Intelligent Display Control Method and System Based on AI Visual Analysis

Abstract

This invention provides an intelligent display control method and system based on AI visual analysis, belonging to the field of artificial intelligence technology. First, it acquires an initial set of visual data streams transmitted from multiple visual acquisition devices in the display area. This set includes multiple continuously acquired visual data frame units with acquisition time sequence markers. Next, it performs display scene structure deconstruction processing on the initial visual data stream set to obtain the temporal observation tendency characteristics of the observed objects and the spatial layout constraint characteristics of the display scene. Then, it invokes a visual attention analysis network to perform joint correlation analysis, generating a description of the distribution of visual attention hotspots and coldspots. Based on this description, it determines the identification information and position migration vector sequence of the exhibits to be adjusted, and then generates a set of display device control instructions and distributes them to the corresponding components. This invention achieves intelligent and automated display control, improving the efficiency and accuracy of display effect optimization.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to an intelligent display control method and system based on AI visual analysis. Background Technology

[0002] In traditional exhibition control processes, optimizing the exhibition effect mainly relies on manual observation and experience-based judgment. Staff need to regularly patrol the exhibition area, visually observing visitor behavior and focus points to evaluate the exhibit's effectiveness. However, this method has several limitations. First, the scope of manual observation is limited, making it difficult to comprehensively cover the entire exhibition area, especially in large exhibition spaces where staff cannot simultaneously monitor the behavior of all visitors and the condition of all exhibits. Second, manual judgment has a certain degree of subjectivity; differences in experience and judgment standards among different staff members lead to inconsistent evaluation results of the exhibition effect, thus affecting the accuracy and effectiveness of exhibit adjustments.

[0003] Furthermore, while some existing exhibition control systems incorporate simple sensor technologies such as infrared and pressure sensors, these sensors can only acquire limited information, such as the number and location of visitors. They cannot provide in-depth insights into visitor behavior patterns and changes in their focus within the exhibition space. Moreover, these systems lack the ability to comprehensively analyze and process multi-source data, and cannot dynamically adjust the display of exhibits based on visitor behavior and focus. They fail to achieve intelligent exhibition control and thus cannot meet the modern exhibition industry's demands for efficient, precise, and personalized displays. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide an intelligent display control method based on AI visual analysis, the method comprising: Acquire an initial set of visual data streams transmitted by multiple visual acquisition devices deployed in the display area, wherein the initial set of visual data streams contains multiple visual data frame units with acquisition time sequence marks acquired continuously. The initial visual data stream set is subjected to a display scene deconstruction process to obtain the temporal observation tendency characteristics of the observed objects in the display scene and the spatial layout constraint characteristics of the display scene. The temporal observation tendency characteristics describe the displacement trajectory of the focus point of the observed object in a continuous visual data frame unit, and the spatial layout constraint characteristics describe the relative relationship between the display position of the exhibits and the activity area of ​​the observed object in the display scene. The pre-built visual attention analysis network is invoked to perform joint correlation analysis on the temporal observation tendency features and the spatial layout constraint features, generating a description of the distribution of hot and cold areas of visual attention in the current display scene; Based on the description of the distribution of hot and cold areas of visual attention, determine the exhibit identification information of the exhibits to be adjusted and the position migration vector sequence of the exhibit attention in the display scene space. Based on the exhibit identification information and the position migration vector sequence, a set of display device control instructions containing exhibit identification codes and angle adjustment parameters is generated, and the set of display device control instructions is distributed to the corresponding exhibit display adjustment components.

[0005] Furthermore, embodiments of the present invention also provide an intelligent display control system based on AI visual analysis, comprising: A processor; a machine-readable storage medium for storing machine-executable instructions of the processor; wherein the processor is configured to execute the aforementioned intelligent display control method based on AI visual analysis by executing the machine-executable instructions.

[0006] Based on the above, by acquiring the initial visual data stream set transmitted by multiple visual acquisition devices deployed in the exhibition area, and fully utilizing the multi-view, continuously acquired visual information, the dynamic situation within the exhibition scene can be captured comprehensively and accurately. The initial visual data stream set is processed to deconstruct the exhibition scene structure, accurately obtaining the temporal observation tendency characteristics of the observed objects within the exhibition scene and the spatial layout constraint characteristics of the exhibition scene. A pre-built visual attention analysis network is then invoked to perform joint correlation analysis on the temporal observation tendency characteristics and spatial layout constraint characteristics, generating a description of the distribution of hot and cold areas of visual attention in the current exhibition scene, avoiding the subjectivity and limitations of manual judgment. Based on this description of the distribution of hot and cold areas, the exhibit identification information of the exhibits to be adjusted and the position migration vector sequence of exhibit attention in the exhibition scene space are determined, achieving accurate evaluation and dynamic adjustment of the exhibit display effect. Finally, based on the exhibit identification information and position migration vector sequence, a set of display device control instructions containing exhibit identification codes and angle adjustment parameters is generated and distributed to the corresponding exhibit display adjustment components, realizing intelligent and automated display control, greatly improving the efficiency and accuracy of display effect optimization, and providing visitors with a higher quality and more personalized exhibition experience. Attached Figure Description

[0007] Figure 1 This is a schematic diagram of the execution flow of the intelligent display control method based on AI visual analysis provided in an embodiment of the present invention.

[0008] Figure 2 This is a schematic diagram of exemplary hardware and software components of an intelligent display control system based on AI visual analysis provided in an embodiment of the present invention. Detailed Implementation

[0009] Figure 1This is a flowchart illustrating an intelligent display control method based on AI visual analysis provided in one embodiment of the present invention, which will be described in detail below.

[0010] Step S110: Obtain the initial visual data stream set transmitted by multiple visual acquisition devices deployed in the display area. The initial visual data stream set includes multiple visual data frame units with acquisition time sequence marks that are acquired continuously.

[0011] For example, in a smart exhibition control scenario applied to a science museum, the exhibition area comprises multiple independent exhibit display units. The initial visual data stream collection includes multiple continuously acquired visual data frame units with acquisition time sequence markers. Each visual data frame unit corresponds to a snapshot at a fixed time interval, containing the distribution and posture information of all observed objects within the exhibition area at that moment. The acquisition of all visual data has been accompanied by informed consent from personnel entering the exhibition area, and biometric features such as faces have been anonymized, complying with data privacy protection regulations. For instance, clear notice boards and electronic screens are deployed at the entrance and in prominent locations within the exhibition area to inform personnel entering the area of ​​the purpose, scope, storage method, and retention period of the visual data acquisition. The acquisition purpose is limited to optimizing the exhibition effect and controlling exhibits, and will not be used for any other commercial purposes. Simultaneously, a QR code link is provided on the electronic screen for visitors to view the full privacy policy text.

[0012] Furthermore, this embodiment employs real-time anonymization processing technology based on face detection. Upon entering the processing pipeline, the raw video streams acquired by all visual acquisition devices undergo a face detection algorithm run through edge computing nodes. Detected face regions are Gaussian blurred with a blur kernel radius set to 0.1 times the pixel width, rendering facial features unrecognizable. For detected body regions, contour simplification is similarly applied, retaining only the coordinates of key joints used for behavior analysis while discarding original clothing textures and surface details. Only the anonymized data stream then proceeds to subsequent steps S111 to S116 for further processing.

[0013] For visual data requiring long-term storage, a tiered privacy protection strategy is employed. Raw anonymized data is retained for no more than 30 days; data exceeding this period is automatically deleted. If offline algorithm training is required, only non-personally identifiable statistical features extracted from the data (such as group density and heatmaps of points of interest) are stored; no trajectory data traceable back to individuals is stored. Data storage employs AES-256 encryption, with the key managed by an independent key management service, and access logs recording all data read operations.

[0014] In addition, a data subject rights response mechanism will be established. If any visitor raises an objection, they can submit a data deletion request through the on-site service desk or online channels. Within 24 hours of receiving the request, all data related to that person will be deleted from all storage media. For anonymized data that cannot be accurately identified, the data for the time period of that person's entry and exit will be excluded from all datasets.

[0015] Furthermore, prior to the deployment of the technical solution, a third-party security assessment agency was commissioned to conduct a compliance audit of the data collection and processing procedures, obtaining a personal information security impact assessment report. All data collection and processing operations were conducted after obtaining filing with the relevant regulatory authorities, and data security self-inspection reports were submitted to the regulatory authorities regularly. The collected visual data was limited to internal use within this system and was not transmitted or shared to any third party, nor used for any other purpose other than display control.

[0016] Step S111: Receive the global field of view visual data stream transmitted by the global coverage visual acquisition device deployed at the top of the display area, and the directional field of view visual data stream transmitted by the directional visual acquisition device deployed in front of the independent exhibit display area. The global field of view visual data stream is composed of global field of view visual data frame units arranged in order according to the global acquisition timing mark, and the directional field of view visual data stream is composed of directional field of view visual data frame units arranged in order according to the directional acquisition timing mark.

[0017] The global field-of-view visual data stream consists of global field-of-view visual data frame units arranged in a fixed global acquisition timing marker (e.g., frame number), covering the macroscopic scene of the entire display area. The global acquisition timing marker is generated by the camera's internal clock and synchronized with the network time protocol server. It also receives directional field-of-view visual data streams transmitted from directional cameras deployed directly in front of each individual exhibit display area. The directional field-of-view visual data frame units in this stream are arranged in the order of the directional acquisition timing marker, focusing on the details of a single exhibit and its surrounding objects. The directional acquisition timing marker is also synchronized with the network time protocol server, ensuring it is on the same time reference as the global acquisition timing marker.

[0018] Step S112: Based on the same global acquisition timing mark and directional acquisition timing mark, establish a correspondence sequence between global visual data frame units and directional visual data frame units. After establishing the correspondence sequence, perform frame-level combination processing on the global field-of-view visual data frame units and directional field-of-view visual data frame units according to the time correspondence to generate a composite visual data frame unit containing global field-of-view content and directional field-of-view content at each acquisition time.

[0019] Iterate through all global field-of-view (GVR) visual data frame units in the global field-of-view (GFVR) visual data stream. For each GVR visual data frame unit, extract its global acquisition timing marker value Tg. Locate the directional field-of-view (DOG) visual data frame unit in the directional field-of-view (DOG) visual data stream whose directional acquisition timing marker value equals Tg. Associate the found GVR visual data frame unit and DOG visual data frame unit to form a pair. Arrange all pairs in ascending order of Tg to form a pair sequence. For each pair in the pair sequence, combine the pixel data from the GVR visual data frame unit and the DOG visual data frame unit in memory. During combination, create a new image buffer with a size equal to the size of the GVR image. Copy all pixels from the GVR image into this buffer. Then, based on the position of the exhibit corresponding to the DOG image in the GVR, scale the DOG image to the appropriate size and overlay it onto the corresponding area of ​​the GVR image. The combined result is a composite visual data frame unit.

[0020] Step S113: Perform observation object group detection processing on the global field of view content in the composite visual data frame unit to obtain the group distribution position contour description data of the observation object group in the global field of view content; and perform observation object single-unit detection processing on the directional field of view content in the composite visual data frame unit to obtain the single-unit pose direction description data of the observation object single unit in the directional field of view content.

[0021] The YOLOv8 object detection algorithm is used to process the global field of view of the composite vision data frame unit. It outputs the bounding box coordinates of all detected objects, with each bounding box represented by its top-left corner coordinates (x1, y1) and bottom-right corner coordinates (x2, y2). The output bounding boxes are then clustered according to their spatial location: the Euclidean distance between the center points of every two bounding boxes is calculated. If d is less than the preset group clustering threshold Dt, the observations corresponding to the two bounding boxes are grouped into the same group. For each group, all bounding boxes contained therein are extracted, and the convex hull of the center points of these bounding boxes is calculated using the Graham scan method. The pixel coordinate sequence of the vertices of the convex hull polygon is used as the contour description data of the group distribution location. The OpenPose pose estimation network is used to process the orientation field content of the composite visual data frame unit. The OpenPose pose estimation network detects the coordinates of human key points, calculates the body orientation angle θb=atan2(y1-y2, x1-x2) based on the left shoulder key point (x1, y1) and the right shoulder key point (x2, y2), calculates the head orientation angle θh=atan2(yn-ye, xn-xe) based on the nose key point (xn, yn) and the eye key point (xe, ye), and calculates the gaze direction vector Vg=(xg-xe, yg-ye, zg-ze) based on the eye key point and the fixation point, and combines them into individual pose direction description data.

[0022] Step S114: Combine the group distribution location contour description data and the individual attitude direction description data at the same acquisition time into the spatiotemporal state description record of the observed object, and write the spatiotemporal state description record of the observed object into the corresponding composite visual data frame unit.

[0023] Create a data structure containing a group field and individual field. The group field stores a list of vertex coordinates of the convex hull polygon, and the individual field stores three components for each observed object: body orientation angle θb, head orientation angle θh, and gaze direction vector (Vx, Vy, Vz). Attach this data structure as metadata to the corresponding composite vision data frame unit.

