Exhibition hall personalized narration adaptive pushing system based on eye movement tracking and gait pre-judgment

Abstract

The application relates to the field of intelligent exhibition hall interaction, and specifically discloses an exhibition hall personalized narration adaptive pushing system based on eye movement tracking and gait prediction, which comprises the following modules: a bimodal perception data fusion and tour state feature extraction module, which realizes cross-device identity consistency binding by constructing an eye movement scanning-gait turning cross-modal space-time coupling fingerprint to analyze real tour intention; a cognitive-movement conflict analysis and tour behavior prediction module, which identifies an inflection point time to predict a stay intention, defines a cognitive window period based on a cognitive intervention readiness index, calculates an implicit interest conduction probability to generate a conduction path; a personalized narration adaptive scheduling and closed-loop optimization module, which adjusts a playing time according to the inflection point time and the conflict index, reorganizes a narration logic according to the implicit interest conduction path, triggers a load reduction mode when cognitive fatigue occurs, and constructs a group implicit interest conduction network reverse iteration optimization model. The application realizes intelligent narration pushing with prediction, accuracy and individuality.

Description

Technical Field

[0001] This invention relates to the field of intelligent exhibition hall interaction technology, and in particular to an adaptive push system for personalized narratives in exhibition halls based on eye tracking and gait prediction. Background Technology

[0002] Currently, the content display and narrative delivery modes of various cultural and museum exhibition halls, science and technology exhibition halls, and corporate exhibition halls are generally homogenized and fixed. Most of them use preset loop playback, fixed voice explanations, or static graphic displays to provide services. They cannot be dynamically adjusted according to the visitor's real-time status, interests, preferences, and tour pace, which shows obvious technical shortcomings.

[0003] Existing intelligent exhibition hall interactive systems suffer from the following deep-seated technical defects: Cross-modal user identity consistency binding drift problem. In exhibition hall scenarios, eye-tracking cameras, gait acquisition cameras, and plantar pressure sensors are typically deployed in a dispersed manner. As users move between exhibition areas, existing systems lack an effective mechanism for continuous cross-device identity binding. Eye-tracking data and gait data are prone to identity mismatch, resulting in the "perceptual ghost" phenomenon, causing subsequent push logic to completely deviate from the true target user. Lack of conflict resolution between visual-driven and movement-driven intentions. In exhibition hall scenarios, users frequently exhibit a "see-walk separation" state, where their eyes are focused on exhibit A while their feet move towards exhibit B. Existing systems typically rely solely on visual perception or crudely weighted averages, lacking a scientific conflict resolution mechanism. This fails to accurately analyze the user's true browsing intentions, leading to misaligned content push targets and narrative logic breaks. Missed cognitive window and sudden changes in cognitive load. Traditional systems rely on passive triggering, only displaying content after a user has lingered in front of an exhibit for a fixed period. This completely ignores the brief cognitive window between initial interest and deep engagement. Pushing content too early interrupts user exploration, while pushing it too late misses the optimal cognitive intervention, causing a sudden shift in cognitive load and severely disrupting the continuity of the immersive narrative experience. Furthermore, existing systems fail to uncover implicit semantic connections between exhibits. Treating each exhibit as an independent information node, they lack the ability to mine implicit knowledge connections between exhibits and cannot predict the implicit transmission paths of deep user interests. This results in fragmented knowledge acquisition and difficulty in forming a systematic cognitive framework. The system also suffers from the implicit accumulation of cognitive fatigue and narrative overload. Existing systems cannot detect implicit cognitive fatigue through subtle changes in eye movement and gait. When users enter a period of cognitive fatigue after prolonged browsing, the system continues to push high-density, in-depth content, causing narrative overload and reducing content dissemination efficiency and user satisfaction. Finally, the system lacks dynamic evolution across user interest networks. Traditional systems collect fragmented behavioral data, recording only the dwell time and browsing trajectory of a single user. They fail to perform correlation mining on multi-user behavioral data to discover implicit semantic connections between exhibits, resulting in a static and rigid exhibit association network in the exhibition hall narrative resource library, which cannot be dynamically optimized as user group behavior patterns evolve.

[0004] Currently, there is no exhibition hall narrative adaptive push system that integrates eye-tracking for precise visual perception and gait dynamic prediction. There is also no complete technical system that can solve deep technical problems such as cross-modal identity binding drift, visual-mobile intention conflict resolution, cognitive window prediction, and latent interest transmission path mining through joint analysis of dual-dimensional human behavioral characteristics. The industry urgently needs a high-precision, predictive, and personalized intelligent exhibition hall narrative interaction system. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a personalized narrative adaptive push system for exhibition halls based on eye tracking and gait prediction.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] The personalized narrative adaptive push system for exhibition halls based on eye tracking and gait prediction includes the following specific steps:

[0008] S1. Dual-modal perception data fusion and tour state feature extraction module:

[0009] This is used to simultaneously collect user eye-tracking data and gait behavior data. The eye-tracking data includes the user's gaze point spatial coordinates, gaze duration, saccade trajectory, pupil diameter change, and blink frequency at each time point. The gait behavior data includes the user's instantaneous walking speed, cadence, walking direction angle, stride length, resting posture, turning trajectory, and start / stop status at each time point.

