclidean distance is the straight-line distance between two points in a plane, and the specific formula is as follows:

[0100] when When this occurs, it is determined to be a subject offset risk; the normalized score of the subject offset risk. The formula is:

[0101] (3) The absolute value of the horizontal tilt angle measured by the roll angle A tilt angle exceeding a preset tilt threshold is considered a risk of horizontal tilt. The roll angle directly reflects the degree of tilt of the terminal relative to the horizontal line; the larger the absolute value, the more obvious the screen tilt. In this embodiment of the invention, the preset tilt threshold is set to 2.5°, and the specific formula is as follows:

[0102] in, The normalized score is the risk of horizontal tilt. Based on the tilt range of common handheld shooting, tilt angles of 12.5° and above are considered to be severely disrupted in terms of image balance. Therefore, in this embodiment, the denominator is set to 10°, which can map angle differences of 12.5° and above to the corresponding risk value of 1.

[0103] (4) If the coordinates of the key point of the subject exceed the boundary of the screen and the confidence of the key point is higher than the preset confidence threshold, it is judged as the risk of the subject going out of bounds. This can avoid false alarms of going out of bounds due to the low confidence output of the key point detection itself.

[0104] For example, for the 68 detected facial key points, if the coordinates of any point exceed the image boundary and the confidence score of that point is greater than 0.8, then an out-of-bounds risk is triggered; the normalized score of the subject out-of-bounds risk is... It is a binary value, which can take the value 1 if it is out of bounds, or the value 0 if it is not out of bounds.

[0105] After normalizing the risks of missing subject, subject offset, horizontal tilt, and subject exceeding boundaries, the maximum value among these sub-risks is taken as the composition risk score. :

[0106] In this embodiment, the composition risk score The feature extraction network is obtained through a pre-trained lightweight neural network classifier, which can use the MobileNetV3-Small classifier as the backbone feature extraction network.

[0107] For example, a total of 50,000 sample images containing normal composition are collected, and 5,000 sample images of each type of defect are collected for the aforementioned various composition defect types. At the same time, the images are labeled according to the preset composition scoring criteria to generate binary labels, namely normal or defective.

[0108] Next, the original fully connected layer at the top of MobileNetV3-Small is removed and replaced with a global average pooling layer and a fully connected output layer with a Softmax activation function. This outputs the probability distribution of whether the current captured image is "normally composed" or "abnormally composed," and the confidence score of the "abnormally composed" category is taken as the composition risk probability of the current captured image. .

[0109] During training, the prediction error is calculated using the cross-entropy loss function, and gradient descent updates are performed using the Adam optimizer. Furthermore, to balance the precision and recall of various risk indicators, this embodiment employs a grid search-based threshold optimization method to determine the threshold for each sub-item. In each round of the search, the current candidate threshold combination is applied to the validation set, and the threshold with the largest harmonic mean (F1-score) of precision and recall is statistically identified as the final risk assessment parameter for the current model.

[0110] Second, the exposure risk model uses the preview frame brightness histogram and the average brightness of the face region. Average brightness of the background area and exposure compensation EV For input.

[0111] (1) The Y channel brightness value is higher than the highlight threshold. pixel ratio An overexposure level exceeding a preset overexposure threshold is considered a risk of highlight clipping. The specific formula is as follows:

[0112] in, k This is a brightness level index, with a value range of [value range missing]. ; Indicates brightness value k The proportion of pixels in the entire image, satisfying .

[0113] For example, take the highlight threshold. The value is 235, and the preset overexposure threshold is 0.12. If this is detected, it is considered a risk of highlight clipping, indicating that the area in the final image will appear pure white and lose detail.

[0114] (2) The brightness of the Y channel is lower than the dark area threshold. A pixel percentage higher than the preset underexposure threshold is considered a risk of completely black shadows. The specific formula is as follows:

[0115] For example, take the dark area threshold. The value is 20, and the preset underexposure threshold is 0.15. When this is detected, it is considered a risk of complete black in the shadows, indicating that the area in the final image will appear pure black and the details cannot be recovered in post-production.

[0116] (3) Calculate the difference between the average brightness of the background and the average brightness of the face. If the exposure value exceeds the preset backlight threshold and the exposure compensation is negative, it is considered a backlight face risk. The specific formula is as follows:

[0117] For example, a preset backlight threshold of 80 is set, that is, when When a backlit face is identified as a risk, the brightness of the background relative to the face can be measured, and the situation can be ruled out by the photographer and by manually applying positive exposure compensation to correct the backlighting.

[0118] (4) Calculate the information entropy of the Y channel brightness histogram E The uniformity of brightness distribution is characterized by the following formula:

[0119] in, It is a very small positive number. For example, when... If the exposure distribution is deemed unreasonable, it indicates an abnormal risk of histogram entropy value.