[0024] Step S115: Arrange the composite visual data frame units containing the spatiotemporal state description records of the observed object at all consecutive acquisition times into a frame sequence according to the global acquisition timing mark, generate an initial visual data stream set containing consecutive visual data frame units, and each visual data frame unit is accompanied by the spatiotemporal state description record of the observed object, extract the global acquisition timing mark of each visual data frame unit in the initial visual data stream set, convert the global acquisition timing mark into an acquisition timing mark value with a unified time reference, and save the acquisition timing mark value with the unified time reference as the unique acquisition timing mark of the visual data frame unit.

[0025] Traverse all composite visual data frame units and read the global acquisition timing marker Tg for each unit. Convert Tg to a millisecond value Tu = Tg + Toff, calculated from 00:00:00 UTC on January 1, 1970, where Toff is the offset between the original clock and UTC. Use this Tu as the unique identifier for the composite visual data frame unit. Sort all composite visual data frame units in ascending order of Tu to form a frame sequence, i.e., the initial visual data stream set.

[0026] Step S116: The initial visual data stream set after unified time reference conversion is divided into storage areas. The initial visual data stream set is divided into multiple visual data stream segments according to the continuous acquisition time segments, and the visual data stream segments are associated with the corresponding unified time reference acquisition time sequence mark range.

[0027] Set the time segment length parameter Le. Starting from the first frame of the initial visual data stream set, divide all frames with timestamps falling within the range [0, Le) into the first visual data stream segment. Divide all frames with timestamps falling within the range [Le, 2Le) into the second visual data stream segment, and so on. Create an index table to record the start timestamp Ts and end timestamp Te of each segment.

[0028] Step S120: Perform a display scene structure deconstruction process on the initial visual data stream set to obtain the temporal observation tendency characteristics of the observed objects in the display scene and the spatial layout constraint characteristics of the display scene. The temporal observation tendency characteristics describe the displacement trajectory of the observation object's point of interest in a continuous visual data frame unit, and the spatial layout constraint characteristics describe the relative relationship between the display position of exhibits and the activity area of ​​the observed objects in the display scene.

[0029] Step S121: Extract the spatiotemporal state description record of the observed object from each visual data frame unit of the initial visual data stream set. The spatiotemporal state description record of the observed object includes the group distribution position contour description data and the individual pose direction description data of the visual data frame unit.

[0030] Traverse each visual data frame unit in the initial visual data stream set, read the spatiotemporal state description record of the observed object in its metadata, and extract the list of vertex coordinates of the convex hull polygon and θb, θh, (Vx, Vy, Vz) of each observed object.

[0031] Step S122: Based on the observation object's head orientation angle parameter and the observation object's line-of-sight vector parameter in the single-object posture direction description data, construct a three-dimensional line-of-sight ray representation of the observation object at each acquisition time. The three-dimensional line-of-sight ray representation includes the ray origin position information and the ray unit direction vector information.

[0032] For each individual observation object, obtain the head orientation angle parameters (yaw angle Yw, pitch angle Pp) and the line-of-sight vector Ve. Calculate the three-dimensional spatial coordinates (Xe, Ye, Ze) of the center points of the observation object's eyes using back projection based on the camera calibration parameters, and use this as the ray origin O. The back projection formula is: Where R is the rotation matrix, T is the translation vector, and d is the depth value. Calculate the unit direction vector of the ray. Rays are represented as .

[0033] Step S123: Perform spatial intersection calculation processing between the three-dimensional line-of-sight ray representation and the preset three-dimensional space model of the display scene to obtain the coordinates of the line-of-sight landing point of the observed object in the three-dimensional space model of the display scene. The coordinates of the line-of-sight landing point include horizontal axis coordinates, horizontal axis coordinates, and vertical height coordinates.

[0034] A triangular mesh of faces is pre-constructed to display the 3D spatial model of the scene. For each individual observed object, a ray... The intersection point is calculated using the Möller-Trumbore ray-triangle intersection algorithm, and its solution equation is... Let V0, V1, and V2 be the vertices of the triangle. Solve for t, u, and v. If t > 0, u >= 0, v >= 0, and u + v <= 1, then an intersection point exists. Traverse all triangular faces and find the intersection point corresponding to the minimum value of t > 0, recording its coordinates (Xh, Yh, Zh).

[0035] Step S124: Collect the coordinates of the line-of-sight points of the same observation object at multiple consecutive acquisition times. String the coordinates of the line-of-sight points in sequence according to the acquisition time to form a line-of-sight point coordinate sequence. Use this line-of-sight point coordinate sequence as the individual attention point displacement trajectory of the observation object. Calculate the spatial position dispersion and motion direction consistency parameters among the individual attention point displacement trajectories of all observation objects within the same observation object group. Merge the individual attention point displacement trajectories into a group attention point displacement trajectory based on the spatial position dispersion and motion direction consistency parameters.

[0036] For each individual observation subject, collect the gaze point coordinates P1, P2, ..., PT at T consecutive moments to form an individual attention point displacement trajectory. For all individuals in the same group, calculate the average spatial distance between the trajectories of any two individuals , which is taken as the spatial position dispersion. Calculate the average value of the included angles between the trajectory direction vectors of any two individuals , which is taken as the motion direction consistency parameter. If Ds<Dth and Ac<Ath, the trajectories of the above individuals are merged: at each moment, the arithmetic mean of the gaze point coordinates of all individuals is taken as the group gaze point coordinate, so as to form a group attention point displacement trajectory.

[0037] Step S125: Perform direction gradient change analysis on the displacement direction vectors of consecutive collection moments in the group attention point displacement trajectory, extract the trajectory inflection point moments and trajectory inflection point coordinates where the displacement direction change angle exceeds the preset angle threshold, generate attention point displacement trajectory morphology description data containing trajectory inflection point moment markers and trajectory inflection point coordinate markers, and take the attention point displacement trajectory morphology description data as the time-series observation tendency feature of the observation subject.

[0038] Calculate the displacement vector Vi=Pi+1-Pi between adjacent moments in the group attention point displacement trajectory. Calculate the included angle θi between adjacent displacement vectors as θi=arccos((Vi·Vi-1) / (|Vi||Vi-1|)). If θi>30 degrees, the moment i is marked as a trajectory inflection point moment, and Pi is marked as a trajectory inflection point coordinate. Record the moments and coordinates of all inflection points.

[0039] Step S126: Extract the region boundary coordinate sequence and the region center point position coordinate of each exhibit display area from the preset three-dimensional space model of the display scene, generate the non-traversable space constraint range description of each exhibit display area according to the region boundary coordinate sequence, and extract the region boundary coordinate sequence of the observation subject activity area from the preset three-dimensional space model of the display scene, and generate the free movement space range description for the observation subject according to the region boundary coordinate sequence of the observation subject activity area.

[0040] Read the boundary polygon vertex coordinate sequence (x1, y1, z1)~(xn, yn, zn) of each exhibit display area from the three-dimensional space model of the display scene, and take the inner area of the polygon as the non-traversable space constraint range description. Read the boundary polygon vertex coordinate sequence of the observation subject activity area, and take the inner area of the polygon as the free movement space range description for the observation subject.

[0041] Step S127: Perform spatial position relationship superposition analysis on the description of the inaccessible spatial constraint range and the description of the freely movable spatial range of the observed object, calculate the shortest spatial distance value between the boundary of the exhibit display area and the boundary of the activity area of ​​the observed object, and calculate the azimuth angle value of the center point of the exhibit display area relative to the boundary of the activity area of ​​the observed object. Combine the shortest spatial distance value and the azimuth angle value to form the spatial relative relationship description parameters between the exhibit display position and the activity area of ​​the observed object, and use the set of spatial relative relationship description parameters corresponding to all exhibit display areas as the spatial layout constraint features of the display scene.

[0042] For each exhibit display area, calculate the shortest Euclidean distance Dm = min|Pa - Pb| between its boundary polygon and the boundary polygon of the activity area of ​​the observed object, where Pa belongs to the exhibit boundary point set and Pb belongs to the activity area boundary point set. Calculate the coordinates of the center point of the exhibit display area Pc = ((x1 + ... + xn) / n, (y1 + ... + yn) / n, (z1 + ... + zn) / n). Calculate the point Pb on the activity area boundary polygon that is closest to Pc, and calculate the direction angle Az from vector Pc to Pb = atan2(Pb_y - Pc_y, Pb_x - Pc_x). Use (Dm, Az) as the spatial relative relationship description parameter for this exhibit. These parameters for all exhibits constitute the spatial layout constraint features.

[0043] Step S130: Call the pre-built visual attention analysis network to perform joint correlation analysis on the temporal observation tendency features and the spatial layout constraint features to generate a description of the distribution of hot and cold areas of visual attention in the current display scene.

[0044] Step S131: Input the data describing the displacement trajectory morphology of the attention point in the temporal observation tendency features into the trajectory feature encoding branch of the visual attention analysis network. The trajectory feature encoding branch consists of multiple cascaded trajectory convolution processing layers and trajectory pooling processing layers. The trajectory convolution processing layer extracts the trajectory curvature change features and trajectory pause duration distribution features of the displacement trajectory of the attention point. The trajectory pooling processing layer compresses the temporal dimension redundancy information of the trajectory features to generate a trajectory feature encoding vector.

[0045] The trajectory feature encoding branch adopts a temporal convolutional network, which includes 3 one-dimensional convolutional layers and 2 max pooling layers. The first convolutional layer uses 64 convolution kernels with a size of 3x1, the second convolutional layer uses 128 convolution kernels with a size of 3x1, and the third convolutional layer uses 256 convolution kernels with a size of 3x1. Each convolutional layer is followed by a batch normalization layer and a rectified linear unit activation layer. The window size of the pooling layer is 2x1, and the stride is 2. The input is a two-dimensional matrix formed by splicing the sequence of trajectory inflection point moments and the sequence of inflection point coordinates, with a size of T×5 (T is the number of trajectory points, and 5 corresponds to x, y, z, moment, and speed). The output is a 256-dimensional trajectory feature encoding vector Vt.

[0046] Step S132: Input the set of spatial relative relationship description parameters in the spatial layout constraint features into the spatial feature encoding branch of the visual attention analysis network. The spatial feature encoding branch is composed of a plurality of cascaded spatial graph convolution processing layers. Each spatial graph convolution processing layer takes the exhibit display area as a graph node, constructs node connection edges according to the shortest spatial distance value in the spatial relative relationship description parameters, extracts the spatial structure correlation features between the exhibit display area nodes through the spatial graph convolution processing layers, and generates a spatial feature encoding vector.

[0047] The spatial feature encoding branch adopts a graph convolutional network, which includes 2 graph convolutional layers. The first layer takes each exhibit display area as a graph node, and the initial feature of the node is (Dm, Az). Construct the adjacency matrix A according to Dm: if Dm < Dc, then Aij=1 / Dm, otherwise Aij=0. The graph convolution operation is , where Ã=A+I, and D̃ is the degree matrix of Ã. The first layer outputs 128 dimensions, and the second layer outputs 128 dimensions. After global average pooling, a 128-dimensional spatial feature encoding vector Vs is obtained.

[0048] Step S133: Input both the trajectory feature encoding vector output by the trajectory feature encoding branch and the spatial feature encoding vector output by the spatial feature encoding branch into the joint correlation analysis layer of the visual attention analysis network, and calculate the correlation response intensity value of each position dimension in the trajectory feature encoding vector and each exhibit node dimension in the spatial feature encoding vector through a multi-head cross-attention processing mechanism.

[0049] The joint correlation analysis layer adopts multi-head cross-attention. Let Vt be the query Q, and Vs be the key K and value V. Calculate the attention weight , and the output Attention=AV. 8 attention heads are used, each head outputs 32 dimensions, and after concatenation, a correlation response intensity value matrix R is obtained through linear transformation, with a dimension of 256×N, where N is the number of exhibit nodes.

[0050] Step S134: Perform weighted fusion processing on the associated response intensity value and preset exhibit visual saliency prior information, where the exhibit visual saliency prior information includes the exhibit size and area parameter, the exhibit surface brightness parameter, and the exhibit-background color difference parameter of the exhibit display area. Through the weighted fusion processing, a corrected associated response intensity value fused with the prior information is generated and input into the attention heat distribution generation layer of the visual attention analysis network, and voxel-by-voxel attention assignment processing is performed on the 3D space model of the display scene according to the corrected associated response intensity value, so as to generate the attention heat value of each spatial voxel unit in the 3D space model of the display scene.

[0051] Exhibit size and area parameter As = width of the exhibit bounding box × height of the exhibit bounding box. Exhibit surface brightness parameter Lb = average pixel gray value of the exhibit area. Exhibit-background color difference parameter Cd = |average RGB of the exhibit - average RGB of the background|. They are combined into a prior vector P=(As, Lb, Cd). Calculate the corrected associated response intensity value , β=0.3. Discretize the 3D space model of the display scene into voxel units with a side length of Lv. For each voxel unit, find the exhibit node j closest to it, and take R'(j) of this node as the attention heat value Hv of this voxel unit.

[0052] Step S135: Perform spatial clustering processing on the attention heat values of all spatial voxel units in the 3D space model of the display scene, aggregate the spatial voxel units whose attention heat values exceed the preset heat threshold and are adjacent in spatial position into attention hot spot regions, and aggregate the spatial voxel units whose attention heat values are lower than the preset cold spot threshold and are connected in spatial position into attention cold spot regions.

