Multi-screen fusion intelligent delivery method and system

Abstract

This invention discloses a multi-screen integrated intelligent advertising method and system, comprising: acquiring audience perception data to construct profiles of interested objects and lingering groups; generating dynamic playback strategies based on profile mapping and comparison; predicting target arrival time based on a spatiotemporal kinematics model and performing semantic parsing on the target advertisement to extract key semantic frames, adjusting the non-key frame scaling to match the rendering timestamp with the arrival time, and generating instructions; controlling the display device group to play according to the instructions, and calculating the risk avoidance guidance sub-screen based on the gaze direction of the risky object to generate and execute warning instructions to smoothly transition it to a risk avoidance guidance image, which is maintained on non-adjacent sub-screens; effectively alleviating the problem of low advertising accuracy, realizing a smooth transition from attracting viewers to guiding risk avoidance, and significantly improving the dynamic adaptation capability of multi-screen display and the safety of the surrounding environment.

Description

Technical Field

[0001] This invention relates to the field of intelligent advertising control, specifically to a multi-screen integrated intelligent advertising delivery method and system. Background Technology

[0002] Existing advertising systems often rely on fixed content rotation or simple visitor flow-triggered playback, lacking in-depth analysis of audience characteristics. This prevents precise, personalized content matching based on the multi-dimensional features of different audiences, resulting in low effective reach. Furthermore, conventional multi-screen advertising devices are functionally limited, lacking flexible multi-screen splicing and split-screen collaborative control mechanisms, making it difficult to meet the needs of delivering various types of advertisements to different groups at different times in complex scenarios. Additionally, existing outdoor advertising screens rarely consider the dynamic safety conditions of the surrounding traffic environment during operation, failing to provide timely warnings and avoidance reminders for potential conflicts between pedestrians and vehicles. Moreover, warnings often involve abrupt global screen switching, failing to consider which screen is currently attracting the attention of pedestrians in danger, lacking a smooth transition mechanism from commercial advertising to safety guidance, posing significant public safety hazards and creating a jarring visual experience. Summary of the Invention

[0003] This invention proposes a multi-screen integrated intelligent advertising method, which involves acquiring audience perception data and constructing profiles of interested objects and lingering groups based on the audience perception data; performing mapping and comparison between the profiles of interested objects and lingering groups, and generating a dynamic playback strategy based on the mapping and comparison results; acquiring the movement trajectory of target pedestrians and predicting the target arrival time of the target pedestrians entering the advertising area based on a spatiotemporal kinematics model; performing video semantic analysis on the target advertisements to extract key semantic frames from the target advertisements; calculating and adjusting the playback multiplier of non-key semantic frames so that the rendering timestamp of the key semantic frames matches the target arrival time, and generating a spatiotemporally synchronized time-sharing advertising instruction.

[0004] Furthermore, according to the spatiotemporal synchronization time-sharing delivery instruction, the control display device group, during the spatiotemporal synchronization playback process, detects the movement status of surrounding people and vehicles through visual analysis cameras. When it is determined that there is a risk of conflict between people and vehicles, based on the physical coordinates of the conflict and the real-time gaze direction of the risky object, it calculates the evacuation guidance sub-screen that attracts the strongest attention of the target evacuation personnel in danger, generates and executes an evacuation warning instruction, controls the evacuation guidance sub-screen to smoothly transition the current advertising content to evacuation guidance images, and controls the other non-adjacent sub-screens to maintain normal advertising playback.

[0005] Furthermore, when the risk of a human-vehicle conflict is determined to exist, the visual analysis camera captures the facial orientation and pupil gaze vector of the person in danger. In the spatial topology mapping table of the display device group, the backward extension of the pupil gaze vector is spatially intersected with the physical display surface of each sub-screen. If there are intersecting sub-screens, the sub-screen with the largest intersection area is selected as the avoidance guidance sub-screen. If there are no intersecting sub-screens, the sub-screen closest to the physical coordinates where the conflict occurs is selected as the avoidance guidance sub-screen.

[0006] Further, the video stream of the target advertisement currently playing on the hazard avoidance guidance sub-screen is extracted, and the main color tone and spatial brightness distribution characteristics of the current rendered frame are obtained; a dynamic guidance layer with hazard avoidance warning semantics is generated, and the initial transparency of the dynamic guidance layer is set to zero; within a preset transition time window, the transparency of the dynamic guidance layer is gradually increased, and the main color tone of the current rendered frame is interpolated and gradually changed to a high-contrast warning color tone, and the spatial brightness distribution characteristics of the image are concentrated towards the area indicated by the guidance arrow, until the hazard avoidance guidance image is fully presented.

[0007] Furthermore, the displacement vector of the target pedestrian is obtained through continuous frame image analysis, and the walking speed of the target pedestrian is combined with the spatiotemporal kinematic model to calculate the target arrival time required for the target pedestrian to travel from the current position to the viewing area directly in front of the advertising screen.