[0120] Furthermore, the sigmoid function is used to normalize the risks of each sub-item and take the maximum value as the exposure risk score. The specific formula is shown below:

[0121] in, This is the standard Sigmoid function, with a value range of 0 to 1; For sensitivity parameters, and , The larger the value, the steeper the jump near the threshold; for i The original value of the sub-item, In order, these refer to the risk of highlight overflow. Risk of dark areas becoming completely black Backlighting risk to faces and the risk of abnormal histogram entropy values ; for i The risk trigger threshold corresponding to the sub-item; the output The normalized values ​​of the risk scores for each sub-item, including the normalized highlight spillover risk score. Normalized dark area dead black risk score Normalized backlight face risk score and the abnormal risk score of normalized histogram entropy value .

[0122] The formula for the exposure risk score is as follows:

[0123] The exposure risk model in this embodiment uses Efficient-B0 as the basic network architecture. Efficient is a series of efficient convolutional neural networks that are scaled in a balanced manner across the network depth, width and input resolution using composite scaling factors. B0 is the smallest basic version in this series, suitable for deployment on the local shooting terminal.

[0124] Next, the MIT-Adobe FiveK dataset was used as a sample. This dataset contains 5,000 RAW format photos taken by professional photographers using high-end cameras, covering various scenes such as indoor, outdoor, and portrait. Each sample photo was labeled as one of four categories: "normal", "overexposed", "underexposed", and "backlit".

[0125] The 20,000 labeled samples were divided into training, validation, and test sets in an 8:1:1 ratio. The Adam optimizer was used with an initial learning rate set, employing a cosine annealing learning rate scheduling strategy. Training consisted of 100 epochs with a batch size of 64. An early stopping strategy was implemented: training was terminated if the validation set loss did not decrease for 10 consecutive epochs. After training, the model weights were quantized to INT8 format to further reduce the model size. The quantized model size was approximately 3MB, with a single-frame inference time of approximately 20 milliseconds on a mobile CPU.

[0126] Third, the focus risk model takes the preview frame, focus frame coordinates, and autofocus status indicator as input.

[0127] (1) Edge gradient magnitudes within the focus frame region that are lower than a preset gradient threshold are considered to be at risk of low contrast. Specifically, the Sobel edge gradient is calculated within the image region covered by the focus frame. First, a convolution is applied to the Y channel image I to obtain the gradient components, as shown in the following formula:

[0128] in, and Coordinates The horizontal and vertical gradient responses at the location; Represents a two-dimensional convolution operation; For the horizontal difference kernel of the Sobel operator, This is the vertical difference kernel of the Sobel operator.

[0129] Gradient mean within the focus area The specific formula is as follows:

[0130] in, This represents the total number of pixels within the focus frame area. For example, a preset gradient threshold of 30 is used. If the condition is identified as low contrast risk, it indicates that the edge response within the focus area is too weak and the image may not be in focus.

[0131] (2) If the intersection-union ratio (IU) of the focus frame and the subject detection frame is lower than a preset IU threshold, it is considered a risk of focus deviation from the subject. Specifically, the area of ​​the focus frame is calculated. Area of ​​the main detection frame intersection ratio :

[0132] For example, a preset intersection-union (IU) threshold of 0.3 is set when... If the focus is off-center, it indicates that the focus point is not on the subject and the subject is out of focus.

[0133] (3) If the image sharpness metric value of the focus frame area is lower than a preset sharpness threshold, it is considered a risk of focus blur. Specifically, for the Y channel image of the focus frame area... I Apply a Laplace transform convolution to obtain the response map, and calculate its variance. The specific formula is as follows:

[0134] in, For Laplace's response in The mean within the region, This is the kernel of the Laplacian operator. For example, a preset sharpness threshold of 150 is used. The system determines the risk of focus blur. A sharp image exhibits rich high-frequency details, a dramatic change in the Laplacian response, and a large variance; a blurred image has smoothed high-frequency details, a gradual change in the Laplacian response, and a small variance.

[0135] This invention uses two publicly available image quality databases, TID2013 and LIVE, as base samples, and additionally collects 10,000 sets of live-streamed photo scene focus samples as training samples, of which 5,000 sets are labeled as in focus and 5,000 sets are labeled as out of focus, to train the XGBoost binary classifier.

[0136] First, construct a four-dimensional feature vector for each training sample. ,in The four-dimensional vector representing the autofocus state after one-hot encoding is concatenated with the first three risk scores to obtain a seven-dimensional feature vector.

[0137] Next, the labels for in-focus samples are set to 0, and the labels for out-of-focus samples are set to 1. The four-dimensional feature vectors are then... Input a pre-trained XGBoost binary classifier and fuse each feature in a non-linear manner to obtain the focus risk probability. That is, the focus risk score Its value range is .

[0138] During training, the binary cross-entropy loss function is used to train the hyperparameters to optimize the classification probability and enable the classifier to have good probability calibration.

[0139] Fourth, the risk model for human status uses time series of facial regions, facial key points, and eye aspect ratio as input.