[0053] Set the heat threshold Th=0.7 and the cold spot threshold Tc=0.3. Adopt a 3D connected component labeling algorithm, mark all voxel units with Hv>Th as hot spot candidates, and mark all voxel units with Hv<Tc as cold spot candidates. Perform 26-neighborhood connected component clustering on the hot spot candidate voxels, and each connected component is a hot spot region. Perform 26-neighborhood connected component clustering on the cold spot candidate voxels, and each connected component is a cold spot region.

[0054] Step S136: Extract the regional boundary voxel coordinate sequence of the attention hot spot region and calculate the average attention heat value inside the attention hot spot region, and write the regional boundary voxel coordinate sequence and the regional internal average attention heat value as hot spot region description parameters into the hot spot region recording unit.

[0055] For each hot spot region, extract the 3D coordinate sequence of its boundary voxels (extract the isosurface using the marching cubes algorithm), and calculate the arithmetic mean Ha of the attention heat values of all voxels in the region. Store the boundary coordinate sequence and Ha.

[0056] Step S137: Extract the region boundary voxel coordinate sequence of the cold attention region and calculate the average attention heat value inside the region of the cold attention region. Write the region boundary voxel coordinate sequence and the average attention heat value inside the region as cold attention region description parameters into the cold attention region recording unit.

[0057] For each cold spot region, extract the 3D coordinate sequence of its boundary voxels, and calculate the arithmetic mean Ca of the heat values ​​of interest for all voxels within the region. Store the boundary coordinate sequence and Ca.

[0058] Step S138: Perform spatial location deduplication and region boundary smoothing on all hot spot region recording units and all cold spot region recording units to obtain a set of non-overlapping hot spot regions and a set of non-overlapping cold spot regions. Then, perform unified data structure encapsulation to generate a cold and hot spot region distribution description result that includes a set of hot spot region description parameters and a set of cold spot region description parameters, as well as visual attention.

[0059] Traverse all hot and cold regions, checking for spatial overlap. If overlap exists, compare the thermal value of the overlapping region with Tc; if it is greater than Tc, it is classified as a hot spot; otherwise, it is classified as a cold spot. Smooth the region boundaries using a 3D morphological closing operation (dilation followed by erosion, with a 3x3x3 cube as the structuring element). Encapsulate all processed regions into a result data structure.

[0060] Step S140: Based on the description of the distribution of hot and cold areas of visual attention, determine the exhibit identification information of the exhibit to be adjusted and the position migration vector sequence of the exhibit attention in the display scene space.

[0061] Step S141: Analyze the set of hot spot region description parameters in the distribution description results of hot and cold spots of visual attention, extract the region boundary voxel coordinate sequence and the average attention heat value inside the region from each hot spot region description parameter, and calculate the spatial coordinates of the region geometric center of each hot spot region based on the region boundary voxel coordinate sequence.

[0062] For each hotspot region, take the minimum x-coordinate xmin and maximum xmax, the minimum y-coordinate ymin and maximum ymax, and the minimum z-coordinate zmin and maximum zmax from the region boundary voxel coordinate sequence, and calculate the geometric center Ch = ((xmin + xmax) / 2, (ymin + ymax) / 2, (zmin + zmax) / 2).

[0063] Step S142: Parse the cold point region description parameter set in the distribution description result of cold and hot spot regions of the visual attention degree, extract the region boundary voxel coordinate sequence and the average attention heat value inside the region in each cold point region description parameter, and calculate the spatial coordinate of the regional geometric center of each cold point region according to the region boundary voxel coordinate sequence.

[0064] Similarly, for each cold point region, the geometric center Cc is calculated.

[0065] Step S143: Perform spatial distance matching calculation between the spatial coordinate of the regional geometric center of each hot spot region and the preset position coordinate of the center of the exhibit display area, select the exhibit display area whose spatial distance value is less than the preset distance matching threshold as the candidate hot spot associated exhibit, and record the exhibit identification information of the candidate hot spot associated exhibit as the exhibit identification information of the to-be-adjusted exhibit; and perform spatial distance matching calculation between the spatial coordinate of the regional geometric center of each cold point region and the preset position coordinate of the center of the exhibit display area, select the exhibit display area whose spatial distance value is less than the preset distance matching threshold as the candidate cold point associated exhibit, and record the exhibit identification information of the candidate cold point associated exhibit as the exhibit identification information of the to-be-adjusted exhibit.

[0066] For each hot spot region center Ch, calculate the distance d=|Ch-E| from each exhibit display area center E. If d is less than Dm, record the exhibit identifier Id as the to-be-adjusted exhibit. The cold point regions are processed in the same manner.

[0067] Step S144: For the hot spot region corresponding to each hot spot associated exhibit, extract the spatial coordinates of the regional geometric center of the hot spot region at a plurality of consecutive collection moments, connect the spatial coordinates of the regional geometric center in the order of collection moments to form a hot spot region center coordinate time series, perform movement direction trend extraction processing on the hot spot region center coordinate time series, calculate the displacement direction unit vector and the displacement distance scalar value between the spatial coordinates of the hot spot region center at adjacent collection moments, and generate a hot spot region movement direction vector sequence including the displacement direction unit vector and the displacement distance scalar value.

[0068] Extract the center coordinates Ch1, Ch2, ..., ChT of the hot spot region at T moments. Calculate the displacement vector Vi=Chi+1-Chi. Calculate the displacement direction unit vector Ui=Vi / |Vi|, and the displacement distance scalar value Di=|Vi|. Generate the movement direction vector sequence {(Ui, Di)}.

[0069] Step S145: Based on the hotspot area movement direction vector sequence, predict the predicted coordinates of the hotspot area center position within a preset time step in the future, connect the spatial coordinates of the hotspot area center at the current acquisition time with the predicted coordinates of the hotspot area center position within a preset time step in the future, and generate the location migration vector of the hotspot area attention.

[0070] Linear prediction is used to predict the displacement vector. Predict center coordinates Where Δt is the future time step and Δts is the sampling interval. The generated location migration vector is Mh = Chp - Ch_current.

[0071] Step S146: For each cold-point associated exhibit, extract the spatial coordinates of the geometric center of the cold-point area at multiple consecutive acquisition times. Serialize these coordinates to form a time series of the cold-point area center coordinates. Perform movement direction trend extraction processing on this time series, calculating the unit vector of displacement direction and the scalar value of displacement distance between the cold-point area center coordinates at adjacent acquisition times. Generate a cold-point area movement direction vector sequence containing the unit vector of displacement direction and the scalar value of displacement distance. Predict the predicted coordinates of the cold-point area center position within a preset time step based on this movement direction vector sequence. Connect the current acquisition time's cold-point area center coordinates with the predicted coordinates of the cold-point area center position within the preset time step to generate a position migration vector of the cold-point area's attention. Combine the position migration vectors of the hot-point area's attention and the cold-point area's attention to form a position migration vector sequence of exhibit attention in the display scene space.

[0072] Similarly, steps S144 and S145 process the cold spot region, generating a cold spot region location migration vector Mc. Mh and Mc are combined into a sequence {Mh1, Mh2, ..., Mc1, Mc2, ...}.

[0073] Step S150: Generate a set of display device control instructions containing the exhibit identification code and angle adjustment parameters based on the exhibit identification information and the position migration vector sequence, and distribute the set of display device control instructions to the corresponding exhibit display adjustment components.

[0074] Step S151: Extract the unique exhibit identification code of each exhibit to be adjusted from the exhibit identification information of the exhibit to be adjusted, query the preset exhibit display adjustment component mapping table according to the unique exhibit identification code, and obtain the network address of the display adjustment component and the set of adjustable parameter types of the display adjustment component bound to the unique exhibit identification code.

[0075] Extract the exhibit identification code (Id) from the exhibit identification information. Query the mapping table to obtain the network address (Addr) of the corresponding display adjustment component and the set of adjustable parameter types (Types), which includes horizontal rotation, vertical tilt, and height adjustment.

[0076] Step S152: Extract the position migration vector corresponding to the exhibit identification information of each exhibit to be adjusted from the position migration vector sequence. The position migration vector includes the spatial coordinates of the migration start point and the spatial coordinates of the migration end point. Calculate the direction vector of the position migration vector based on the spatial coordinates of the migration start point and the migration end point. Determine the target horizontal orientation angle of the horizontal rotation adjustment component in the exhibit display adjustment assembly based on the direction vector of the position migration vector. Calculate the difference between the target horizontal orientation angle and the current horizontal orientation angle of the exhibit display adjustment assembly, using it as a horizontal rotation angle parameter, and determine the corresponding horizontal rotation direction parameter. The horizontal rotation direction parameter is determined by the positive or negative sign value of this difference. Determine the target vertical pitch angle of the vertical pitch adjustment component in the exhibit display adjustment assembly based on the direction vector of the position migration vector. Calculate the difference between the target vertical pitch angle and the current vertical pitch angle of the exhibit display adjustment assembly, using it as a vertical pitch angle parameter, and determine the corresponding vertical pitch direction parameter. The vertical pitch direction parameter is determined by the positive or negative sign value of this difference.

[0077] For each exhibit to be adjusted, obtain its position migration vector M = (Ps, Pe). Calculate the direction vector Dm = Pe - Ps. The target horizontal orientation angle Az_t = atan2(Dm_y, Dm_x), the horizontal rotation angle parameter dAz = Az_t - Az_c, and the horizontal rotation direction parameter Sz = sign(dAz). The target vertical pitch angle El_t = atan2(Dm_z, The vertical pitch angle parameter dEl = El_t - El_c, and the vertical pitch direction parameter Se = sign(dEl).

[0078] Step S153: Extract the vertical height coordinates of the starting point and ending point from the location migration vector. Calculate the vertical height difference between the ending point and the starting point. Use the absolute value of the vertical height difference as the height rise / fall distance parameter. Determine the height rise / fall direction parameter based on the sign of the vertical height difference. Encapsulate the horizontal rotation direction parameter, horizontal rotation angle parameter, vertical pitch direction parameter, vertical pitch angle parameter, height rise / fall direction parameter, and height rise / fall distance parameter with the corresponding unique exhibit identifier to generate an independent control instruction data packet for each exhibit to be adjusted.

[0079] Extract the Z-coordinate Z_s of Ps and the Z-coordinate Z_e of Pe. Calculate the vertical height difference dZ = Z_e - Z_s. Calculate the height rise / fall distance parameter Hd = |dZ| and the height rise / fall direction parameter Sh = sign(dZ). Encapsulate (Sz, dAz, Se, dEl, Sh, Hd) and the exhibit identification code Id into an independent control command data packet.

[0080] Step S154: Write the control parameter values ​​in the independent control instruction data packet into the instruction parameter field that matches the set of controllable parameter types of the display control component, and add instruction execution timing mark and instruction priority mark to form a set of display device control instructions containing exhibit identification code and spatial coordinate offset.

[0081] Parameters are written to the corresponding fields according to Types: horizontal rotation is written to dAz, vertical pitch is written to dEl, and height increase / decrease is written to Hd. A current timestamp is added as an execution timing marker, and priority is set to 1 (high priority). All independent instructions constitute the display device control instruction set.

[0082] Step S155: Establish a bidirectional data transmission channel between the display adjustment component and the exhibit display adjustment component based on the network address of the display adjustment component. Send the independent control instruction data packet corresponding to the control instruction set of the display device to the instruction receiving buffer of the exhibit display adjustment component through the bidirectional data transmission channel. The exhibit display adjustment component reads the independent control instruction data packet from the instruction receiving buffer and parses the control parameter value. Drive the horizontal rotation adjustment component to perform horizontal rotation movement, drive the vertical pitch adjustment component to perform vertical pitch movement, and drive the height adjustment component to perform height adjustment movement to adjust the orientation of the exhibit to be consistent with the direction of the attention position migration vector.

[0083] A transmission control protocol connection is established based on the network address Addr. Independent control command data packets are sent through this connection. Upon receiving the data, the exhibit display adjustment component parses the parameters and controls the servo motors to execute the corresponding movements. These include the horizontal rotation motor's rotation angle dAz, the vertical pitch motor's rotation angle dEl, and the lifting motor's movement distance Hd.

[0084] Step S210: After the preset adjustment stabilization period ends following the execution of the adjustment action by the exhibit display adjustment component, the set of initial visual data streams after adjustment transmitted by multiple visual acquisition devices deployed in the display area is reacquired. The set of initial visual data streams after adjustment includes multiple visual data frame units with adjustment acquisition time sequence markers that are continuously acquired.

[0085] Set the adjustment stabilization period length parameter Td, for example, Td = 10 seconds. After the exhibit display adjustment component completes the adjustment action in step S155, wait for the Td time. Once the exhibit's physical position is stable, repeat steps S110 to S116 to obtain a new set of adjusted initial visual data streams. Each adjusted visual data frame unit in this set of adjusted initial visual data streams has an adjusted acquisition timing marker.

[0086] Step S211: Perform adjusted display scene construction deconstruction processing on the adjusted initial visual data stream set to obtain adjusted temporal observation tendency features of the observed objects in the adjusted display scene and spatial layout constraint features of the adjusted display scene. The adjusted temporal observation tendency features describe the adjusted attention point displacement trajectory of the observed objects in continuous adjusted visual data frame units, and the spatial layout constraint features describe the relative relationship between the display position of exhibits and the activity area of ​​the observed objects in the display scene.