[0008] Further, the video stream of the target advertisement is input into a pre-trained semantic segmentation model to identify and label image frames containing product close-ups and brand climaxes as the key semantic frames; the first timestamp of the key semantic frames in the original video stream is extracted, and the target arrival time of the target pedestrian is obtained as the second timestamp; the time difference between the first timestamp and the second timestamp is calculated using a dynamic time warping algorithm, and the target compression ratio of non-key semantic frames is determined based on the time difference; the non-key semantic frames are subjected to frame extraction or acceleration processing according to the target compression ratio so that the actual rendering time of the key semantic frames is aligned with the target arrival time.

[0009] Furthermore, in the step of calculating and adjusting the playback ratio of non-critical semantic frames, the target compression ratio is determined by the formula... Calculated; where, This represents the target compression ratio. This represents the time difference between the first timestamp and the second timestamp. This represents the original total playback duration of the non-critical semantic frames.

[0010] Furthermore, in the step of generating the spatiotemporal synchronization time-sharing delivery instruction, the rendering discard threshold of the non-critical semantic frame is extracted according to the target compression ratio. The rendering discard threshold is obtained by formula... The calculation yields the result; where `round()` represents the rounding function. This represents the rendering discard threshold. The total original frame rate of the non-critical semantic frames. The target compression ratio is represented; after performing equal-interval frame extraction on the non-critical semantic frames according to the rendering discard threshold, the spatiotemporal synchronization time-sharing delivery instruction is generated.

[0011] Furthermore, this application also proposes a non-volatile storage medium storing a computer program, which, when executed by a processor, implements the aforementioned multi-screen fusion intelligent projection method.

[0012] Furthermore, this application also proposes a multi-screen integrated intelligent advertising system, including: a profile construction module, used to perform the processing of obtaining profiles of interested objects and lingering groups from audience perception data; a dynamic strategy generation module, used to perform the processing of profile mapping and comparison to generate dynamic playback strategies; and a spatiotemporal synchronous advertising module, used to perform the processing of spatiotemporal kinematic model prediction, video semantic parsing, non-critical semantic frame playback rate adjustment to generate spatiotemporal synchronous time-sharing advertising instructions.

[0013] This invention proposes a multi-screen integrated intelligent advertising method. By constructing profiles from multi-dimensional sensory data and mapping them to filter target advertisements, it effectively alleviates the problem of low advertising accuracy and significantly improves audience reach. By extracting key semantic frames and adjusting the scaling of non-key frames to achieve spatiotemporal synchronous advertising, it effectively improves the alignment between key nodes of advertisement playback and audience arrival time. By accurately calculating the gaze direction of dangerous personnel to determine the hazard guidance sub-screen and achieving a smooth visual transition from commercial advertisements to hazard guidance images, it effectively alleviates public safety hazards around outdoor advertising screens, avoids visual abruptness and hazard delay caused by abrupt screen transitions, and significantly improves the safety and visual friendliness of the system operation. Attached Figure Description

[0014] Figure 1 A flowchart illustrating a multi-screen integrated intelligent advertising method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the timing alignment of the spatiotemporal synchronization time-division delivery command generation in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the calculation principle of the sub-screen spatial mapping for risk avoidance guidance based on gaze vectors in an embodiment of the present invention. Figure 4This is a schematic diagram illustrating the smooth transition of the avoidance guidance sub-screen in an embodiment of the present invention. Detailed Implementation

[0015] This invention proposes a multi-screen integrated intelligent advertising method, comprising: S1. Obtain audience perception data, and construct a profile of interested objects and a profile of staying groups based on the audience perception data.

[0016] The audience perception data refers to a composite data set including pedestrian gaze focus time, gender, age, crowd cluster information, voice feedback data, and a physical dwell trajectory heatmap. The physical dwell trajectory heatmap is a two-dimensional spatial mapping generated by continuously tracking the distribution density of pedestrian coordinate positions in front of the display device group over time. The interest object profile refers to a set of feature vectors representing a high degree of interest in a specific advertisement, constructed based on historical audience data. The lingering group profile refers to a set of feature vectors representing the characteristics of the current audience group, constructed based on real-time lingering crowd data in front of the screen.

[0017] Specifically, the system deploys image and voice acquisition sensors below the projection screen to simultaneously acquire video and audio streams. It uses image recognition algorithms to extract pedestrian facial and posture features, and uses a voice analysis model to extract voiceprint and semantic features. At the same time, it uses a multi-target tracking algorithm to record the coordinate trajectory of pedestrians in front of the screen and generate a heat map. After fusing the above features, it performs vectorized encoding to form a basic data structure for the portrait that includes spatial location weights.