[0140] (1) An eye aspect ratio lower than a preset closed-eye threshold for a continuous frame count reaching a preset closed-eye frame count is considered a risk of closed-eye. Specifically, the eye aspect ratio is calculated from the coordinates of six key points around the eyes output by the face key point detector. These six key points are denoted as... ,in, For the left corner of the eye, For the right corner of the eye, and These are two points on the upper eyelid. and The two points on the lower eyelid are the aspect ratio of the eye. The definition of is:

[0141] in, This represents the Euclidean distance between two points. For example, a preset eye-closing threshold of 0.18 is set, and a preset number of eye-closing frames is set to 3. Furthermore, if this state persists for 3 consecutive frames, it is considered a risk of eye closure.

[0142] (2) The sum of probabilities of preset negative expression categories exceeding a preset expression threshold is judged as an abnormal expression risk. Specifically, the face region is input into a pre-trained facial expression recognition model, which outputs the probability distribution of seven expression categories. These seven categories correspond one-to-one with anger, disgust, fear, happiness, sadness, surprise, and neutrality, and satisfy the following conditions: , where c is the target index for the seven types of expressions.

[0143] Negative expression probability The sum of the probabilities of disgust, fear, and sadness: .

[0144] For example, a preset expression threshold of 0.5 is set when... It was determined to be an abnormal expression risk.

[0145] (3) A standard deviation exceeding a preset line-of-sight threshold for the yaw angles of multiple heads is considered a risk of inconsistency between the viewing angle and the line of sight. Specifically, when multiple faces are detected in the image, i.e., the number of faces is considered high. Greater than 2, independently estimate the yaw angle of head pose for each face. And calculate the standard deviation of the yaw angle of multiple faces. :

[0146] in, The mean yaw angle for each face.

[0147] For example, a preset viewing threshold of 15° is set. The time was determined to be a risk of inconsistency between the film and television production lines.

[0148] Next, the raw risk values ​​of the above three sub-items are normalized using sigmoid and the maximum value is taken as the character's status risk score. .

[0149] This invention employs Mini-Xception as the base network and uses two publicly available facial expression datasets, AffectNet and RAF-DB, as base samples. The MTCNN face detector can be used to crop facial regions from images and align them to standard poses.

[0150] Two parallel output branches are set at the end of the Mini-Xception network: the expression classification branch outputs a seven-dimensional softmax probability distribution; the eye aspect ratio regression branch outputs a single-value EAR prediction. The predicted target value of the EAR is calculated offline on the training set by the face landmark detector and used as the regression label. The total loss function is the weighted sum of the two branches.

[0151] Fifth, the occlusion risk model takes the preview frame, face detection box, and semantic segmentation results as input.

[0152] (1) If the intersection-union ratio of the segmented region of the occlusion category and the face detection box is higher than the preset occlusion threshold, it is considered as a face occlusion risk.

[0153] Specifically, the preview frame is input into the semantic segmentation network BiSeNet, which outputs a segmentation label map of the same size as the original image. Where C is the total number of semantic categories, each pixel is classified into a specific semantic category, and a set of occlusion categories is defined. When the intersection-union ratio of the occlusion area and the face detection box is greater than the preset occlusion threshold of 0.1, it is judged as a face occlusion risk.

[0154] (2) Low variance regions that appear continuously at a fixed position in multiple consecutive frames are identified as lens smudge risk. Specifically, inter-frame difference is performed on consecutive frames. If a low variance region appears continuously at a fixed position in the figure, with a variance of less than 50 and the low variance state persists for more than 10 consecutive frames, the position is identified as a lens smudge risk.

[0155] This invention uses the COCO-Stuff dataset as the base sample, combined with 5000 photos of live streaming business scenarios as training samples.

[0156] Next, the two sub-items are normalized, and the maximum value is taken as the occlusion risk score. .

[0157] The five independent sub-models designed in this invention provide a comprehensive evaluation of the image presentation dimensions. Through a reduction strategy that takes the maximum value after independent evaluation of multiple sub-items within each sub-model, it comprehensively covers all risks related to shooting quality, a feature unique in the field. Furthermore, the design of the thresholds set in this invention, which can be trained and optimized based on labeled datasets, makes the model's judgment criteria more universal and accurate than fixed rules based on pure experience.

[0158] exist Figure 1 Based on the illustrated embodiment, as one implementation of this invention, the multidimensional quality risk model further includes a key link coverage risk model and an equipment anomaly risk model constructed based on the shooting execution dimension, specifically including the following two points: First, the key link coverage risk model takes the current link identifier, the number of qualified photos taken in the current link, the minimum number of photos to be covered in the current link, and the duration of the current link as inputs, and uses the ratio of the number of qualified photos taken to the minimum number of photos to be covered as the coverage completion rate.

[0159] When the duration of a process exceeds the preset high-risk ratio of 70% and the coverage completion rate is lower than the preset high-risk coverage threshold of 50%, a high coverage risk is output. When the coverage completion rate is lower than 80% of the preset medium-risk coverage threshold and the remaining time of the process is lower than 5 minutes of the preset remaining time threshold, the medium coverage risk is output; and the high coverage risk and medium coverage risk are normalized into the key process coverage risk score.