[0087] Steps S120 to S127 are repeated on the adjusted initial visual data stream set to obtain the adjusted temporal observation tendency features and the adjusted spatial layout constraint features of the display scene. The displacement trajectory of the attention point in the adjusted temporal observation tendency features is denoted as the adjusted attention point displacement trajectory.

[0088] Step S212: Call the pre-built visual attention analysis network to perform joint correlation analysis on the adjusted temporal observation tendency features and the spatial layout constraint features, generate the adjusted cold and hot spot area distribution description results of visual attention in the current display scene, and calculate the visual attention distribution improvement degree evaluation parameters based on the adjusted cold and hot spot area distribution description results. The visual attention distribution improvement degree evaluation parameters include the cold spot area reduction ratio, the hot spot area attention heat value increase ratio, and the hot spot area overlap with the exhibit display area increase ratio.

[0089] Steps S130 to S138 are repeated for the adjusted temporal observation trend characteristics and spatial layout constraint characteristics to obtain the adjusted description results of the distribution of hot and cold areas. The total area Ab of cold areas before adjustment and the total area Aa of cold areas after adjustment are extracted, and the cold area reduction ratio Rc = (Ab - Aa) / Ab is calculated. The average attention heat value Hb of hot areas before adjustment and the average attention heat value Ha of hot areas after adjustment are extracted, and the attention heat value increase ratio Rh = (Ha - Hb) / Hb of hot areas is calculated. The overlapping area Ob between hot areas and exhibit display areas before adjustment and the overlapping area Oa after adjustment are extracted, and the overlap increase ratio Ro = (Oa - Ob) / Ob is calculated.

[0090] Step S213: Input the visual attention distribution improvement assessment parameters into the preset display effect feedback adjustment strategy module. The display effect feedback adjustment strategy module iteratively corrects the angle adjustment parameters in the display device control command set based on the cold spot area reduction ratio, the hot spot area attention heat value increase ratio, and the hot spot area overlap with the exhibit display area increase ratio. Based on the iteratively corrected angle adjustment parameters, the horizontal rotation angle parameters, vertical pitch angle parameters, and height rise and fall distance parameters of the exhibit display adjustment component are regenerated, generating the display device control command set after feedback correction.

[0091] The feedback adjustment strategy module for the demonstration effect adopts a PID control algorithm. Let the target improvement level be Rc_t, Rh_t, and Ro_t. Calculate the errors ec = Rc_t - Rc, eh = Rh_t - Rh, and eo = Ro_t - Ro. Calculate the correction amount. , The dAz, dEl, and dHd are superimposed on the original angle adjustment parameters to generate a new set of corrected control commands.

[0092] Step S214: The set of control commands for the display device after feedback correction is redistributed to the corresponding exhibit display adjustment components. The exhibit display adjustment components perform secondary adjustment actions based on the corrected horizontal rotation angle parameters, vertical pitch angle parameters, and height rise and fall distance parameters.

[0093] The revised set of display device control commands is sent to the exhibit display adjustment component in accordance with step S155 to perform secondary adjustment.

[0094] Step S215: Record the description results of the distribution of hot and cold areas after each round of adjustment and the evaluation parameters of the improvement degree of visual attention distribution before and after each round of adjustment action, forming a historical evaluation parameter record sequence of the exhibit display adjustment effect. Input the historical evaluation parameter record sequence into the preset exhibit display layout optimization analysis module. The exhibit display layout optimization analysis module generates exhibit static display position adjustment suggestion information based on the changing trend of hot and cold area distribution in the historical evaluation parameter record sequence.

[0095] The coordinates (Rc, Rh, Ro) of each round, along with the corresponding boundary coordinates of hot and cold areas, are stored in the historical record sequence. The exhibit display layout optimization analysis module uses time series analysis to calculate the movement trajectory of the center position of the cold area with each adjustment round, fitting a linear trend line. If the center of the cold area continues to move in a certain direction, it is recommended to move the exhibits near that area in that direction. The module generates suggested exhibit identification codes and the coordinates of the center point of the suggested adjusted exhibit display area.

[0096] Step S216: Send the proposed adjustment information for the static display position of the exhibits to the display management terminal device. The proposed adjustment information for the static display position of the exhibits includes the identifier code of the exhibit to be adjusted and the coordinates of the center point of the display area of ​​the exhibit after the proposed adjustment.

[0097] The suggested information for adjusting the static display position of exhibits generated in step S215 is sent to the exhibition management terminal device through a message queue or hypertext transfer protocol interface for exhibition management personnel to refer to and rearrange the physical position of exhibits.

[0098] Step S310: Perform observation object attribute classification processing on each visual data frame unit in the initial visual data stream set, extract the facial region visual feature data, body shape outline visual feature data and clothing color visual feature data of the observation object, and generate the observation object attribute feature vector based on the facial region visual feature data, body shape outline visual feature data and clothing color visual feature data.

[0099] For each visual data frame unit in the initial visual data stream set, a multi-task convolutional neural network (MTCNN) is first used to detect face regions in the image. MTCNN consists of three cascaded networks: a proposal network (P-Net) for quickly generating candidate windows, a refinement network (R-Net) for filtering out erroneous candidate windows, and an output network (O-Net) for outputting the final face bounding box and the coordinates of five facial key points (left eye, right eye, nose, left corner of mouth, and right corner of mouth). For the detected face regions, they are cropped and scaled to 160x160 pixels and input into the FaceNet network for facial feature extraction. The FaceNet network uses the Inception-ResNet-v1 architecture, trained using a triplet loss function, and outputs a 128-dimensional facial feature vector V_face. For body contour features, a DeepLabv3+ semantic segmentation network is used to segment the observed objects and extract their binary masks. The ratio of height to shoulder width, head-to-body ratio (the ratio of head height to body height), and body width-to-height ratio are calculated from the binary mask. These parameters are combined into a 32-dimensional body shape feature vector, V_body. For clothing color features, within the instance segmentation mask region, pixels are converted from the red-green-blue color space to the hue-saturation-lightness color space. Histograms of 8 bins are calculated for the hue channel, saturation channel, and lightness channel, resulting in a 24-dimensional clothing color feature vector, V_cloth. V_face, V_body, and V_cloth are concatenated into a 184-dimensional observation object attribute feature vector, V_attr = [V_face, V_body, V_cloth].

[0100] Step S311: Input the observation object attribute feature vectors at different acquisition times into the observation object re-identification and tracking network. The observation object re-identification and tracking network calculates the feature similarity distance value between the observation object attribute feature vectors at different acquisition times, and performs continuous identity binding processing for the same observation object across visual data frame units based on the feature similarity distance value. Establish a mapping association table between the continuous identity of the observation object and the temporal observation tendency characteristics of the observation object in each visual data frame unit. The mapping association table records the one-to-one correspondence between the continuous identity of the observation object and the individual attention point displacement trajectory.

[0101] The object re-identification and tracking network employs a metric learning-based architecture. For object A detected in frame i and object B detected in frame j, their attribute feature vectors V_A and V_B are extracted, respectively. Cosine similarity is then calculated. A similarity threshold T_sim = 0.65 is set. If S_AB > T_sim, then observation objects A and B are considered to be observations of the same person at different times. The Hungarian algorithm is used for cross-frame matching: a similarity matrix M is constructed, where M_ij represents the similarity between the p-th observation object in the i-th frame and the q-th observation object in the j-th frame. The maximum weight matching is solved, and successfully matched observation objects are assigned the same persistent identity ID. For newly appearing observation objects, a new ID is assigned. A mapping association table is established, with each ID corresponding to a list storing the sequence of individual attention point displacement trajectory coordinates for that ID at each time step.

[0102] Step S312: Based on the individual attention point displacement trajectory corresponding to the continuous identity identifier of each observed object in the mapping association table, extract the exhibit display area identifier sequence through which the observation object's line of sight coordinates passes during the complete visit period, and arrange the exhibit display area identifier sequence according to the time sequence of the line of sight to form an individual visit path sequence.

[0103] For each persistent identity ID, extract the coordinates (X, Y, Z) of all line-of-sight points from the displacement trajectory of its associated individual focus point. For each line-of-sight coordinate, traverse the pre-defined list of polygons representing the boundaries of the exhibit display area, and use a ray casting method to determine whether the line-of-sight point is located inside a certain exhibit display area. The ray casting method rule is: emit a horizontal ray to the right from the line-of-sight point, calculate the number of intersections between the ray and the polygon edges; if the number of intersections is odd, the line-of-sight point is inside the polygon, and if the number is even, the line-of-sight point is outside. Record the exhibit identification code corresponding to the line-of-sight point. Arrange all recorded exhibit identification codes in timestamp order to form an individual visitor path sequence Path=[Id1, Id2, ..., IdM].

[0104] Step S313: Calculate the path similarity parameters between individual visitor path sequences of all observed objects in the display scene, and classify the individual visitor path sequences into multiple visitor flow categories based on the path similarity parameters.

[0105] The Longest Common Subsequence (LCS) algorithm is used to calculate the similarity between two path sequences, Path_a and Path_b. The recursive formula for the LCS algorithm is: LCS(i,j)=0 if i=0 or j=0; LCS(i,j)=LCS(i-1,j-1)+1 if a_i=b_j; LCS(i,j)=max(LCS(i-1,j), LCS(i,j-1)) if a_i≠b_j. The path similarity Sim=LCS(Path_a, Path_b) / max(len(Path_a), len(Path_b)). The DBSCAN clustering algorithm is used to cluster the paths, setting the neighborhood radius ε=0.3 and the minimum number of samples MinPts=2. Paths with a distance less than ε are grouped into the same cluster, and each cluster corresponds to a visitor flow category.

[0106] Step S314: Count the number of observation objects in each visitor flow category, calculate the proportion of the number of observation objects in each visitor flow category among all observation objects, and mark the visitor flow categories whose proportion of the number of observation objects exceeds the preset mainstream flow proportion threshold as mainstream visitor flow.

[0107] Let Nk be the number of observations for the k-th streamline category, and Ntotal be the total number of observations. Calculate the proportion of this streamline category, Pk = Nk / Ntotal. Set the mainstream streamline proportion threshold, P_main = 0.3. If Pk > P_main, then mark this streamline category as the mainstream visitor streamline.

[0108] Step S315: Perform spatial overlay analysis on the sequence of exhibit display area identifiers along the main visitor flow and the exhibit display locations in the current display scene to identify exhibit display area identifiers located at the beginning and end of the main visitor flow.

[0109] Take the first exhibit identifier code from all path sequences in the main visitor flow, count the frequency of each exhibit identifier code, and take the exhibit identifier code with the highest frequency as the starting exhibit Id_start. Take the last exhibit identifier code from all path sequences, and take the exhibit identifier code with the highest frequency as the ending exhibit Id_end.

[0110] Step S316: Generate an exhibit display position pre-adjustment strategy according to the exhibit display area identifier located at the start end of the main visiting flow line, wherein the exhibit display position pre-adjustment strategy comprises an exhibit position exchange scheme for adjusting exhibits with a historical average attention heat value higher than a preset high attention threshold from a non-start end area to the start end area of the main visiting flow line.

[0111] Obtain the historical average attention heat value H_avg of each exhibit from historical statistical data. Set the high attention threshold H_high=0.75. Screen out exhibits with H_avg>H_high and the exhibit identification code not equal to Id_start, which are recorded as a candidate high-attention exhibit set S_high. Perform position exchange between the first exhibit in S_high and the exhibit corresponding to Id_start, so as to generate the pre-adjustment strategy.

[0112] Step S317: Generate an exhibit display position post-adjustment strategy according to the exhibit display area identifier located at the tail end of the main visiting flow line, wherein the exhibit display position post-adjustment strategy comprises an exhibit position exchange scheme for adjusting exhibits with a historical average attention heat value lower than a preset low attention threshold from the start end area to the tail end area of the main visiting flow line.

[0113] Set the low attention threshold L_low=0.25. Screen out exhibits with H_avg<L_low and the exhibit identification code not equal to Id_end, which are recorded as a candidate low-attention exhibit set S_low. Perform position exchange between the first exhibit in S_low and the exhibit corresponding to Id_end, so as to generate the post-adjustment strategy.

[0114] Step S318: Package the exhibit display position pre-adjustment strategy and the exhibit display position post-adjustment strategy to generate an exhibit static layout optimization scheme data packet, and transmit the exhibit static layout optimization scheme data packet to a display management terminal device to assist the adjustment decision of the physical display layout of exhibits.

[0115] Encapsulate the pre-adjustment strategy and the post-adjustment strategy into a data packet in JSON format, which comprises fields: adjustment type type (pre or post), source exhibit identification code src_id, target exhibit identification code dst_id, and suggested adjustment position coordinate pos. Send the data packet to the display management terminal device through a hypertext transfer protocol interface.

[0116] Step S410: Extract the region boundary voxel coordinate sequence from all cold spot region description parameters from the description results of the distribution of hot and cold spots in visual attention. Determine the display scene space range covered by the cold spot region based on the region boundary voxel coordinate sequence. Perform spatial position overlap judgment processing on the display scene space range covered by the cold spot region and the preset exhibit display area location information, and filter out the set of cold spot covered exhibit identifiers that have a spatial position overlap relationship with the cold spot region.