[0018] In one specific implementation scenario, a binocular camera and microphone array are deployed below a multi-screen splicing advertising wall at the entrance of a large shopping mall. When a pedestrian enters the sensing area, the camera captures the pedestrian's facial contours and estimates their age and gender, while the microphone array collects audio clips of the pedestrian's conversation and extracts acoustic features. Simultaneously, the system records the duration of the pedestrian's stay in front of each sub-screen of the advertising wall, generating a heatmap. These features are then fused in real time to provide data support for subsequent profile matching. The system stores and vectorizes the extracted multi-dimensional features according to the structured format shown in Table 1. Table 1 is a multi-dimensional feature vectorization mapping table for audience perception data.

[0019] S2. Based on the mapping comparison between the profile of the object of interest and the profile of the lingering group, a dynamic playback strategy is generated according to the mapping comparison result.

[0020] The mapping comparison refers to calculating the feature similarity between the profile of the lingering group and the profile of the object of interest corresponding to each advertisement, with spatial location weight serving as an adjustment factor in the similarity calculation. The dynamic playback strategy refers to a set of screen control instructions generated based on the mapping results and environmental time information, covering playback content, splicing layout, and split-screen window allocation schemes.

[0021] Specifically, the system uses a cosine similarity algorithm to calculate the overlap between the feature vectors of two sets of images, combines spatial location weights for weighted summation, and selects the advertisement with the highest overlap as the target advertisement. Simultaneously, the system obtains ambient light levels through a light sensor and acquires calendar and weather data through a network interface, comprehensively adjusting screen brightness and switching holiday themes.

[0022] In one specific implementation scenario, when the system identifies that the current lingering group is mainly young women who are mostly gathered in front of the left sub-screen, and the similarity of the target audience profile of a certain beauty advertisement is the highest, if it happens to be Mother's Day, the system will automatically generate a dynamic playback strategy, allocate the beauty advertisement to the left main screen for playback, and adjust the screen color temperature to adapt to the warm color atmosphere of the holiday.

[0023] S3. Obtain the movement trajectory of the target pedestrian, and predict the target arrival time of the target pedestrian entering the advertising area based on the spatiotemporal kinematics model; perform video semantic parsing on the target advertisement, and extract the key semantic frames in the target advertisement; calculate and adjust the playback multiplier of the non-key semantic frames so that the rendering timestamp of the key semantic frames matches the target arrival time, and generate a spatiotemporally synchronized time-sharing delivery instruction.

[0024] The spatiotemporal kinematic model refers to the kinematic equations established based on the pedestrian's current speed, direction, and distance, used to calculate arrival time. The target arrival time refers to the estimated time required for the pedestrian to travel from their current location to the viewing area directly in front of the advertising screen. Video semantic parsing refers to the process of classifying and labeling video frames using a deep learning model. Key semantic frames are high-information video frames containing product close-ups, brand logos, or core selling points. Non-key semantic frames are low-information video frames such as transitional scenes or background introductions. The rendering timestamp refers to the actual time the video frame is displayed on the screen. The spatiotemporal synchronized time-sharing delivery instruction refers to a control instruction containing the target compression ratio and timestamp alignment parameters.

[0025] Specifically, the system obtains pedestrian displacement vectors through continuous frame image analysis, combines their walking speed with kinematic equations to estimate arrival time; simultaneously, it inputs the video stream into a semantic segmentation network to extract keyframes, calculates the difference between the original timestamp and arrival time of each keyframe, and dynamically extracts or speeds up playback of non-keyframes based on this difference to ensure that keyframes are rendered precisely when the pedestrian stops. Figure 2As shown, the system dynamically normalizes and aligns the original video timeline with the physical pedestrian arrival timeline, and performs compression or stretching operations on non-critical semantic frames within the time difference window.

[0026] In one specific implementation scenario, when the system predicts that a target VIP customer will arrive in front of the screen in 15 seconds, and the key frame of the product display in the current target advertisement is originally scheduled to be rendered at the 20th second, the system extracts the non-key transition frames in the first 19 seconds and increases their playback multiplier, so that the key frame of the product display is presented on time at the 15th second.

[0027] Furthermore, it also includes: S4. According to the time-sharing delivery instruction, the display device group is controlled to detect the movement status of people and vehicles in the surrounding area through the visual analysis camera during the time-sharing playback process. When it is determined that there is a risk of conflict between people and vehicles, the evacuation guidance sub-screen that attracts the strongest attention of the target evacuation personnel is calculated based on the physical coordinate position of the conflict and the real-time gaze direction of the risk object. An evacuation warning instruction is generated and executed, and the evacuation guidance sub-screen is controlled to smoothly transition the current advertising content to the evacuation guidance image. At the same time, the other non-adjacent sub-screens are controlled to maintain normal advertising playback.