[0160] Second, the device anomaly risk model uses terminal battery power, remaining storage space, network upload latency, and camera module temperature as input indicators. When any of these indicators exceeds the corresponding safe range, it is determined to be a device anomaly risk.

[0161] This invention transforms activity progress risks from traditional manual experience-based judgments into quantifiable real-time indicators through a key link coverage risk model; and incorporates equipment operating status into a unified risk assessment framework through an equipment anomaly risk model, thus avoiding interruptions and losses caused by equipment problems during critical shooting moments.

[0162] exist Figure 1Based on the illustrated embodiment, as one implementation of the present invention, a quality metric for active correction actions is calculated. The quality metric includes correction response time, correction effectiveness, quality improvement degree, and correction thoroughness.

[0163] First, the difference between the moment when the active correction action is detected and the moment when the corresponding warning information is triggered is used as the correction response time to characterize the photographer's response speed to the warning.

[0164] Second, the number of times the warning for the same quality dimension was not triggered again within the preset validity period after the proactive corrective action was taken is accumulated, and the ratio of this number to the total number of proactive corrective actions is used as the correction effectiveness rate, which is used to characterize the degree to which the corrective action suppresses the same type of risk.

[0165] Specifically, a preset validity period timer is started after each correction action. In this embodiment, the validity period is set to 30 minutes. If the warning for the same quality dimension is not triggered again within the validity period, the correction is recorded as a valid correction. At the end of this activity, the number of valid corrections is accumulated and divided by the total number of corrections to obtain the correction effectiveness rate.

[0166] Third, the difference between the comprehensive risk score before the warning is triggered and the comprehensive risk score after the corrective action is completed under the same camera position or the same scenario is used as the quality improvement degree. A positive difference indicates that the risk has decreased after the correction.

[0167] Fourth, count the number of times the same risk label is triggered in a single event, and use the reciprocal of the number of triggers after normalization as the degree of correction. The fewer the number of triggers, the higher the degree of correction.

[0168] The embodiments of this invention construct a complete quantitative evaluation system for corrective behaviors from four perspectives: response time, effectiveness, degree of improvement, and thoroughness. This system overcomes the crude model of traditional photographer competence assessment that relies on subjective scoring and random observation, and provides a traceable and comparable data foundation for subsequent portrait accumulation and personalized training suggestions.

[0169] exist Figure 1 Based on the illustrated embodiment, as one implementation of this invention, the warning information, proactive correction actions, and quality metrics are stored in a structured document format. The photographer's identifier is written into the photographer's behavioral profile storage using the photographer's identifier as the primary key. Each structured document of the photographer's behavioral profile includes the following data levels: 1. Basic information layer, including photographer ID, registration time, commonly used equipment models, and commonly used focal length range.

[0170] 2. Capability index layer: The historical scores of each capability dimension are stored in the form of time-series array and are calculated monthly. The capability dimensions include at least backlight handling capability, focus stability, composition stability, control of closed eyes in group photos, coverage of key links and speed of problem correction. In addition, the trend data of each capability dimension is retained for the most recent 12 months.

[0171] 3. Error mode layer, which is a list of warning types sorted by trigger frequency. Each warning type records the warning type identifier, trigger frequency, and the corresponding correction effectiveness rate.

[0172] 4. Activity history layer: records the types of activities the photographer has participated in, the number of events, the total number of photos taken, and the overall pass rate.

[0173] exist Figure 1 Based on the illustrated embodiment, as one implementation of the present invention, the photographer's behavioral profile is stored in a three-layer hierarchical storage consisting of a hot data layer, a warm data layer, and a cold data layer.

[0174] First, the hot data layer: uses an in-memory database to store real-time snapshots of the current activity. An in-memory database is a database system in which the data subject resides in memory rather than on disk. Its query response time can reach the millisecond level, making it suitable for real-time monitoring scenarios that require high-frequency read and write operations.

[0175] The real-time profile snapshot includes the early warning trigger count and real-time correction effectiveness rate of the event. After each early warning trigger or correction action is detected, the corresponding field of the hot data layer is incremented or updated in real time, and it is retained for a preset time after the event ends before becoming invalid. In this embodiment, the retention time is set to 24 hours.

[0176] Second, the warm data layer: The full set of structured portrait data is stored in a relational database, with tables partitioned by photographer identifier. This relational database refers to a database system that organizes data in tabular form and supports complex filtering and sorting through structured query statements.

[0177] The full-scale profile structured data includes four data layers as described in the above embodiment: the basic information layer, the historical score time series array of each capability dimension of the capability index layer, the list of warning types and correction effectiveness rate sorted by trigger frequency of the error mode layer, and the activity type, number of sessions, total number of images and overall pass rate of the activity history layer.