[0117] Iterate through the description parameters of each cold point region in the cold and hot point region distribution description results, and extract its region boundary voxel coordinate sequence. Calculate the minimum axis-aligned bounding box of the cold point region: x_min=min(x_i), x_max=max(x_i), y_min=min(y_i), y_max=max(y_i), z_min=min(z_i), z_max=max(z_i). For each exhibit display area, calculate its spatial bounding box as well. Determine if two bounding boxes intersect: if the projection intervals of two bounding boxes overlap in the X, Y, and Z directions, then the two regions are considered to have spatial overlap. Collect all exhibit identifiers that overlap with cold point regions into the cold point covered exhibit identifier set S_cold.

[0118] Step S411: For each cold-spot covered exhibit in the cold-spot covered exhibit identifier set, extract the exhibit description text data, exhibit image data set, and exhibit multimedia presentation material data set of the cold-spot covered exhibit from the preset exhibit information database.

[0119] For each exhibit identifier (Id) in S_cold, query the exhibit information database to retrieve the exhibit's description text field (Txt), image URL list (Img_urls), and multimedia presentation material URL list (Video_urls). Download or read the corresponding data content from the URLs.

[0120] Step S412: Invoke the pre-built dynamic content generation network to convert the exhibit description text data into a speech synthesis parameter sequence, convert the exhibit image data set into a dynamic image carousel control command sequence, and convert the exhibit multimedia demonstration material data set into a demonstration content trigger signal sequence.

[0121] The presentation content dynamic generation network comprises three sub-modules: **Text-to-Speech Sub-module:** This sub-module inputs the introductory text Txt into the Tacotron2 model, which consists of an encoder, an attention mechanism, a decoder, and a post-processing network. The encoder converts the text into a feature sequence, the decoder generates a Mel spectrogram through the attention mechanism, the post-processing network converts the Mel spectrogram into a linear spectrum, and finally, an audio waveform is synthesized using a WaveGlow vocoder to obtain the speech synthesis parameter sequence Sp (containing audio sampling point data). **Image Carousel Control Sub-module:** This sub-module arranges the images in the image URL list Img_urls sequentially, generating a carousel control instruction sequence Cp. Each instruction includes the image URL and display duration (e.g., 3 seconds per image). **Video Trigger Sub-module:** This sub-module generates a presentation content trigger signal sequence Vt, containing the video URL and playback start time offset.

[0122] Step S413: Determine the device identifier of the directional audio output device deployed near the cold spot area based on the spatial coordinates of the cold spot area, and send the speech synthesis parameter sequence to the directional audio output device. The directional audio output device plays the audio content introducing the exhibits according to the speech synthesis parameter sequence.

[0123] Calculate the geometric center coordinates Cc of the cold spot area: Cc = ((x_min+x_max) / 2, (y_min+y_max) / 2, (z_min+z_max) / 2). Iterate through all directional audio output devices deployed in the display area, calculate the Euclidean distance between each device's position coordinates and Cc, select the closest device, and obtain its device identifier Did. Send the speech synthesis parameter sequence Sp to the device via a network socket. Upon receiving the signal, the device decodes the audio data, converts it into an analog signal using a digital signal processor, and then drives the speaker to play the speech via a power amplifier.

[0124] Step S414: Determine the screen identifier of the auxiliary visual presentation screen deployed above the cold spot area based on the spatial coordinates of the cold spot area, and send the dynamic image carousel control command sequence and the demonstration content trigger signal sequence to the auxiliary visual presentation screen. The auxiliary visual presentation screen switches to display exhibit-related images according to the dynamic image carousel control command sequence, and starts the playback process of exhibit multimedia demonstration materials according to the demonstration content trigger signal sequence. During the playback of exhibit introduction voice content by the directional audio output device and the display of exhibit-related images by the auxiliary visual presentation screen, continuously acquire the local visual data stream transmitted by the directional visual acquisition device deployed near the cold spot area, and perform observation object attention state detection processing on the local visual data stream.

[0125] Similarly, based on the geometric center coordinates Cc of the cold spot area, the distance to each auxiliary visual presentation screen is calculated, the nearest screen is selected, and its screen identifier Sid is obtained. The dynamic image carousel control command sequence Cp and the presentation content trigger signal sequence Vt are sent to this screen. After receiving the signals, the screen loads and displays images sequentially according to the order in Cp, and starts video playback according to the time offset in Vt. Simultaneously, directional visual acquisition devices deployed near the cold spot area are activated to continuously acquire local visual data streams.

[0126] Step S415: The observation object attention state detection processing includes extracting the head orientation angle change rate parameter and the gaze direction stabilization duration parameter of the observation object from the local visual data stream. Based on the head orientation angle change rate parameter and the gaze direction stabilization duration parameter, it is determined whether the observation object's attention is attracted to the cold spot covered exhibit area. When the observation object's attention is attracted to the cold spot covered exhibit area, a content intervention fade-out control command sequence is generated to reduce the playback volume of the directional audio output device and the display brightness of the auxiliary visual presentation screen according to a preset attenuation curve. The content intervention fade-out control command sequence is sent to the directional audio output device and the auxiliary visual presentation screen to perform the fade-out operation.

[0127] Extract the head orientation angle sequence θ(t) for each observed object from the local visual data stream. Calculate the rate of change of head orientation angle dθ / dt = (θ(t+Δt)-θ(t)) / Δt. If dθ / dt < 5 degrees / second for 5 consecutive frames, the head orientation is considered stable. Extract the angle between the gaze direction vector and the normal vector of the exhibit area. If the angle is < 30 degrees and the duration exceeds 2 seconds, the gaze is considered stable. When both head orientation stability and gaze stability are satisfied, attention is considered to be attracted. Generate a fade-out control command: the volume decays exponentially. The brightness decreases according to a linear decay curve. Decrease, where τ = 2 seconds and T = 5 seconds. Send the fade-out command to the directional audio output device and the auxiliary visual display screen.

[0128] Step S416: After the content intervention fade-out operation is completed, update the cold spot covered exhibit area to the exhibit area that has been intervened in the attention enhancement process, and remove the cold spot area description parameter record corresponding to the cold spot covered exhibit from the cold spot area queue to be processed.

[0129] After the fade-out operation is completed, the exhibit identifier code of the cold spot covered is recorded in the processed set S_done. The corresponding cold spot area description parameter record is deleted from the cold spot area queue to be processed.

[0130] For example, the method may further include: step S510: obtaining the distribution description results of hot and cold areas of visual attention in different time periods within the display scene, constructing a time series hot and cold area distribution data set containing time period markers and corresponding hot and cold area distribution description results, inputting the time series hot and cold area distribution data set into a preset display effect periodic evaluation module, and the display effect periodic evaluation module calculating the rate of change of hot area area, the rate of change of average attention heat value of hot area, and the spatial movement speed value of hot area location under adjacent time periods.

[0131] The 24 hours of a day are divided into 24 time periods, each hourly. For each time period, steps S110 to S138 are performed to obtain the distribution description of hot and cold areas for that time period. For adjacent time periods t and t+1, the rate of change of the area of ​​the hot spot is calculated as R_area=(Area_{t+1}-Area_t) / Area_t, the rate of change of the average heat value of the hot spot is calculated as R_heat=(Heat_{t+1}-Heat_t) / Heat_t, and the spatial movement speed of the geometric center of the hot spot is calculated as V_move=Distance(Center_{t+1},Center_t) / 3600 seconds.

[0132] Step S511: Generate a time evolution feature vector of visual attention in the display scene based on the rate of change of the area of ​​the hot spot, the rate of change of the average attention heat value of the hot spot, and the spatial movement speed of the hot spot. Input the time evolution feature vector into a pre-constructed display strategy adaptive learning network. The display strategy adaptive learning network includes a gated recurrent processing layer and a fully connected output layer. The gated recurrent processing layer extracts the long-term evolution law features and short-term fluctuation features of attention in the time evolution feature vector. The fully connected output layer generates suggested values ​​for adjusting the display device control parameters based on the long-term evolution law features and short-term fluctuation features of attention. The suggested values ​​for adjusting the display device control parameters include suggested values ​​for adjusting the target horizontal orientation angle, the target vertical pitch angle, and the height rise and fall distance parameters calculated by the exhibit display adjustment component based on the position migration vector.

[0133] Arrange (R_area, R_heat, V_move) in chronological order to form a 24×3 temporal evolution feature matrix F. Input F into the presentation policy adaptive learning network. The gated recurrent processing layer of this network uses Gated Recurrent Units (GRUs) with 64 units. The update formulas for the GRUs are: z_t=σ(W_z·[h_{t-1}, x_t]), r_t=σ(W_r·[h_{t-1}, x_t]). The fully connected output layer is a 64×3 linear transform, outputting an adjustment suggestion vector [dAz_rec, dEl_rec, dHd_rec].

[0134] Step S512: Combine the target horizontal orientation angle adjustment amount and target vertical pitch angle adjustment amount indicated by the adjustment suggestion value of the display device control parameters with the current horizontal orientation angle and current vertical pitch angle of the current exhibit display adjustment component to calculate the new target horizontal orientation angle and target vertical pitch angle, and then generate a set of display device control instructions for the next display period. The smoothing weight coefficient in the calculation process is dynamically determined by the fluctuation amplitude value of the short-term fluctuation feature of attention in the time evolution feature vector.

[0135] Let the current horizontal orientation angle be Az_c, and the current vertical pitch angle be El_c. Calculate the short-term fluctuation characteristic of attention σ=std(R_area, R_heat, V_move). The smoothing weighting coefficient α=1 / (1+σ). The new target horizontal orientation angle... New target vertical pitch angle New altitude rise and fall distance Encapsulate (Az_new, El_new, Hd_new) into a set of display device control instructions for the next time period.

[0136] Step S513: Obtain the pedestrian density distribution data within the display scene. The pedestrian density distribution data is jointly generated by the counting visual acquisition device deployed at the entrance of the display area and the partition density statistics module deployed inside the display area, and includes the number of people in each display sub-area within a unit area.

[0137] A counting camera is deployed at the entrance of the exhibition area. The background subtraction method is used to count the number of people entering and leaving, and the total number of people in the area is calculated in real time. The exhibition area is divided into K sub-areas. A wide-angle camera is deployed in each sub-area, and the YOLOv8 algorithm is used to count the number of people in the sub-area. The population density of the sub-area is calculated as ρ_k = number of people_k / area_k.

[0138] Step S514: Perform correlation analysis on the crowd density distribution data and the time series hot and cold spot distribution data set, calculate the correlation response coefficient between the number of people in each display sub-area and the heat value of the hot spot area in the corresponding area, and generate a crowd density-guided exhibit control strategy based on the correlation response coefficient; the crowd density-guided exhibit control strategy includes setting the target horizontal orientation angle of the exhibit display adjustment component to the direction towards the area where the crowd density is lower than the preset density threshold, and calculating the deflection direction parameter and deflection angle parameter based on the current orientation, so as to guide the observation object from the area where the crowd density is higher than the preset density threshold to the area where the crowd density is lower than the preset density threshold through the orientation of the exhibit.

[0139] For each display sub-region k, obtain its population statistics N_k and the average attention heat value H_k of the hot spots within that region. Calculate the correlation response coefficient. Where Cov(N, H) is the covariance. If γ > 0.3, there is a positive correlation, indicating that areas with high pedestrian density also attract more attention. Determine the direction D_low of the sub-area with the lowest pedestrian density, and set the target horizontal orientation angle of the exhibit display adjustment components to point towards D_low. Calculate the angle θ between the current orientation and D_low, and generate the deflection direction parameter (clockwise or counterclockwise) and the deflection angle parameter θ.

[0140] Step S515: Perform strategy fusion processing on the crowd density-guided exhibit control strategy and the set of display device control instructions for the next display period to generate a set of display device control instructions fused with crowd density guidance, and distribute the set of display device control instructions fused with crowd density guidance to the corresponding exhibit display adjustment components.

[0141] The deflection angle parameter in the crowd density guidance strategy is weighted and fused with the angle parameter in the control command generated in step S512, with a fusion weight w = γ / (1+γ). The final target horizontal orientation angle is then determined. Az_final, El_new, and Hd_new are encapsulated into a set of display device control instructions that integrate crowd density guidance and are sent to the exhibit display adjustment component.

[0142] Step S610: Perform observation object group behavior pattern recognition processing on the continuous visual data frame units in the initial visual data stream set, extract the individual posture direction description data of all individual observation objects in the same observation object group, calculate the spatial discrete angle value of the individual observation object gaze direction vector in the group based on the individual observation object gaze direction vector parameter in the individual posture direction description data, and compare the spatial discrete angle value with a preset group co-view judgment angle threshold. When the spatial discrete angle value is less than the preset group co-view judgment angle threshold, mark the observation object group at the corresponding acquisition time as a co-view group state. When the spatial discrete angle value is not less than the preset group co-view judgment angle threshold, mark the observation object group at the corresponding acquisition time as a dispersed attention group state.

[0143] For each observation group, extract the unit vector of the gaze direction for all members within the group: Vi = (vx_i, vy_i, vz_i). Calculate the angle θ_ij between each pair of all vectors: θ_ij = arccos((Vi·Vj) / (|Vi||Vj|)). Take the arithmetic mean of all angles. Let the threshold for determining the shared view angle be θ_c = 15 degrees. If θ_avg < θ_c, mark it as a shared view group state; otherwise, mark it as a dispersed attention group state.