[0028] The display device group consists of multiple vertically oriented display screens arranged horizontally in sequence. The movement status includes the movement trajectory, speed, and relative position of personnel and vehicles. The pedestrian-vehicle conflict risk refers to the physical state where the relative distance between personnel and vehicles is less than the safe avoidance threshold and the relative speed direction constitutes a collision tendency. The target avoidance personnel refers to pedestrians on the path where the pedestrian-vehicle conflict risk occurs. The avoidance guidance sub-screen refers to the display unit in the multi-screen splicing matrix where the current visual focus of the target avoidance personnel in danger is located or the closest physical distance. The avoidance warning command refers to a composite control signal that includes screen smooth transition parameters, gaze focus resolution results, and state maintenance instructions. The avoidance guidance image refers to an avoidance instruction screen that smoothly transitions from a commercial advertisement and includes high-contrast warning colors and dynamic guide arrows.

[0029] Specifically, the system sends the generated playback command to the controllers of each display screen, synchronously refreshing the displayed content. Simultaneously, the system utilizes an edge computing module to perform real-time frame difference analysis on surrounding traffic monitoring video. Once a sudden decrease in the distance between a pedestrian and a vehicle is detected, the system immediately captures the pupil gaze direction of the at-risk pedestrian and calculates the intersection surface in the screen topology mapping table to determine the sub-screen they are currently viewing. Subsequently, the system calls the image rendering engine to perform gradual interpolation on the hue and brightness of the advertisement currently playing on that sub-screen, overlaying gradually appearing avoidance guidance arrows to achieve a seamless visual transition from commercial attraction to safety guidance, while maintaining normal playback on other sub-screens.

[0030] In a specific implementation scenario, the spatiotemporal synchronized playback mechanism and the gaze focus smooth transition early warning mechanism produce a synergistic effect in system resource scheduling and physical spatiotemporal coordinate control. In terms of system resource scheduling, video semantic parsing and non-keyframe scaling require GPU resources for real-time video stream decoding and frame extraction rendering, while pupil gaze vector calculation and smooth transition rendering also require high-performance facial feature tracking and image interpolation. The system, through a unified multi-task scheduling kernel, reuses computational power for feature vector extraction from pedestrian trajectory prediction and facial orientation and pupil gaze vector extraction from risky objects. Same-source video frames only need to undergo decoding and head pose feature extraction once to simultaneously serve advertising keyframe alignment calculation and risk avoidance guidance sub-screen judgment, significantly reducing the overall computational load of the system. This allows high-concurrency spatiotemporal synchronized playback and real-time security monitoring to run smoothly on the same edge computing node. In the physical spatiotemporal coordinate control dimension, the generated spatiotemporally synchronized time-sharing delivery instructions are precisely bound to the rendering timestamp of the target advertisement's key semantic frames and the physical arrival time of the target pedestrians, achieving a micro-scale spatiotemporal alignment between the screen display content and the pedestrians' viewing points in the physical space. Meanwhile, by calculating the evacuation guidance sub-screen based on the real-time gaze direction of the at-risk object and executing a smooth transition warning, a macro-scale spatial mapping is formed between the physical sub-screen units of the screen display area and the visual cognitive focus of the dangerous pedestrians in the physical space. When an evacuation warning instruction is triggered, the system utilizes the established video stream frame-level control channel to gradually overwrite the rendering parameters only for the evacuation guidance sub-screen, while simultaneously using a timestamp alignment mechanism to ensure that the rendering sequence of the advertisement's key semantic frames on other non-adjacent sub-screens is not disturbed. This deep coupling of micro-timeline alignment and macro-visual focus mapping enables the video frame scheduling in digital space and the safety warning response in physical space to achieve time-division multiplexing and spatial segmentation control on a multi-screen matrix. This ensures both the integrity of individual audience ad reach and instantaneous, directional, and smooth intervention for public safety in physical space, forming a highly collaborative system that combines commercial display efficiency with public safety defense capabilities in complex urban scenarios. Combination Figure 1 As shown, this constitutes a complete closed-loop process from audience perception, spatiotemporal playback to risk intervention.

[0031] In one specific implementation scenario, a multi-screen advertising wall in a square is playing synchronous advertisements to pedestrians who are about to stop. At this moment, a delivery truck suddenly drives into a densely populated pedestrian area. The system detects that a pedestrian in danger is looking at the advertisement on the upper right sub-screen. The system immediately controls the upper right sub-screen to smoothly transition the current beauty advertisement to a red warning screen with dynamic avoidance arrows within two seconds. Meanwhile, the rest of the advertising wall continues to play the original synchronous advertisements, ensuring that the target pedestrian receives avoidance guidance without losing focus.

[0032] Furthermore, when the risk of a human-vehicle conflict is determined to exist, the visual analysis camera captures the facial orientation and pupil gaze vector of the person in danger. In the spatial topology mapping table of the display device group, the backward extension of the pupil gaze vector is spatially intersected with the physical display surface of each sub-screen. If there are intersecting sub-screens, the sub-screen with the largest intersection area is selected as the avoidance guidance sub-screen. If there are no intersecting sub-screens, the sub-screen closest to the physical coordinates where the conflict occurs is selected as the avoidance guidance sub-screen.