[0010] The preprocessing and spatiotemporal alignment of the eye-tracking data and the gait behavior data includes using filtering algorithms to remove environmental noise and data jitter, eliminating invalid data frames, and unifying timestamps.

[0011] This is used to construct a cross-modal spatiotemporal coupled fingerprint of eye-tracking saccades and gait turning, and to generate a user cross-device consistent identity identifier by analyzing the time-delay coupling relationship between the saccade trajectory and the turning trajectory.

[0012] Used to construct a visual-mobile intent conflict resolution matrix, when the visual attention area where the eye gaze point is located is inconsistent with the mobile target area pointed to by the direction of travel, the visual dominant intent weight and mobile dominant intent weight are calculated based on historical behavior patterns and current tour status, and the conflict resolution coefficient is output.

[0013] This is used to perform weighted fusion of eye-tracking focus features, visual interest region features, gait rhythm features, and walking trend features, and combine them with the visual dominant intent weight, the mobile dominant intent weight, and the conflict resolution coefficient to generate a user consistency binding tour state feature vector.

[0014] S2. Cognitive-Motor Conflict Analysis and Tourist Behavior Prediction Module:

[0015] This is used to receive the user's consistent bound tour state feature vector and input it into a lightweight deep learning prediction model. Based on the continuous time series data of the gait rhythm features and walking trend features, as well as the offset data of the eye movement focus features and visual interest area features, the model predicts the user's walking position, the probability of focusing on the target exhibit, and the expected dwell time within a preset time period.

[0016] This method is used to perform antagonistic relationship analysis on eye movement and gait time series data within each sliding time window, based on spatiotemporally aligned eye movement and gait time series data. The antagonistic relationship analysis includes: extracting the offset vector of the spatial coordinates of the fixation point relative to the user's head position within the current time window; calculating the time series standard deviation of the offset vector as a visual lock-on strength index, which characterizes the degree of stability of the user's visual attention system in locking onto external exhibits within the current time period; simultaneously extracting the time series coefficient of variation of the instantaneous walking speed within the current time window as a gait propulsion strength index, which characterizes the driving strength of the user's motor control system in maintaining forward movement within the current time period; and calculating the ratio of the visual lock-on strength index to the gait propulsion strength index within the same time window as a visual-motor conflict index. When the visual lock-on strength index increases and the gait propulsion strength index increases synchronously, it is determined to be a high-conflict state, indicating that the user's visual attention system and motor control system are in a mutually antagonistic game process.

[0017] The visual-motion conflict index is used to identify the inflection point when it changes from an upward trend to a downward trend based on the change trajectory of the visual-motion conflict index based on multiple consecutive sliding time windows. The inflection point is determined as the burst point of the user's intention to stay. Based on the time difference between the occurrence time of the inflection point and the current time, the probability and duration of the user's stay on the current exhibit within a preset time period in the future are predicted.

[0018] This is used to construct a cognitive intervention readiness index based on the rate of change of pupil dilation frequency and gait deceleration characteristics, and to predict the time window for a user to enter a deep cognitive intervention state from the current state.

[0019] Based on the semantic association network of exhibits in the exhibition hall, this method calculates the implicit interest transmission probability of the target exhibit's attention probability among associated exhibits, generates a target exhibit prediction sequence including direct target exhibits and implicitly associated exhibits, and constructs an implicit interest transmission path.

[0020] This tool combines gaze duration, pupil changes, blink frequency, posture, and walking speed to quantify user interest in exhibits into three levels: high interest, medium interest, and low interest.

[0021] This is used to construct a dynamic three-dimensional user profile based on the user's interest preferences for each exhibit, tour pace, content acceptance habits, and the cognitive engagement readiness index, including preference tags, pace tags, cognitive level tags, and cognitive fatigue index.

[0022] Used to generate behavior prediction results, the behavior prediction results include target exhibit identifier, dwell probability value, predicted dwell time, visual-motor conflict index, target exhibit prediction sequence, implicit interest transmission path, interest level and dynamic three-dimensional user profile;

[0023] S3. Personalized Narrative Adaptive Scheduling and Closed-Loop Optimization Module:

[0024] Used to receive the behavior prediction results and match the corresponding narrative content from the pre-generated multi-level narrative resource library. The multi-level narrative resource library reserves four types of narrative content for each exhibit: basic popular science version, in-depth detailed explanation version, interesting story version and professional academic version.