[0144] Step S611: For the observation object group marked as a shared viewing group, extract the displacement trajectory of the group attention point of the observation object group, and extract the group distribution position contour description data of the observation object group in the display scene space. Calculate the average station center coordinates of the group and the spatial distribution radius of the group members based on the group distribution position contour description data.

[0145] For a shared-view group, extract the displacement trajectory of its group focus point. Calculate the average station center coordinates Cg = (1 / m)ΣPi from the group distribution location contour description data (the list of convex hull polygon vertex coordinates P1, P2, ..., Pm). Calculate the spatial distribution radius of group members Rg = (1 / m)Σ|Pi - Cg|.

[0146] Step S612: For the observation object group marked as a dispersed attention group, extract the individual attention point displacement trajectory of each individual observation object in the observation object group, and calculate the trajectory angle value and trajectory spatial separation distance value between each pair of individual attention point displacement trajectories. Mark the subset of observation object individuals whose trajectory angle value exceeds the preset dispersed angle threshold and whose trajectory spatial separation distance value exceeds the preset separation distance threshold as the observation object subgroup to be evaluated.

[0147] For each individual observation object in the dispersed attention group, extract its individual attention point displacement trajectory. For any two individuals A and B, calculate the trajectory angle θ_AB = average(|direction_A - direction_B|), and calculate the trajectory spatial separation distance d_AB = average(|position_A - position_B|). Set the dispersion angle threshold θ_d = 30 degrees and the separation distance threshold d_s = 1.5 meters. Individual pairs satisfying θ_AB > θ_d and d_AB > d_s are marked as candidate split pairs. The subset consisting of the most frequently occurring individuals among the candidate split pairs is marked as the observation object subgroup to be evaluated.

[0148] Step S613: Input the individual attention point displacement trajectory of the observed object subgroup into the pre-constructed group splitting trend prediction network. The group splitting trend prediction network extracts the separation trend features of the individual attention point displacement trajectory through the temporal convolution processing layer, calculates the mutual influence intensity value of the separation trend features between different observed object individuals through the attention weighting processing layer, and outputs the prediction result of the complete splitting time of the group and the motion direction prediction vector of the split subgroup.

[0149] The group split trend prediction network employs a temporal convolutional network (TCN) and an attention mechanism. The TCN consists of 4 dilated convolutional layers with dilation rates of 1, 2, 4, 8 respectively, and a kernel size of 3. The input is the trajectory sequence of each individual within the sub-group (T×D, T=30 frames, D=2-dimensional position coordinates). The TCN outputs a split trend feature map. The attention weighting processing layer calculates the mutual influence weight between individuals α_ij=exp(e_ij) / Σexp(e_ik), where . The output is the predicted full group split time T_split (in seconds) and the predicted motion direction vector V_split of the sub-groups after splitting.

[0150] Step S614: When the predicted full group split time result is less than the preset intervention time threshold, a coordinated control instruction set for display devices oriented to the spatial area where the observed object group is located is generated. The coordinated control instruction set for display devices includes a coordinated rotation direction parameter sequence and a coordinated rotation angle parameter sequence for a plurality of exhibit display adjustment components located within the field of view in front of the observed object group; the coordinated rotation direction parameter sequence and the coordinated rotation angle parameter sequence for the plurality of exhibit display adjustment components are obtained by solving a multi-exhibit joint orientation optimization problem, whose optimization objective is: for each exhibit display adjustment component, calculate the direction vector from the center point of its exhibit display area to the geometric midpoint of the line connecting the center positions of the observed object sub-groups, take this direction vector as the target orientation of the exhibit, and then calculate the coordinated rotation direction parameter and coordinated rotation angle parameter according to its current orientation.

[0151] Set the intervention time threshold T_int=8 seconds. If T_split<T_int, generate coordinated control instructions. Let the positions of each individual in the observed object sub-group be Q1, ..., Qm, and calculate the geometric midpoint of the center point connection Pc=(1 / m)ΣQi. For each exhibit within the field of view in front of the group, let the center point of its exhibit display area be E_i. Calculate the target orientation vector D_i=Pc-E_i. Calculate the included angle θ_i between the current orientation of the exhibit and D_i, and generate the coordinated rotation direction parameter (clockwise or counterclockwise) and the coordinated rotation angle parameter θ_i. Form a sequence with the parameters of all exhibits.

[0152] Step S615: Send the coordinated control instruction set for display devices to all exhibit display adjustment components located within the field of view in front of the observed object group. All the exhibit display adjustment components perform rotation actions synchronously according to the coordinated rotation direction parameter sequence and the coordinated rotation angle parameter sequence, so that the display surfaces of the multiple exhibits collectively face the area of the line connecting the center positions of the observed object sub-groups.

[0153] The coordinated control commands are sent to the display and adjustment components of each exhibit via the network. Upon receiving the commands, each component synchronously starts its servo motor at the same time reference, performing rotational movements in the specified direction and angle.

[0154] Step S616: During the preset group aggregation observation period after the exhibit display adjustment component performs synchronous rotation, continuously acquire the local visual data stream of the spatial area where the observed object group is located, perform group aggregation state detection processing on the local visual data stream, extract the line-of-sight vector parameters of the individual observed objects in the group, and calculate the spatial discrete angle value of the group line-of-sight after the synchronous rotation is performed.

[0155] Let the observation period T_obs = 10 seconds. During this period, the visual data stream of the area where the group is located is continuously acquired. In each frame of the image, the gaze direction vector of each member in the group is extracted, and a new spatial discrete angle value θ_avg_new is calculated.

[0156] Step S617: Calculate the difference between the spatial discrete angle value before the synchronous rotation action and the group line-of-sight spatial discrete angle value after the synchronous rotation action, generate group aggregation intervention effect evaluation parameters, and store the group aggregation intervention effect evaluation parameters and the corresponding set of collaborative control instructions for display devices into the group intervention strategy experience database for subsequent parameter optimization iteration of group splitting intervention strategies.

[0157] Calculate the intervention effectiveness evaluation parameter Δθ = θ_avg_old - θ_avg_new. If Δθ > 0, it indicates that the intervention is effective and the group aggregation degree has increased. Store (Δθ, coordinating control instructions) in an experience database. Subsequently, the rotation angle parameter in the coordinating control instructions can be optimized based on the stored experience using reinforcement learning algorithms (such as Q-learning).

[0158] Step S710: Extract the global position coordinate sequence of all observed objects in the display scene from the global field of view visual data frame unit in the initial visual data stream set. Calculate the instantaneous movement speed value and instantaneous movement direction angle value of each observed object based on the position change of each observed object at continuous acquisition time in the global position coordinate sequence. Input the instantaneous movement speed value and instantaneous movement direction angle value into the pre-constructed observed object motion intention prediction network. The observed object motion intention prediction network extracts the motion pattern latent features in the observed object movement speed value sequence and movement direction angle value sequence through a long short-term memory processing layer. Output the predicted identifier of the next moving target area of ​​the observed object based on the motion pattern latent features through a fully connected classification processing layer.

[0159] From the global field-of-view visual data frame unit, the pixel coordinates (u, v) of each observed object in the image are obtained through an object detection algorithm. The pixel coordinates are converted to 3D position coordinates (X, Y, Z) in the world coordinate system according to camera calibration parameters. For each observed object, its position sequence P1, P2, ..., PT at consecutive times T=10 is extracted. The instantaneous velocity V_t=|P_{t+1}-P_t| / Δt and the instantaneous direction angle θ_t=atan2(P_{t+1, y}-P_{t, y}, P_{t+1, x}-P_{t, x}) are calculated. The velocity sequence and direction angle sequence are input into a Long Short-Term Memory (LSTM) network with a hidden layer dimension of 128. The output of the LSTM passes through a fully connected layer (128×K, where K is the number of target region categories) and a Softmax activation function, outputting the predicted probability of each target region. The region identifier corresponding to the highest probability is taken as the prediction result Zid.

[0160] Step S711: Determine the target display sub-area that the observed object will arrive at within a preset time period based on the next moving target area prediction identifier, and extract the exhibit identification information and the coordinates of the center point of the exhibit display area of ​​all exhibits in the target display sub-area from the preset three-dimensional space model of the display scene.

[0161] Based on the predicted region identifier Zid, the corresponding target display sub-region is queried from the preset region identifier and sub-region boundary mapping table. The exhibit identifier Id_i and the coordinates E_i of the center point of the exhibit display area for all exhibits within the sub-region are read from the 3D spatial model of the display scene.

[0162] Step S712: For each exhibit within the target display sub-area, extract the prior information of the exhibit's visual salience from the preset exhibit information database. The prior information of the exhibit's visual salience includes the exhibit's size area parameters, the exhibit's surface brightness parameters, and the exhibit's color difference parameters from the background. Calculate the exhibit's initial attractiveness score based on the prior information of the exhibit's visual salience.

[0163] Exhibit size area parameter As = exhibit bounding box width × exhibit bounding box height. Exhibit surface brightness parameter Lb = average grayscale value of pixels within the exhibit area. Exhibit and background color difference parameter Cd = sqrt((Rexhibit - Rbackground)) Exhibits - G Background Exhibit - B Background Initial attraction rating , where w1=0.3, w2=0.4, w3=0.3.

[0164] Step S713: Obtain the displacement trajectory of the group attention point of other observation object groups currently staying in the target display sub-area, calculate the coordinates of the gaze landing point position of the group attention point displacement trajectory at the current moment, count the duration value and cumulative number of observations of each exhibit as the current gaze landing point in the target display sub-area, convert the duration value and cumulative number of observations into the social conformity attraction increment parameter of the exhibit, and perform weighted superposition processing with the initial attraction score value of the exhibit to generate the comprehensive attraction prediction score value of the exhibit. The weight coefficient of the weighted superposition processing is dynamically adjusted by the motion intention confidence parameter output by the observation object motion intention prediction network.

[0165] For each exhibit within the target display sub-area, calculate its duration as a viewpoint (T_dwell) and the cumulative number of views (N_view). This is the incremental parameter for social conformity attraction. Where α=0.02, β=0.1. The motion intention confidence parameter C_conf (the maximum probability value output by Softmax) is obtained from the observed object motion intention prediction network. The overall attractiveness prediction score is then calculated. .

[0166] Step S714: Sort all exhibits in the target display sub-area in descending order according to the comprehensive attraction prediction score of the exhibits, and select the top N exhibits in terms of comprehensive attraction prediction score as the pre-activated exhibit set, where N is an integer predetermined based on the area of ​​the target display sub-area and the average size of the exhibits. The exhibits in the pre-activated exhibit set will enter the display pre-activated state before the observation object arrives at the target display sub-area.

[0167] Sort all exhibits within the target display sub-region in descending order by S_final. Let N=3, then select the top 3 exhibits as the pre-activated exhibit set S_pre.

[0168] Step S715: For each pre-activated exhibit in the pre-activated exhibit set, generate a pre-activation control command containing a pre-pointing direction parameter and a pre-pointing angle parameter of the exhibit display adjustment component. The pre-pointing direction parameter is determined by the direction vector of the line connecting the center point coordinates of the exhibit display area of ​​the pre-activated exhibit and the predicted entrance coordinates of the target display sub-area that the observed object is about to enter. The pre-pointing angle parameter is determined by the angle between the normal direction of the exhibit display surface and the direction vector of the line connecting the two points.

[0169] Obtain the predicted entry point coordinates P_entry of the observed object as it is about to enter the target display sub-area (this can be obtained by extrapolating the motion trajectory predicted in step S710). For each exhibit in the pre-activated exhibit set, its center coordinates are E. Calculate the direction vector D_pre = E - P_entry. The normal direction of the exhibit's current display surface is N_current, and the included angle θ_pre = arccos((N_current·D_pre) / (|N_current||D_pre|)). Generate a pre-activation control command, which includes the direction vector D_pre and the included angle θ_pre.

[0170] Step S716: The pre-activation control command is sent to the exhibit display adjustment component of the corresponding pre-activated exhibit by a preset advance time before the observed object arrives at the target display sub-area. The exhibit display adjustment component performs a pre-rotation action according to the pre-pointing direction parameter and the pre-pointing angle parameter, so that the display surface of the pre-activated exhibit is exactly facing the predicted entrance position direction of the observed object when the observed object enters the target display sub-area.

[0171] Let the advance time T_adv = 3 seconds. Based on the observed object's movement speed V_obj predicted in step S710, calculate its estimated arrival time T_arrival in the target display sub-area. At the time T_arrival - T_adv, send the pre-activation control command to the exhibit display adjustment component. The component performs a pre-rotation, orienting the exhibit display surface toward the P_entry direction.

[0172] Step S717: During the preset attraction effect observation period after the observed object actually enters the target display sub-area, acquire the local visual data stream transmitted by the directional visual acquisition device deployed in the target display sub-area, perform the first gaze point detection processing on the local visual data stream, extract the coordinates of the first gaze point within the preset time after the observed object enters the target display sub-area, and determine whether the coordinates of the first gaze point fall within the display area of ​​a certain pre-activated exhibit in the pre-activated exhibit set. Record the judgment result as the pre-activated exhibit attraction verification data, and feed the pre-activated exhibit attraction verification data back to the observed object motion intention prediction network to optimize the network weight parameters of the fully connected classification processing layer of the motion intention prediction network.