[0033] The pupil gaze vector refers to a spatial ray that represents the direction of human eye gaze, fitted from 3D facial key points. The spatial topology mapping table is a lookup table recording the parameters of six faces of each sub-screen's physical installation location in the 3D world coordinate system. The spatial intersection calculation refers to the geometric operation of solving for the intersection point of the spatial ray and the 3D plane. The intersection area refers to the area of ​​the region where the projection circle of the gaze ray on the sub-screen plane coincides with the physical boundary of the sub-screen.

[0034] Specifically, the system extracts 2D facial key points of pedestrians at risk using a visual analysis camera, maps them to a 3D world coordinate system using a pre-calibrated camera intrinsic and extrinsic parameter matrix, and constructs a gaze vector ray based on the relative position of the pupils. Then, it traverses the planar equations of each sub-screen in the spatial topology mapping table, uses a ray-plane intersection algorithm to solve for the intersection points, determines whether the intersection points fall within the physical boundaries of the sub-screen, and calculates the intersection coverage. Figure 3 As shown, the gaze vector ray is extended in reverse in the multi-screen matrix space and intersects with the physical display surface of the sub-screen, so that the system can accurately lock onto the risk avoidance guidance sub-screen.

[0035] In one specific implementation scenario, an electric bicycle enters the square, and a pedestrian looking down at a mobile phone is in front of it. The system calculates that the movement trajectories of the two will intersect in a short time. The system immediately captures the pedestrian's facial posture and calculates that the extension line of the pupil gaze vector intersects with the sub-screen at the lower left of the advertising wall. The system then identifies the lower left sub-screen as the risk avoidance guidance sub-screen.

[0036] Further, the video stream of the target advertisement currently playing on the hazard avoidance guidance sub-screen is extracted, and the main color tone and spatial brightness distribution characteristics of the current rendered frame are obtained; a dynamic guidance layer with hazard avoidance warning semantics is generated, and the initial transparency of the dynamic guidance layer is set to zero; within a preset transition time window, the transparency of the dynamic guidance layer is gradually increased, and the main color tone of the current rendered frame is interpolated and gradually changed to a high-contrast warning color tone, and the spatial brightness distribution characteristics of the image are concentrated towards the area indicated by the guidance arrow, until the hazard avoidance guidance image is fully presented.

[0037] The main color tone of the image refers to the hue and saturation of the color cluster with the highest pixel proportion in the current rendered frame. The spatial brightness distribution feature refers to the average brightness matrix of each grid region after the screen image is divided into grids. The dynamic guidance layer refers to a transparent overlay layer containing graphic elements such as avoidance arrows and warning stripes. The interpolation gradient refers to the image processing process of weighted mixing of the pixel values ​​of the original image and the pixel values ​​of the target warning image within a linear transition time window.

[0038] Specifically, after identifying the hazard guidance sub-screen, the system pauses the original spatiotemporal synchronization and time-sharing delivery instructions for that sub-screen, and extracts the hue and brightness matrix of the current frame. Subsequently, the system generates a fully transparent guidance layer and overlays it on the original image. During the next two-second transition window, the alpha channel value of the guidance layer is increased frame by frame. Simultaneously, a linear interpolation algorithm is used to approximate the RGB values ​​of the original image towards a high-contrast red-yellow warning hue, and the overall brightness peak of the image is concentrated in the direction indicated by the guidance arrow, ultimately completely replacing it with a pure hazard guidance image. Combined with... Figure 4 As shown, within the transition time window, the image gradually transitions from the original commercial advertisement to a warning layer, ultimately presenting a pure safety guidance image. The specific parameters of the transparency and hue transitioning over time are shown in Table 2. Table 2 shows the interpolation control parameters for smooth transition of avoidance guidance images.

[0039] In one specific implementation scenario, a sub-screen is playing an advertisement for a blue-packaged mineral water. When a smooth transition is triggered, the system gradually changes the blue screen to a high-contrast red and yellow warning color within two seconds. At the same time, a white dynamic avoidance arrow gradually appears from transparent and occupies the center of the screen, guiding pedestrians to move to the safe area on the left.

[0040] Furthermore, the displacement vector of the target pedestrian is obtained through continuous frame image analysis, and the walking speed of the target pedestrian is combined with the spatiotemporal kinematic model to calculate the target arrival time required for the target pedestrian to travel from the current position to the viewing area directly in front of the advertising screen.

[0041] The continuous frame image analysis refers to a method that extracts the pixel displacement of the target pedestrian in adjacent video frames and calculates the actual physical distance by combining it with camera calibration parameters. The spatiotemporal kinematic model integrates the pedestrian's instantaneous velocity vector with environmental obstacle constraints to estimate the time required for unobstructed straight-line movement to the target area.