[0178] Furthermore, the warm data layer supports conditional queries and sorting by capability and activity type dimensions. For example, one can query "a list of photographers whose composition stability energy level is greater than 0.8 in the last three months" or "the top ten photographers in terms of effectiveness in correcting group photo scenes," enabling filtering and sorting across multiple fields.

[0179] Third, the cold data layer: This layer uses object storage to store historical raw shooting logs and a distributed storage system that stores unstructured data in key-value pairs. Its storage cost is far lower than relational databases and it is suitable for long-term archiving of large volumes of data. Historical raw shooting data is partitioned and stored by combining shooting time and photographer identification, for example, using "year / month / photographer identification / " as the partition path. This facilitates quick data location by time or photographer dimension and automatically archives the data to low-cost storage media after a preset retention period, further reducing the cost burden of long-term storage.

[0180] Incremental updates write the event data to the hot data layer and the warm data layer after the event ends. Batch recalculation reads the cumulative data from the warm data layer and the cold data layer at a preset period and writes the recalculation results back to the warm data layer.

[0181] The collaborative relationship between the above three-tier storage architecture and the dual-track update strategy is as follows: 1. Incremental updates are performed after the event ends, simultaneously writing the event data to both the hot data layer and the warm data layer: the hot data layer updates the real-time count and effectiveness field of the current event; the warm data layer appends the warning trigger records, correction action records, pass rate statistics and event history generated by the event to the corresponding photographer's data table.

[0182] 2. Batch recalculation is performed every Sunday morning at the preset time: Read the cumulative data of the photographer in the warm data layer within the most recent preset time period, and recalculate the trend value of each capability dimension according to the time decay weight.

[0183] Among them, time decay weighting refers to a weighting method that assigns higher weight to recent data and lower weight to distant data. Its purpose is to make the photographer's behavioral profile more accurately reflect the photographer's current ability status rather than being flattened by long-term historical data.

[0184] Next, the recalculation results are written back to the time series array of capability indicators in the warm data layer, thus forming a complementary maintenance mechanism of "incremental updates to ensure timeliness and batch recalculation to ensure accuracy".

[0185] This invention supports millisecond-level real-time queries through a memory database in the hot data layer, meeting the high-frequency access needs of monitoring terminals during activities; it supports complex structured queries and sorting through a relational database in the warm data layer, ensuring the needs of profile analysis and team management; and it uses object storage in the cold data layer to archive raw logs long-term at low cost, balancing cost control and data recoverability. The dual-track collaboration of incremental updates and batch recalculation ensures both the real-time nature of profile data and the accuracy of long-term trends, solving the inherent challenge of traditional single-layer storage solutions in balancing real-time storage, query capabilities, cost control, and data integrity.

[0186] Secondly, embodiments of the present invention also provide a real-time shooting quality early warning system 20 based on shooting behavior characteristics, comprising: The shooting behavior acquisition module 210 is used to collect multi-source shooting behavior data and identify activity categories, and perform time alignment and data cleaning based on the preview frame timestamp to obtain standard shooting behavior data; the multi-source shooting behavior data is a multi-dimensional time-series dataset containing terminal status, framing environment and operation characteristics. The edge-side risk identification module 220 is used to input standard shooting behavior data into a multi-dimensional quality risk model to obtain normalized risk scores for each quality dimension; the multi-dimensional quality risk model is a set of models for normalized risk assessment of the image presentation dimension and shooting execution dimension of the shooting scene. The continuous frame warning trigger module 230 is used to acquire an activity template corresponding to the activity category and storing weight allocation parameters for each quality dimension; to perform a weighted summation of the normalized risk scores to obtain a comprehensive risk score; to map the comprehensive risk score to a corresponding warning segment based on a preset warning threshold; the warning segment includes a non-warning segment, a first-level warning segment, a second-level warning segment, and a third-level warning segment, with the risk severity corresponding to each level of warning segment increasing sequentially; when the number of consecutive frames within a preset time window reaches a preset frame number threshold, and the comprehensive risk score of each frame is mapped to the same warning segment, a warning message is triggered. The behavior profile update module 240 is used to calculate the difference in standard shooting behavior data across each shooting behavior dimension within a preset number of frames before and after the warning trigger time as the behavior change amount; if the behavior change amount exceeds a preset change threshold and automatic device adjustment is excluded, it is identified as an active correction action, and the quality metric of the active correction action is calculated; the warning information, active correction action, and quality metric are associated and stored as a photographer behavior profile; after each event, an incremental update is performed based on the photographer behavior profile newly added for this event, and a batch recalculation is performed on the stored data of the photographer behavior profile when the preset period arrives.

[0187] Thirdly, embodiments of the present invention also provide an electronic device, comprising: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method described in any of the above embodiments.

[0188] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention without departing from the principles and spirit of the present invention.