[0173] Within T_obs=2 seconds after the observed object enters the target display sub-area, the coordinates F_gaze of its first gaze point are extracted from the local visual data stream. It is determined whether F_gaze is located within the display area of ​​an exhibit in the pre-activated exhibit set. If so, positive feedback reward=1 is recorded; otherwise, negative feedback reward=0 is recorded. (reward, observed object motion feature sequence) is used as training samples and fed into the observed object motion intent prediction network. The policy gradient algorithm REINFORCE is used to update the weight parameters of the fully connected classification layer of the network.

[0174] In one exemplary embodiment, an intelligent display control system based on AI visual analysis is provided. This intelligent display control system can be a terminal, a server, etc., and its internal structure diagram can be as follows: Figure 2 As shown, this AI-based intelligent display control system includes a processor, memory, input / output interfaces, a communication interface, a display unit, and input devices. The processor, memory, and input / output interfaces are connected via a system bus, and the communication interface, display unit, and input devices are also connected to the system bus via the input / output interfaces. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides the environment for the operation of the operating system and computer programs in the non-volatile storage media. The input / output interfaces are used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, near-field communication, or other technologies. When the computer program is executed by the processor, it implements an AI-based intelligent display control method. The display unit is used to form a visually visible image and can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device can be a touch layer covering the display screen, or a button, trackball, or touchpad set on the shell of an intelligent display control system based on AI visual analysis, or an external keyboard, touchpad, or mouse, etc.

[0175] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. An intelligent display control method based on AI visual analysis, characterized in that, The method includes: Acquire an initial set of visual data streams transmitted by multiple visual acquisition devices deployed in the display area, wherein the initial set of visual data streams contains multiple visual data frame units with acquisition time sequence marks acquired continuously. The initial visual data stream set is subjected to a display scene deconstruction process to obtain the temporal observation tendency characteristics of the observed objects in the display scene and the spatial layout constraint characteristics of the display scene. The temporal observation tendency characteristics describe the displacement trajectory of the focus point of the observed object in a continuous visual data frame unit, and the spatial layout constraint characteristics describe the relative relationship between the display position of the exhibits and the activity area of ​​the observed object in the display scene. The pre-built visual attention analysis network is invoked to perform joint correlation analysis on the temporal observation tendency features and the spatial layout constraint features, generating a description of the distribution of hot and cold areas of visual attention in the current display scene; Based on the description of the distribution of hot and cold areas of visual attention, determine the exhibit identification information of the exhibits to be adjusted and the position migration vector sequence of the exhibit attention in the display scene space. Based on the exhibit identification information and the position migration vector sequence, a set of display device control instructions containing exhibit identification codes and angle adjustment parameters is generated, and the set of display device control instructions is distributed to the corresponding exhibit display adjustment components.

2. The intelligent display control method based on AI visual analysis according to claim 1, characterized in that, The acquisition of the initial visual data stream set transmitted by multiple visual acquisition devices deployed in the display area includes: The system receives a global field-of-view visual data stream transmitted by a global coverage visual acquisition device deployed at the top of the display area, and a directional field-of-view visual data stream transmitted by a directional visual acquisition device deployed directly in front of the independent exhibit display area. The global field-of-view visual data stream is composed of global field-of-view visual data frame units arranged in order according to the global acquisition timing mark, and the directional field-of-view visual data stream is composed of directional field-of-view visual data frame units arranged in order according to the directional acquisition timing mark. Based on the same global acquisition timing mark and directional acquisition timing mark, establish a correspondence sequence between global visual data frame units and directional visual data frame units. Then, perform frame-level combination processing on the global field-of-view visual data frame units and directional field-of-view visual data frame units after establishing the correspondence sequence according to the time correspondence to generate a composite visual data frame unit containing global field-of-view content and directional field-of-view content at each acquisition time. Perform observation object group detection processing on the global field of view content in the composite visual data frame unit to obtain the group distribution position contour description data of the observation object group in the global field of view content; and perform observation object single-unit detection processing on the directional field of view content in the composite visual data frame unit to obtain the single-unit pose direction description data of the observation object single unit in the directional field of view content. The group distribution location contour description data and the individual attitude direction description data at the same acquisition time are combined into the spatiotemporal state description record of the observed object, and the spatiotemporal state description record of the observed object is written into the corresponding composite visual data frame unit. All composite visual data frame units containing spatiotemporal state descriptions of the observed object at all consecutive acquisition times are arranged into a frame sequence according to the global acquisition timing mark. An initial visual data stream set containing continuous visual data frame units, with each visual data frame unit accompanied by a spatiotemporal state description of the observed object, is generated. The global acquisition timing mark of each visual data frame unit in the initial visual data stream set is extracted, and the global acquisition timing mark is converted into an acquisition timing mark value with a unified time reference. The acquisition timing mark value with the unified time reference is used as the unique acquisition timing mark of the visual data frame unit and saved. The initial visual data stream set after conversion to a unified time base is divided into storage areas. The initial visual data stream set is divided into multiple visual data stream segments according to the time segments of continuous acquisition, and the visual data stream segments are associated with the corresponding unified time base acquisition time sequence mark range.

3. The intelligent display control method based on AI visual analysis according to claim 1, characterized in that, The process of performing a deconstruction of the display scene on the initial visual data stream set yields the temporal observation tendency characteristics of the observed objects within the display scene and the spatial layout constraint characteristics of the display scene, including: From each visual data frame unit of the initial visual data stream set, extract the spatiotemporal state description record of the observed object. The spatiotemporal state description record of the observed object includes the group distribution position contour description data and the individual pose direction description data of the visual data frame unit. Based on the head orientation angle parameter and line-of-sight vector parameter of the observed object in the single-object posture direction description data, a three-dimensional line-of-sight ray representation of the observed object at each acquisition time is constructed. The three-dimensional line-of-sight ray representation includes the ray origin position information and the ray unit direction vector information. The three-dimensional line-of-sight ray representation is spatially intersected with the preset three-dimensional space model of the display scene to obtain the coordinates of the line-of-sight landing point of the observed object in the three-dimensional space model of the display scene. The coordinates of the line-of-sight landing point include the horizontal axis coordinate value, the horizontal axis coordinate value, and the vertical height coordinate value. The coordinates of the line-of-sight points of the same observation object at multiple consecutive acquisition times are collected. The coordinates of the line-of-sight points are concatenated in the order of acquisition times to form a line-of-sight point coordinate sequence. The line-of-sight point coordinate sequence is used as the individual attention point displacement trajectory of the observation object. The spatial position dispersion and motion direction consistency parameters of the individual attention point displacement trajectories of all observation objects in the same observation object group are calculated. Based on the spatial position dispersion and motion direction consistency parameters, the individual attention point displacement trajectories are merged into the group attention point displacement trajectory. The displacement direction vectors of the continuously collected moments in the displacement trajectory of the group of focus points are analyzed for directional gradient changes. The trajectory inflection point time and trajectory inflection point coordinates when the displacement direction change angle exceeds a preset angle threshold are extracted. Focus point displacement trajectory morphology description data containing trajectory inflection point time markers and trajectory inflection point coordinate markers are generated. The focus point displacement trajectory morphology description data is used as the temporal observation tendency characteristics of the observation object. Extract the region boundary coordinate sequence and region center point coordinate of all exhibit display areas from the preset 3D space model of the display scene, generate an inaccessible spatial constraint range description for each exhibit display area based on the region boundary coordinate sequence, and extract the region boundary coordinate sequence of the observation object's activity area from the preset 3D space model of the display scene, generate a description of the observation object's freely movable spatial range based on the region boundary coordinate sequence of the observation object's activity area. The spatial relationship of the description of the inaccessible spatial constraint range and the description of the freely movable spatial range of the observed object is superimposed and analyzed. The shortest spatial distance between the boundary of the exhibit display area and the boundary of the activity area of ​​the observed object is calculated, and the azimuth angle of the center point of the exhibit display area relative to the boundary of the activity area of ​​the observed object is calculated. The shortest spatial distance and the azimuth angle are combined to form the spatial relative relationship description parameters between the exhibit display position and the activity area of ​​the observed object. The set of spatial relative relationship description parameters corresponding to all exhibit display areas is used as the spatial layout constraint feature of the display scene.

4. The intelligent display control method based on AI visual analysis according to claim 1, characterized in that, The process involves invoking a pre-built visual attention analysis network to perform joint correlation analysis on the temporal observation tendency features and the spatial layout constraint features, generating a description of the distribution of hot and cold areas of visual attention in the current display scene, including: The data describing the displacement trajectory morphology of the attention point in the time-series observation tendency features are input into the trajectory feature encoding branch of the visual attention analysis network. The trajectory feature encoding branch consists of multiple cascaded trajectory convolution processing layers and trajectory pooling processing layers. The trajectory convolution processing layer extracts the trajectory curvature change features and trajectory pause duration distribution features of the displacement trajectory of the attention point. The trajectory pooling processing layer compresses the redundant information of the time dimension of the trajectory features to generate the trajectory feature encoding vector. The set of spatial relative relationship description parameters in the spatial layout constraint features is input into the spatial feature encoding branch of the visual attention analysis network. The spatial feature encoding branch consists of multiple cascaded spatial graph convolutional processing layers. Each spatial graph convolutional processing layer takes the exhibit display area as a graph node and constructs node connection edges according to the shortest spatial distance value in the spatial relative relationship description parameters. The spatial structural association features between nodes in the exhibit display area are extracted through the spatial graph convolutional processing layer to generate a spatial feature encoding vector. The trajectory feature encoding vector output by the trajectory feature encoding branch and the spatial feature encoding vector output by the spatial feature encoding branch are simultaneously input into the joint correlation analysis layer of the visual attention analysis network. The correlation response strength value between each position dimension in the trajectory feature encoding vector and each exhibit node dimension in the spatial feature encoding vector is calculated through a multi-head cross-attention processing mechanism. The associated response intensity value is weighted and fused with the preset prior information on the visual saliency of the exhibits. The prior information on the visual saliency of the exhibits includes the size and area parameters of the exhibits in the exhibit display area, the surface brightness parameters of the exhibits, and the color difference parameters between the exhibits and the background. The weighted fusion process generates a corrected associated response intensity value of the fused prior information and inputs it into the attention heat distribution generation layer of the visual attention analysis network. Based on the corrected associated response intensity value, the three-dimensional spatial model of the display scene is subjected to voxel-by-voxel attention assignment processing to generate the attention heat value of each spatial voxel unit in the three-dimensional spatial model of the display scene. Spatial clustering is performed on the attention heat values ​​of all spatial voxel units in the three-dimensional spatial model of the display scene. Spatial voxel units with attention heat values ​​exceeding a preset heat threshold and adjacent spatial positions are aggregated into attention hot spots, and spatial voxel units with attention heat values ​​below a preset cold spot threshold and connected spatial positions are aggregated into attention cold spot areas. Extract the region boundary voxel coordinate sequence of the hot spot region and calculate the average heat value of the hot spot region. Write the region boundary voxel coordinate sequence and the average heat value of the hot spot region as hot spot region description parameters into the hot spot region recording unit. Extract the voxel coordinate sequence of the region boundary of the cold attention area and calculate the average attention heat value inside the region. Write the region boundary voxel coordinate sequence and the average attention heat value inside the region as cold attention area description parameters into the cold attention area recording unit. All hot spot and cold spot recording units are deduplicated in spatial location and their boundaries are smoothed to obtain non-overlapping sets of hot spot and cold spot regions. Then, a unified data structure is used for encapsulation to generate a description of the distribution of hot and cold spot regions, which includes the description parameters of hot spot and cold spot regions and the visual attention.

5. The intelligent display control method based on AI visual analysis according to claim 1, characterized in that, The step of determining the exhibit identification information and the position migration vector sequence of exhibit attention in the display scene space based on the description results of the distribution of hot and cold areas of visual attention includes: The hot spot region description parameter set in the cold and hot spot region distribution description results of the visual attention is analyzed, and the region boundary voxel coordinate sequence and the average attention heat value inside the region are extracted from each hot spot region description parameter. The region geometric center spatial coordinates of each hot spot region are calculated based on the region boundary voxel coordinate sequence. The cold spot region description parameter set in the cold spot region distribution description result of the visual attention is analyzed, the region boundary voxel coordinate sequence and the average attention heat value inside the region are extracted from each cold spot region description parameter, and the region geometric center spatial coordinates of each cold spot region are calculated based on the region boundary voxel coordinate sequence. Spatial distance matching calculation is performed between the spatial coordinates of the geometric center of each hot spot area and the coordinates of the center point of the preset exhibit display area. Exhibit display areas with spatial distance values ​​less than a preset distance matching threshold are selected as candidate hot spot associated exhibits, and the exhibit identification information of the candidate hot spot associated exhibits is recorded as the exhibit identification information of the exhibit to be adjusted. For each hotspot-related exhibit, the spatial coordinates of the geometric center of the hotspot area are extracted at multiple consecutive acquisition times. The spatial coordinates of the geometric center of the hotspot area are concatenated according to the order of acquisition times to form a time series of hotspot area center coordinates. The movement direction trend extraction processing is performed on the time series of hotspot area center coordinates. The displacement direction unit vector and displacement distance scalar value between the spatial coordinates of the hotspot area center at adjacent acquisition times are calculated to generate a hotspot area movement direction vector sequence containing the displacement direction unit vector and displacement distance scalar value. Based on the hot spot area movement direction vector sequence, predict the center coordinates of the hot spot area within a preset time step in the future, connect the spatial coordinates of the center of the hot spot area at the current acquisition time with the predicted coordinates of the center of the hot spot area within a preset time step in the future, and generate the location migration vector of the hot spot area attention. For each cold point associated exhibit, the spatial coordinates of the geometric center of the cold point region are extracted at multiple consecutive acquisition times. The spatial coordinates of the geometric center of the region are concatenated in the order of acquisition times to form a time series of the center coordinates of the cold point region. The time series of the center coordinates of the cold point region is processed to extract the movement direction trend. The unit vector of displacement direction and the scalar value of displacement distance between the spatial coordinates of the center of the cold point region at adjacent acquisition times are calculated to generate a cold point region movement direction vector sequence containing the unit vector of displacement direction and the scalar value of displacement distance. Based on the predicted coordinates of the center position of the cold spot area within a preset time step according to the vector sequence of the moving direction of the cold spot area, the spatial coordinates of the center position of the cold spot area at the current acquisition time are connected with the predicted coordinates of the center position of the cold spot area within the preset time step to generate the position migration vector of the cold spot area attention. The position migration vector of the hot spot area attention and the position migration vector of the cold spot area attention are combined to form the position migration vector sequence of the exhibit attention in the display scene space.