[0042] Specifically, the system uses a multi-target tracking algorithm to lock onto the target pedestrian, extracts the difference in centroid coordinates between two consecutive frames, and converts it into the actual displacement of the ground plane by combining the intrinsic and extrinsic parameter matrices of the camera. This displacement is then divided by the frame interval to obtain the walking speed, and the arrival time is further calculated by combining the straight-line distance between the pedestrian and the front of the screen.

[0043] In one specific implementation scenario, the system identifies a pedestrian walking towards the screen at a speed of one step per second. Based on the pedestrian's physical position of five meters away from the screen, it calculates that the pedestrian will reach the optimal viewing area directly in front of the advertising screen in ten seconds.

[0044] Further, the video stream of the target advertisement is input into a pre-trained semantic segmentation model to identify and label image frames containing product close-ups and brand climaxes as the key semantic frames; the first timestamp of the key semantic frames in the original video stream is extracted, and the target arrival time of the target pedestrian is obtained as the second timestamp; the time difference between the first timestamp and the second timestamp is calculated using a dynamic time warping algorithm, and the target compression ratio of non-key semantic frames is determined based on the time difference; the non-key semantic frames are subjected to frame extraction or acceleration processing according to the target compression ratio so that the actual rendering time of the key semantic frames is aligned with the target arrival time.

[0045] The pre-trained semantic segmentation model refers to a model trained based on a deep learning architecture, capable of pixel-level classification and scene recognition of video frame content. The first timestamp refers to the absolute position of the key semantic frame on the original video timeline. The dynamic time warping algorithm is used to calculate the optimal matching path and time difference between two time series. The target compression ratio refers to the ratio coefficient calculated to shorten or lengthen the total playback duration of non-key frames to fill the time difference.

[0046] Specifically, the system decodes the video stream and inputs it frame by frame into the model. The model outputs semantic tags for each frame and filters out frames containing product identifiers, recording their timestamps. Then, it reads the pedestrian arrival time, calculates the difference between the two, divides the difference by the original duration of the non-critical frames to obtain the target compression ratio, and finally uses a video processing chip to perform frame extraction or variable-speed rendering on the non-critical frames.

[0047] In one specific implementation scenario, the system identifies that the product packaging display frame in an advertisement is at the 20th second, while the target pedestrian will arrive at the 15th second. The system calculates that the non-critical frames in the first 19 seconds need to be compressed to 14 seconds, obtains the target compression ratio, and accelerates the non-critical frames accordingly, so that the product packaging display frame is rendered on time at the 15th second.

[0048] Furthermore, the target compression ratio is determined by the formula... Calculated; where, This represents the target compression ratio. This represents the time difference between the first timestamp and the second timestamp. This represents the original total playback duration of the non-critical semantic frames.

[0049] The formula establishes a quantitative mapping relationship between physical space waiting time and digital space video playback rate by converting absolute time difference values ​​into relative playback rate change coefficients, thus ensuring the mathematical rigor of compression processing.

[0050] Specifically, after obtaining the first and second timestamps, the system directly calls the calculation unit to perform a division operation to obtain the dimensionless multiplier. This serves as the frequency conversion parameter for the subsequent video processing chip. To ensure the closed-loop playback logic under extreme conditions, the system adopts differentiated processing strategies for different positive and negative time difference relationships, as shown in Table 3: Table 3 is a comparison table of the adaptive playback control logic for the target compression ratio.

[0051] In a specific implementation scenario, when the time difference is 5 seconds and the original total playback duration is 19 seconds, the system calculates the target compression ratio. The value is approximately 0.263, meaning that non-keyframes need to be played at about 26% of the original speed or have frames skipped.

[0052] Further, the rendering discard threshold for the non-critical semantic frames is extracted based on the target compression ratio, and the rendering discard threshold is obtained using the formula... Calculated; where, This represents the rendering discard threshold. The total original frame rate of the non-critical semantic frames. The target compression ratio is represented; after performing equal-interval frame extraction on the non-critical semantic frames according to the rendering discard threshold, the spatiotemporal synchronization time-sharing delivery instruction is generated.

[0053] The rendering discard threshold defines a quantitative standard for how many frames in a continuous frame sequence should be discarded to achieve overall duration compression. The formula, by introducing a floor function and a maximum value protection function, ensures that at least one frame of rendering data is retained even at extremely high compression ratios, preventing complete loss of the image.

[0054] Specifically, after calculating the target compression ratio, the system substitutes the result into the formula to obtain the threshold. Then iterate through the non-keyframe sequence, reading each... Each frame is discarded, and the remaining frames are repackaged into a new video stream with timestamp alignment parameters appended to generate the final control commands.

[0055] In a specific implementation scenario, when the original frame rate is 300 frames and the target compression ratio is 0.5, the system calculates that the rendering drop threshold is 2, that is, 1 frame is dropped every 2 frames, and 150 frames are actually rendered, thereby accurately compressing the total duration of non-critical frames by half.

[0056] Furthermore, this application also proposes a non-volatile storage medium storing a computer program, which, when executed by a processor, implements the aforementioned multi-screen fusion intelligent projection method.