Claims

1. A real-time shooting quality early warning method based on shooting behavior characteristics, characterized in that, include: Collect multi-source shooting behavior data and identify activity categories. Using the preview frame timestamp as the time reference, perform time alignment and data cleaning to obtain standard shooting behavior data. The multi-source shooting behavior data is a multi-dimensional time-series dataset that includes terminal status, framing environment and operation characteristics. Standard shooting behavior data is input into a multidimensional quality risk model to obtain normalized risk scores for each quality dimension; The multidimensional quality risk model is a set of models that normalize the risk assessment of the image presentation dimension and the shooting execution dimension of the captured image. Obtain the activity template corresponding to the activity category, which stores the weight allocation parameters for each quality dimension, and perform a weighted summation on the normalized risk scores to obtain the comprehensive risk score; Based on a preset warning threshold, the comprehensive risk score is mapped to a corresponding warning zone. The warning zones include a non-warning zone, a first-level warning zone, a second-level warning zone, and a third-level warning zone, and the severity of the risk corresponding to each level of warning zone increases sequentially. When the number of consecutive frames within a preset time window reaches a preset frame threshold, and the comprehensive risk score of each frame is mapped to the same warning segment, a warning message is triggered. The difference between standard shooting behavior data in each shooting behavior dimension within a preset number of frames before and after the warning trigger time is calculated as the behavior change amount; if the behavior change amount exceeds the preset change threshold and automatic device adjustment is excluded, it is identified as an active correction action, and the quality metric index of the active correction action is calculated. The warning information, proactive corrective actions, and quality metrics are linked and stored to create a photographer's behavioral profile. After each event, an incremental update is performed based on the newly added photographer behavior profiles for that event, and a batch recalculation is performed on the stored data of the photographer behavior profiles when a preset period arrives.

2. The method according to claim 1, characterized in that, The process of collecting multi-source shooting behavior data and identifying activity categories includes: The multi-source shooting behavior data includes preview frames, focus frame positions, exposure parameters, terminal posture data, lens orientation, and camera position. The system reads preview frames in real time from the viewfinder preview channel at a first sampling frequency, and obtains the focus frame position according to the autofocus system callback trigger method. The sampling period of the focus frame position is less than a preset time threshold. The system reads exposure parameters consisting of shutter speed, aperture, ISO, and exposure compensation from the camera application interface at a second sampling frequency. The system reads terminal attitude data consisting of pitch angle, roll angle, and yaw angle from the inertial measurement unit at a third sampling frequency, and reads the change state of lens direction in an event-triggered manner. The system reads the camera position at a fourth sampling frequency. The third sampling frequency is higher than the first sampling frequency, the first sampling frequency is higher than the second sampling frequency and the fourth sampling frequency; and when the terminal computing power is lower than a preset computing power threshold, the first sampling frequency is adaptively reduced to a preset minimum frame rate. Based on the scene classifier built into the acquisition terminal, the current framing environment is classified in real time and labeled with the corresponding activity category. For each quality dimension under the current activity category, record the number of samples collected by the sensor; when the number of samples is lower than a preset sample size threshold, trigger a scene supplement prompt; calculate the coefficient of variation for the sensor data corresponding to a preset number of consecutive frames; when the coefficient of variation exceeds a preset variation threshold, determine that the current acquisition environment is unstable, and reduce the risk judgment weight within the time window corresponding to the consecutive number of frames by a preset amount.

3. The method according to claim 1, characterized in that, The process of using the preview frame timestamp as a time base to perform time alignment and data cleaning includes: Using the timestamp sequence of the preview frames as the reference time axis, the focus frame position, exposure parameters, terminal posture data, lens direction and camera position are aligned to the reference timestamp using the nearest neighbor interpolation method, and the alignment tolerance does not exceed the preset time deviation range. A sliding window mean filter is applied to the inertial measurement unit angular velocity sequence in the terminal attitude data. The window is gradually slid over the angular velocity sequence with a preset window length, and the arithmetic mean of all sampled values ​​in each window is used as the filtered output at the current window time. Kalman filtering is applied to the machine position, and the plane coordinates and moving speed components of the machine position are used as state variables. A state transition matrix is ​​constructed based on the uniform motion model. Prediction is performed at each sampling time to obtain a prior estimate. The prior estimate is then weighted and corrected with new sampled values ​​to obtain a posterior estimate. If the sensitivity value of one frame exceeds the sensitivity range allowed by the terminal hardware specifications, or the absolute value of the angular velocity output by the inertial measurement unit exceeds the preset angular velocity threshold, then the frame is marked as a sensor abnormality and discarded from the standard shooting behavior data. If, within a preset time window, the duration of missing valid sampling data for any sensor dimension exceeds a preset missing duration threshold, the comprehensive risk score corresponding to that time window will be determined as uncertain.