6. The intelligent display control method based on AI visual analysis according to claim 1, characterized in that, The process of generating a set of display device control commands containing display device identification codes and angle adjustment parameters based on the exhibit identification information and the position migration vector sequence, and distributing the set of display device control commands to the corresponding exhibit display adjustment components, includes: Extract the unique exhibit identifier code of each exhibit to be adjusted from the exhibit identifier information, query the preset exhibit display adjustment component mapping table according to the unique exhibit identifier code, and obtain the network address of the display adjustment component and the set of adjustable parameter types of the display adjustment component bound to the unique exhibit identifier code; From the position migration vector sequence, extract the position migration vector corresponding to the exhibit identification information of each exhibit to be adjusted. The position migration vector includes the spatial coordinates of the migration start point and the spatial coordinates of the migration end point. Calculate the direction vector of the position migration vector based on the spatial coordinates of the migration start point and the migration end point. Determine the target horizontal orientation angle of the horizontal rotation adjustment component in the exhibit display adjustment assembly based on the direction vector of the position migration vector. Calculate the difference between the target horizontal orientation angle and the current horizontal orientation angle of the exhibit display adjustment assembly, using it as a horizontal rotation angle parameter, and determine the corresponding horizontal rotation direction parameter. The horizontal rotation direction parameter is determined by the positive or negative sign value of this difference. Determine the target vertical pitch angle of the vertical pitch adjustment component in the exhibit display adjustment assembly based on the direction vector of the position migration vector. Calculate the difference between the target vertical pitch angle and the current vertical pitch angle of the exhibit display adjustment assembly, using it as a vertical pitch angle parameter, and determine the corresponding vertical pitch direction parameter. The vertical pitch direction parameter is determined by the positive or negative sign value of this difference. Extract the vertical height coordinates of the starting point and ending point from the location migration vector. Calculate the vertical height difference between the ending point and the starting point. Use the absolute value of the vertical height difference as the height rise / fall distance parameter. Determine the height rise / fall direction parameter based on the sign of the vertical height difference. Encapsulate the horizontal rotation direction parameter, horizontal rotation angle parameter, vertical pitch direction parameter, vertical pitch angle parameter, height rise / fall direction parameter, and height rise / fall distance parameter with the corresponding unique exhibit identifier to generate an independent control instruction data packet for each exhibit to be adjusted. Write the control parameter values ​​in the independent control instruction data package into the instruction parameter field that matches the set of controllable parameter types of the display control component, and add instruction execution timing mark and instruction priority mark to form a set of display device control instructions containing exhibit identification code and spatial coordinate offset. A bidirectional data transmission channel is established between the display adjustment component and the exhibit display adjustment component based on the network address of the display adjustment component. Through the bidirectional data transmission channel, the independent control instruction data packets corresponding to the control instruction set of the display device are sent to the instruction receiving buffer of the exhibit display adjustment component. The exhibit display adjustment component reads the independent control instruction data packets from the instruction receiving buffer and parses the control parameter values. It then drives the horizontal rotation adjustment component to perform horizontal rotation, the vertical pitch adjustment component to perform vertical pitch, and the height adjustment component to perform height adjustment, so as to adjust the orientation of the exhibit to be consistent with the direction of the attention position migration vector.

7. The intelligent display control method based on AI visual analysis according to claim 1, characterized in that, The method further includes: After the preset adjustment stabilization period ends following the execution of the adjustment action by the exhibit display adjustment component, the set of initial visual data streams after adjustment transmitted by multiple visual acquisition devices deployed in the display area is reacquired. The set of initial visual data streams after adjustment includes multiple adjusted visual data frame units with adjustment acquisition time sequence marks that are continuously acquired. The adjusted initial visual data stream set is subjected to adjusted display scene deconstruction processing to obtain adjusted temporal observation tendency features of the observed objects in the adjusted display scene and spatial layout constraint features of the adjusted display scene. The adjusted temporal observation tendency features describe the adjusted attention point displacement trajectory of the observed objects in continuous adjusted visual data frame units. The spatial layout constraint features describe the relative relationship between the display position of exhibits and the activity area of ​​the observed objects in the display scene. A pre-built visual attention analysis network is invoked to perform joint correlation analysis on the adjusted temporal observation tendency characteristics and the spatial layout constraint characteristics, generating an adjusted description of the distribution of hot and cold areas of visual attention in the current display scene. Based on the adjusted description of the distribution of hot and cold areas, evaluation parameters for the degree of improvement of visual attention distribution are calculated. The evaluation parameters for the degree of improvement of visual attention distribution include the cold area area reduction ratio, the hot area attention heat value increase ratio, and the hot area overlap with the exhibit display area increase ratio. The visual attention distribution improvement assessment parameters are input into the preset display effect feedback adjustment strategy module. The display effect feedback adjustment strategy module iteratively corrects the angle adjustment parameters in the display device control command set based on the cold spot area reduction ratio, the hot spot area attention heat value increase ratio, and the hot spot area overlap with the exhibit display area increase ratio. Based on the iteratively corrected angle adjustment parameters, the horizontal rotation angle parameters, vertical pitch angle parameters, and height rise and fall distance parameters of the exhibit display adjustment component are regenerated, and the display device control command set after feedback correction is generated. The set of control commands for the display device that has been corrected by feedback is redistributed to the corresponding exhibit display adjustment components. The exhibit display adjustment components perform secondary adjustment actions based on the corrected horizontal rotation angle parameters, vertical pitch angle parameters, and height rise and fall distance parameters. Record the distribution of hot and cold areas after each round of adjustment and the evaluation parameters of the improvement of visual attention distribution, forming a historical evaluation parameter record sequence of the exhibit display adjustment effect. Input the historical evaluation parameter record sequence into the preset exhibit display layout optimization analysis module. The exhibit display layout optimization analysis module generates static display position adjustment suggestions based on the changing trend of hot and cold area distribution in the historical evaluation parameter record sequence. The proposed adjustment information for the static display position of the exhibits is sent to the display management terminal device. The proposed adjustment information includes the exhibit identification code and the coordinates of the center point of the proposed display area.

8. The intelligent display control method based on AI visual analysis according to claim 1, characterized in that, The method further includes: For each visual data frame unit in the initial visual data stream set, perform observation object attribute classification processing, extract facial region visual feature data, body contour visual feature data and clothing color visual feature data of the observation object, and generate observation object attribute feature vector based on facial region visual feature data, body contour visual feature data and clothing color visual feature data; The observation object attribute feature vectors at different acquisition times are input into the observation object re-identification and tracking network. The observation object re-identification and tracking network calculates the feature similarity distance value between the observation object attribute feature vectors at different acquisition times, and performs continuous identity binding processing for the same observation object across visual data frame units based on the feature similarity distance value. A mapping association table is established between the continuous identity of the observation object and the temporal observation tendency characteristics of the observation object in each visual data frame unit. The mapping association table records the one-to-one correspondence between the continuous identity of the observation object and the individual attention point displacement trajectory. Based on the individual attention point displacement trajectory corresponding to the continuous identity identifier of each observed object in the mapping association table, the sequence of exhibit display area identifiers passed by the coordinates of the eye landing point during the complete visit period is extracted, and the sequence of exhibit display area identifiers is arranged in the order of the eye landing point to form an individual visit path sequence. Calculate the path similarity parameter between individual visit path sequences of all observed objects in the display scene, and classify the individual visit path sequences into multiple visit flow categories based on the path similarity parameter; The number of observation objects included in each visitor flow category is counted, the proportion of the number of observation objects in each visitor flow category among all observation objects is calculated, and visitor flow categories whose proportion of the number of observation objects exceeds the preset mainstream flow proportion threshold are marked as mainstream visitor flow. The sequence of exhibit display area identifiers along the main visitor flow is spatially overlaid with the exhibit display locations in the current display scene to identify exhibit display area identifiers located at the beginning and end of the main visitor flow. Based on the exhibit display area identification located at the beginning of the main visitor flow, an exhibit display position adjustment strategy is generated. The exhibit display position adjustment strategy includes an exhibit position exchange scheme that adjusts exhibits with historical average attention heat values ​​higher than a preset high attention threshold from non-starting area to the starting area of ​​the main visitor flow. Based on the exhibit display area markers located at the end of the main visitor flow, a post-adjustment strategy for exhibit display positions is generated. The post-adjustment strategy for exhibit display positions includes an exhibit position exchange scheme that adjusts exhibits with historical average attention heat values ​​below a preset low attention threshold from the starting area to the end area of ​​the main visitor flow. The pre-adjustment strategy and post-adjustment strategy for exhibit display positions are packaged into a static layout optimization scheme data package, and the static layout optimization scheme data package is transmitted to the display management terminal device to assist in the decision-making on the adjustment of the physical display layout of exhibits.

9. The intelligent display control method based on AI visual analysis according to claim 1, characterized in that, The method further includes: Extract the region boundary voxel coordinate sequence from all cold point region description parameters from the description results of the distribution of cold and hot points of visual attention. Determine the display scene space range covered by the cold point region based on the region boundary voxel coordinate sequence. Perform spatial position overlap judgment processing on the display scene space range covered by the cold point region and the preset exhibit display area location information to filter out the set of cold point covered exhibit identifiers that have a spatial position overlap relationship with the cold point region. For each cold-spot covered exhibit in the cold-spot covered exhibit identifier set, extract the exhibit description text data, exhibit image data set, and exhibit multimedia presentation material data set of the cold-spot covered exhibit from the preset exhibit information database; The system calls a pre-built dynamic content generation network to convert the exhibit description text data into a speech synthesis parameter sequence, the exhibit image data set into a dynamic image carousel control command sequence, and the exhibit multimedia demonstration material data set into a demonstration content trigger signal sequence. The device identifier of the directional audio output device deployed near the cold spot area is determined based on the spatial coordinates of the cold spot area, and the speech synthesis parameter sequence is sent to the directional audio output device. The directional audio output device plays the audio content of the exhibit introduction according to the speech synthesis parameter sequence. The screen identifier of the auxiliary visual presentation screen deployed above the cold spot area is determined based on the spatial coordinates of the cold spot area. The dynamic image carousel control command sequence and the demonstration content trigger signal sequence are sent to the auxiliary visual presentation screen. The auxiliary visual presentation screen switches to display exhibit-related images according to the dynamic image carousel control command sequence and starts the playback process of exhibit multimedia demonstration materials according to the demonstration content trigger signal sequence. During the playback of exhibit introduction voice content by the directional audio output device and the display of exhibit-related images by the auxiliary visual presentation screen, the local visual data stream transmitted by the directional visual acquisition device deployed near the cold spot area is continuously acquired, and the observation object attention state detection processing is performed on the local visual data stream. The observation object attention state detection and processing includes extracting the head orientation angle change rate parameter and the gaze direction stabilization duration parameter of the observation object from the local visual data stream. Based on the head orientation angle change rate parameter and the gaze direction stabilization duration parameter, it is determined whether the observation object's attention is attracted to the cold spot covered exhibit area. When the observation object's attention is attracted to the cold spot covered exhibit area, a content intervention fade-out control command sequence is generated to reduce the playback volume of the directional audio output device and the display brightness of the auxiliary visual presentation screen according to a preset attenuation curve. The content intervention fade-out control command sequence is sent to the directional audio output device and the auxiliary visual presentation screen to perform the fade-out operation. After the content fade-out operation is completed, the cold spot covered exhibit area is updated to the exhibit area that has undergone attention enhancement processing, and the cold spot area description parameter record corresponding to the cold spot covered exhibit is removed from the cold spot area queue to be processed.

10. An intelligent display control system based on AI visual analysis, characterized in that, include: processor; A machine-readable storage medium for storing machine-executable instructions of the processor; The processor is configured to execute the intelligent display control method based on AI visual analysis as described in any one of claims 1 to 9 by executing the machine-executable instructions.

Images