[0057] The non-volatile storage medium refers to a computer storage medium that retains stored data even after the external power supply is disconnected, including but not limited to read-only memory, erasable programmable read-only memory, flash memory, solid-state drive, or optical disk. The computer program includes computer-executable instruction code segments for executing the above S1 to S4.

[0058] Specifically, in actual edge computing advertising terminals, this non-volatile storage medium is either built into or externally attached to the system's main control board. Developers encapsulate compiled image recognition algorithms, spatiotemporal kinematic prediction models, dynamic time warping algorithms, and smooth transition rendering engines into machine-executable code and burn it into this storage medium. When the terminal device powers on, the main control processor reads and loads the computer program from the storage medium, and then continuously executes the entire process of audience perception, profile matching, spatiotemporal synchronized delivery, and risk avoidance smooth transition warnings on the multi-screen advertising matrix. Even if the device experiences an unexpected restart, the core logic program in the storage medium will not be lost, ensuring the stability and reliability of the system operation.

[0059] Furthermore, this application also proposes a multi-screen integrated intelligent advertising system, including: a profile construction module, used to perform the processing of obtaining profiles of interested objects and lingering groups from audience perception data; a dynamic strategy generation module, used to perform the processing of profile mapping and comparison to generate dynamic playback strategies; and a spatiotemporal synchronous advertising module, used to perform the processing of spatiotemporal kinematic model prediction, video semantic parsing, non-critical semantic frame playback rate adjustment to generate spatiotemporal synchronous time-sharing advertising instructions.

[0060] The system adopts a modular architecture design, with each module interacting with the others via a high-speed internal bus or microservice interface. The profile building module receives multimodal perception signals collected by an external sensor matrix and performs vectorized fusion; the dynamic strategy generation module has a built-in high-concurrency feature comparison engine that calculates profile similarity in real time and issues scheduling instructions; the spatiotemporal synchronous delivery module integrates a video stream encoding / decoding chip and an AI computing unit, specifically designed to handle high-load semantic frame parsing and physical spatiotemporal timestamp alignment calculations.

[0061] In one specific implementation scenario, the system is deployed on a large outdoor multi-screen advertising tower in a smart city transportation hub. The profile building module uses a dual-camera array at the top of the tower to capture real-time pedestrian flow characteristics in the plaza and construct a group profile; the dynamic strategy generation module quickly matches high-priority target advertisements from the cloud-based advertising library; and the spatiotemporal synchronous delivery module combines pedestrian walking speed to predict the time it takes for them to reach the optimal viewing point, and performs real-time decoding and frame-by-frame rendering of the target advertisement video to ensure that the brand's core keyframes are displayed precisely at the moment a pedestrian stops. Furthermore, the system's modular design allows for flexible addition or removal of functional modules for specific scenarios, improving the system's engineering deployment capabilities.

[0062] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A multi-screen integrated intelligent delivery method, characterized in that, include: S1. Obtain audience perception data, and construct a profile of interested objects and a profile of staying groups based on the audience perception data; The audience perception data refers to a composite data set including pedestrian eye focus time, gender, age, crowd cluster information, voice feedback data, and physical dwell trajectory heat map. The physical dwell trajectory heat map is a two-dimensional spatial mapping map generated by continuously tracking the distribution density of the coordinate position of pedestrians in front of the display device group over time. S2. Based on the profile of the object of interest and the profile of the lingering group, perform a mapping comparison, and generate a dynamic playback strategy according to the mapping comparison result; the mapping comparison refers to calculating the feature similarity between the profile of the lingering group and the profile of the object of interest corresponding to each advertisement, with spatial location weight as an adjustment factor in the similarity calculation, and selecting the advertisement with the highest overlap as the target advertisement. S3. Obtain the movement trajectory of the target pedestrian and predict the target arrival time of the target pedestrian entering the advertising area based on the spatiotemporal kinematics model; Perform video semantic analysis on the target advertisement to extract key semantic frames from the target advertisement; The playback ratio of non-critical semantic frames is calculated and adjusted to match the rendering timestamp of the critical semantic frames with the target arrival time, generating a spatiotemporally synchronized time-sharing delivery instruction. The spatiotemporal kinematic model refers to the kinematic equations established based on the pedestrian's current speed, direction, and distance. Video semantic parsing refers to the process of classifying and labeling video frames using a deep learning model. The critical semantic frames are high-information video frames containing product close-ups, brand logos, or core selling points, while non-critical semantic frames are low-information video frames such as transitional scenes or background introductions. The rendering timestamp refers to the actual moment the video frame is displayed on the screen. The spatiotemporally synchronized time-sharing delivery instruction includes the target compression ratio and timestamp pair. The control instructions for the parameters are as follows: The video semantic parsing is specifically as follows: The video stream of the target advertisement is input into a pre-trained semantic segmentation model, image frames containing product close-ups and brand climaxes are identified and labeled as the key semantic frames, the first timestamp of the key semantic frames in the original video stream is extracted, and the target arrival time of the target pedestrian is obtained as the second timestamp. The time difference between the first timestamp and the second timestamp is calculated by a dynamic time warping algorithm. Based on the time difference, the target compression ratio of non-key semantic frames is determined, and the non-key semantic frames are subjected to frame extraction or acceleration processing according to the target compression ratio so that the actual rendering time of the key semantic frames is aligned with the target arrival time.