4. The method according to claim 1, characterized in that, The process of obtaining the activity template corresponding to the activity category and storing the weight allocation parameters for each quality dimension includes: Different activity templates are pre-configured for different activity categories and stored in the template repository. Each activity template contains a set of weight allocation parameters and a set of stage configurations. The weight allocation parameter set consists of weight coefficients corresponding to each quality dimension. Each weight coefficient is greater than zero and the sum of all weight coefficients is equal to 1. The weight coefficients are configured differently according to the emphasis of the activity category on the delivery quality. The configuration set of the stages consists of the names of the stages included in the activity, the minimum number of qualified coverage shots for each stage, and the start and end time information of each stage, which is used to provide a benchmark for the coverage risk model judgment of key stages in the shooting execution dimension. When creating an activity, the corresponding activity template is read from the repository and loaded into the runtime configuration area; so that when the multidimensional quality risk model calculates the comprehensive risk score, it directly calls the weight coefficients in the runtime configuration area to perform a weighted summation of the normalized risk scores of each quality dimension.

5. The method according to claim 1, characterized in that, When the number of consecutive frames within a preset time window reaches a preset frame threshold, and the comprehensive risk score of each frame is mapped to the same warning segment, a warning message is triggered, including: Obtain the time window length and frame threshold corresponding to the current activity category; The time window length and the frame count threshold are determined based on the category of the current activity; the time window length is negatively correlated with the scene dynamics corresponding to the activity category, and the frame count threshold is positively correlated with the scene dynamics. Within the time window, the number of frames whose comprehensive risk scores are continuously mapped to the same warning segment is read and counted frame by frame; when the number of consecutive frames reaches the frame number threshold, the warning information for that warning segment is triggered. If the warning segment mapped by the comprehensive risk score of a frame is different from that of the previous frame during the statistical process, the current accumulated number of consecutive frames will be cleared to zero and the number of consecutive frames will be re-accumulated from that frame. When the normalized risk score of any quality dimension exceeds the preset mandatory upgrade threshold, the warning segment corresponding to the current warning information will be upgraded by one level in the direction of increasing risk severity, and the upgraded warning segment will not exceed the level three warning segment.

6. The method according to claim 1, characterized in that, The multidimensional quality risk model includes a composition risk model, an exposure risk model, a focus risk model, a person's state risk model, and an occlusion risk model, all constructed based on the dimensions of image presentation. The composition risk model takes the preview frame, the coordinates of the subject detection box, and the terminal posture angle as inputs. It determines the risk of subject missing if the proportion of the subject detection box area to the total screen area is lower than a preset proportion threshold; it determines the risk of subject offset if the normalized Euclidean distance between the center point of the subject detection box and the preset composition reference point of the screen is higher than a preset offset threshold; and it determines the risk of horizontal tilt if the absolute value of the horizontal tilt angle measured by the roll angle is higher than a preset tilt threshold. If the coordinates of a key point of a subject exceed the boundary of the image and the confidence level of that key point is higher than a preset confidence threshold, it is considered a risk of the subject going out of bounds. The maximum value of the normalized risk of each of the above sub-items is taken as the construction risk score; The exposure risk model takes the preview frame brightness histogram, average brightness of the face area, average brightness of the background area, and exposure compensation as inputs. It determines the risk of highlight clipping when the proportion of pixels with brightness higher than the highlight threshold is higher than the preset overexposure threshold; it determines the risk of black shadow when the proportion of pixels with brightness lower than the shadow threshold is higher than the preset underexposure threshold; and it determines the risk of backlight face when the difference between the average brightness of the background and the average brightness of the face is higher than the preset backlight threshold and the exposure compensation is negative. The maximum value of each sub-item risk after normalization is taken as the exposure risk score; The focus risk model takes the preview frame, focus frame coordinates, and autofocus status indicator as input. It determines low contrast risk if the average edge gradient magnitude in the focus frame area is lower than a preset gradient threshold; determines focus deviation risk if the intersection-union ratio (IU) between the focus frame and the subject detection frame is lower than a preset IU threshold; and determines focus blur risk if the image sharpness metric in the focus frame area is lower than a preset sharpness threshold. The model calculates a comprehensive probability based on the XGBoost binary classifier as the focus risk score. The character state risk model takes the face region, facial key points, and eye aspect ratio time series as input. It judges the risk of closed eyes as the eye aspect ratio being lower than the preset closed eye threshold and the number of consecutive frames reaching the preset closed eye frame number. It judges the risk of abnormal expression as the sum of the probabilities of preset negative expression categories being higher than the preset expression threshold. If the standard deviation of the head yaw angle of multiple figures is higher than the preset line of sight threshold, it is judged as a risk of inconsistency between the line of sight and the line of sight. The maximum value of each sub-item risk after normalization is taken as the character's status risk score; The occlusion risk model takes the preview frame, face detection box and semantic segmentation result as input. It determines face occlusion risk if the intersection-union ratio of the segmented region of the occlusion category and the face detection box is higher than the preset occlusion threshold, and determines lens smudge risk if the low variance region that appears continuously at a fixed position in multiple consecutive frames appears continuously. The maximum value of each sub-item risk after normalization is taken as the occlusion risk score.

7. The method according to claim 1, characterized in that, The multidimensional quality risk model also includes a key link coverage risk model and an equipment anomaly risk model built based on the shooting execution dimension; The critical link coverage risk model takes the current link identifier, the number of qualified photos taken in the current link, the minimum number of photos covered in the current link, and the duration of the current link as inputs. The ratio of the number of qualified photos taken to the minimum number of photos covered is used as the coverage completion rate. When the duration of the link exceeds the preset high-risk ratio of the preset duration of the link and the coverage completion rate is lower than the preset high-risk coverage threshold, a high coverage risk is output. When the coverage completion rate is lower than the preset medium-risk coverage threshold and the remaining duration of the link is lower than the preset remaining duration threshold, a medium coverage risk is output and normalized into a critical link coverage risk score. The device anomaly risk model uses terminal battery power, remaining storage space, network upload latency, and camera module temperature as input indicators. When any indicator exceeds the corresponding safe range, it is judged as a device anomaly risk.

8. The method according to claim 1, characterized in that, The calculation of the quality metrics for the active corrective action includes corrective response time, corrective effectiveness, quality improvement, and corrective thoroughness, among others: The difference between the time when the active correction action is detected and the time when the corresponding warning information is triggered is used as the correction response time to characterize the photographer's response speed to the warning. The cumulative number of times that a warning for the same quality dimension is not triggered again within the preset validity period after an active corrective action is taken, and the ratio of this number to the total number of active corrective actions is used as the corrective effectiveness rate, which is used to characterize the degree to which corrective actions suppress similar risks. The quality improvement degree is obtained by subtracting the comprehensive risk score after the correction action is completed from the comprehensive risk score before the warning is triggered at the same camera position or in the same scenario. A positive difference indicates that the risk has decreased after the correction. The number of times the same risk label is triggered in a single event is counted. The reciprocal of the number of triggers is normalized and used as the degree of correction thoroughness. The fewer the number of triggers, the higher the degree of correction thoroughness.

9. The method according to claim 1, characterized in that, Also includes: The photographer's behavioral profile is stored in a three-tiered storage system consisting of a hot data layer, a warm data layer, and a cold data layer. The hot data layer stores real-time profile snapshots of the current activity in an in-memory database. The real-time profile snapshots include the early warning trigger count and real-time correction effectiveness of the activity, and are retained for a preset time after the activity ends before becoming invalid. The warm data layer uses a relational database to store the full-scale portrait structured data in separate tables according to photographer identification. The full-scale portrait structured data includes a time series array of historical scores for each capability dimension, a list of error patterns sorted by trigger frequency, and an activity history. It also supports querying and sorting by capability dimension and activity type. The cold data layer stores historical original shooting behavior logs as objects. The historical original shooting behavior is stored in partitions according to the shooting time and photographer's identifier, and is automatically archived to low-cost storage media after a preset retention period. Specifically, the incremental update writes the event data to the hot data layer and the warm data layer after the event ends, and the batch recalculation reads the cumulative data of the warm data layer and the cold data layer at a preset period and writes the recalculation result back to the warm data layer.

10. A real-time shooting quality early warning system based on shooting behavior characteristics, characterized in that, include: The shooting behavior acquisition module is used to collect multi-source shooting behavior data and identify activity categories. It uses the preview frame timestamp as the time base to perform time alignment and data cleaning to obtain standard shooting behavior data. The multi-source shooting behavior data is a multi-dimensional time-series dataset that includes terminal status, framing environment and operation characteristics. The edge-side risk identification module is used to input standard shooting behavior data into a multi-dimensional quality risk model to obtain normalized risk scores for each quality dimension. The multidimensional quality risk model is a set of models that normalize the risk assessment of the image presentation dimension and the shooting execution dimension of the captured image. The continuous frame warning triggering module is used to obtain an activity template corresponding to the activity category and storing the weight allocation parameters of each quality dimension, and to perform a weighted summation of the normalized risk scores to obtain a comprehensive risk score. Based on a preset warning threshold, the comprehensive risk score is mapped to the corresponding warning segment. The warning segment includes a non-warning segment, a first-level warning segment, a second-level warning segment, and a third-level warning segment, and the severity of the risk corresponding to each level of warning segment increases sequentially. When the number of consecutive frames within a preset time window reaches a preset frame number threshold, and the comprehensive risk score of each frame is mapped to the same warning segment, a warning message is triggered. The behavior profile update module is used to calculate the difference in standard shooting behavior data in each shooting behavior dimension within a preset number of frames before and after the warning trigger time as the behavior change amount; if the behavior change amount exceeds the preset change threshold and automatic device adjustment is excluded, it is identified as an active correction action, and the quality metric of the active correction action is calculated; the warning information, active correction action and quality metric are associated and stored as a photographer behavior profile. After each event, an incremental update is performed based on the newly added photographer behavior profiles for that event, and a batch recalculation is performed on the stored data of the photographer behavior profiles when a preset period arrives.

Images