2. The method according to claim 1, characterized in that, Also includes: S4. According to the time-sharing delivery instruction, the display device group is controlled to detect the movement status of people and vehicles in the surrounding area through the visual analysis camera during the time-sharing playback process. When it is determined that there is a risk of conflict between people and vehicles, the evacuation guidance sub-screen that attracts the strongest attention of the target evacuation personnel is calculated based on the physical coordinate position of the conflict and the real-time gaze direction of the risk object. An evacuation warning instruction is generated and executed, and the evacuation guidance sub-screen is controlled to smoothly transition the current advertising content to the evacuation guidance image. At the same time, the other non-adjacent sub-screens are controlled to maintain normal advertising playback.

3. The method according to claim 2, characterized in that, The step of calculating the evacuation guidance sub-screen that most strongly attracts the attention of personnel seeking refuge in danger, based on the physical coordinates of the conflict and the real-time gaze direction of the risky object, specifically includes: When the risk of a human-vehicle conflict is determined to exist, the visual analysis camera captures the facial orientation and pupil gaze vector of the person in danger who is trying to avoid danger. In the spatial topology mapping table of the display device group, the backward extension of the pupil gaze vector is spatially intersected with the physical display surface of each sub-screen; If intersecting subscreens exist, the subscreen with the largest intersecting area is selected as the risk avoidance guidance subscreen; if no intersecting subscreens exist, the subscreen closest to the physical coordinates where the conflict occurred is selected as the risk avoidance guidance subscreen.

4. The method according to claim 2, characterized in that, The step of controlling the hazard avoidance guidance sub-screen to smoothly transition the current advertising content to the hazard avoidance guidance image specifically includes: Extract the video stream of the target advertisement currently playing on the risk avoidance guidance sub-screen, and obtain the main color tone and spatial brightness distribution characteristics of the current rendered frame; Generate a dynamic guide layer with risk avoidance warning semantics, wherein the initial transparency of the dynamic guide layer is set to zero; Within the preset transition time window, the transparency of the dynamic guide layer is gradually increased, and the main color tone of the current rendering frame is interpolated and gradually changed to a high-contrast warning color tone. The spatial brightness distribution characteristics of the image are concentrated towards the area indicated by the guide arrow until the danger avoidance guide image is fully presented.

5. The method according to claim 1, characterized in that, The step of obtaining the movement trajectory of the target pedestrian and predicting the target arrival time of the target pedestrian entering the advertising area based on the spatiotemporal kinematics model specifically includes: obtaining the displacement vector of the target pedestrian through continuous frame image analysis, and combining the walking speed of the target pedestrian with the spatiotemporal kinematics model to calculate the target arrival time required for the target pedestrian to move from the current position to the viewing area in front of the advertising screen.

6. The method according to claim 1, characterized in that, In the step of calculating and adjusting the playback ratio of non-critical semantic frames, the target compression ratio is determined by the formula... Calculated; where, This represents the target compression ratio. This represents the time difference between the first timestamp and the second timestamp. This represents the original total playback duration of the non-critical semantic frames.

7. The method according to claim 1, characterized in that, In the step of generating the time-sharing delivery instruction for spatiotemporal synchronization, the rendering discard threshold of the non-critical semantic frames is extracted based on the target compression ratio. The rendering discard threshold is obtained by formula... Calculated; where, This represents the rounding function. This represents the rendering discard threshold. The total original frame rate of the non-critical semantic frames. The target compression ratio is represented; after performing equal-interval frame extraction on the non-critical semantic frames according to the rendering discard threshold, the spatiotemporal synchronization time-sharing delivery instruction is generated.

8. A non-volatile storage medium, characterized in that, The non-volatile storage medium stores a computer program, which, when executed by a processor, implements the multi-screen fusion intelligent projection method as described in any one of claims 1 to 7.

9. A multi-screen integrated intelligent advertising system, used to implement the multi-screen integrated intelligent advertising method as described in any one of claims 1 to 7, characterized in that, include: The profile building module is used to perform the processing of obtaining profiles of interested objects and lingering groups from audience perception data; The dynamic strategy generation module is used to perform the process of mapping and comparing the portrait to generate the dynamic playback strategy; The spatiotemporal synchronous delivery module is used to perform spatiotemporal kinematic model prediction, video semantic parsing, and processing of non-critical semantic frame playback rate adjustment to spatiotemporal synchronous time-sharing delivery instructions.

Images