[0025] When the probability of stopping exceeds a preset threshold, the preloading and preplay of the narrative content corresponding to the current exhibit is initiated within a first preset time window after the inflection point, and the narrative content of implicitly related exhibits is preloaded within the pre-visit interval before the user arrives at the direct target exhibit according to the implicit interest transmission path.

[0026] It is used to match narrative content of corresponding length and depth from the multi-level narrative resource library according to the predicted dwell time and the interest level, and dynamically adjust the start playback time and content entry rhythm of the pushed content according to the real-time value of the visual-motion conflict index.

[0027] The system is used to dynamically adjust the playback speed, switching speed and display duration of the narrative content based on the user's gait rhythm, eye movement focus and the cognitive fatigue index. When the cognitive fatigue index exceeds a preset threshold, a cognitive deload mode is triggered to automatically simplify the depth of the narrative content and shorten the display duration.

[0028] It is used to dynamically adjust the narrative order and content association logic of exhibits based on the user's browsing trajectory and interest shift trend, and automatically connect the related knowledge points of adjacent exhibits according to the implicit interest transmission path to form a coherent personalized tour narrative chain.

[0029] It is used to output adapted narrative content to various terminal devices for display and to collect user content browsing feedback data in real time.

[0030] This is used to construct a group implicit interest transmission network, iteratively optimize the edge weights of the semantic association network of the exhibition hall exhibits based on multi-user browsing feedback data, and iteratively optimize the lightweight deep learning prediction model and content matching strategy in reverse to form a closed-loop intelligent service system.

[0031] The beneficial effects of this invention are as follows:

[0032] 1. Construct a cross-modal spatiotemporal coupled fingerprint of eye-tracking saccades and gait turning to achieve consistent binding of user identity across devices and eliminate the phenomenon of perception ghosts; through the visual-movement intent conflict resolution matrix and visual-motion conflict index, accurately analyze the true intent in the "see-walk separation" state, and predict the burst point of dwell intent based on the inflection point moment to achieve content preloading and seamless push, solving the problems of traditional push delay and target misalignment.

[0033] 2. Based on pupil dilation and gait deceleration features, a cognitive intervention readiness index is constructed to accurately define the deep cognitive window period, so that the timing of push notifications is synchronized with the user's cognitive state; the probability of implicit interest transmission is calculated through semantic association network and transmission paths are generated to automatically connect implicitly related exhibits; a cognitive deload mode is triggered based on the cognitive fatigue index to avoid narrative overload and improve the efficiency of content dissemination and the tour experience.

[0034] 3. Construct a group implicit interest transmission network, iteratively optimize the edge weights of the semantic association network based on multi-user browsing feedback, and inject the updated network back into the prediction model so that the association relationship of exhibits evolves dynamically with the group behavior; form a closed-loop system of perception-prediction-push-feedback, realize the continuous evolution of system accuracy and adaptation effect, and have strong scene adaptability and scalability. Attached Figure Description

[0035] Figure 1 This is a framework diagram of the system proposed in this invention;

[0036] Figure 2 This is a flowchart of the system proposed in this invention;

[0037] Figure 3 This is an architecture diagram of the lightweight deep learning prediction model of this invention. Detailed Implementation

[0038] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0039] Please see the appendix Figure 1 -Appendix Figure 3 The personalized narrative adaptive push system for exhibition halls, based on eye tracking and gait prediction, includes the following specific steps:

[0040] S1. Dual-modal perception data fusion and tour state feature extraction module, used to simultaneously collect eye tracking and gait behavior data, construct a cross-modal spatiotemporal coupled fingerprint of eye saccades and gait turning to generate a consistent identity identifier, construct a visual-mobility intent conflict resolution matrix to output conflict resolution coefficients, and generate a consistent bound tour state feature vector; specifically including:

[0041] S11. Data Synchronization Acquisition Unit, the specific working steps are as follows:

[0042] Eye-tracking data is collected using an infrared eye-tracking camera array. The cameras are deployed 1.5m to 2.5m above the display case, with a sampling frequency of 60Hz to 120Hz, preferably 90Hz. The spatial coordinates of the gaze point are based on the exhibition hall floor coordinate system, and after calibration with intrinsic and extrinsic parameters, two-dimensional coordinates are output with an accuracy of 5mm. The duration of gaze is obtained by detecting valid gaze events with a gaze point offset of less than 1° of visual angle and a duration exceeding 100ms; the start and end times are recorded and the duration is calculated. The saccade trajectory tracks the rapid eye movement path between adjacent gaze events, with a sampling interval of 11ms to 17ms. The pupil diameter change is extracted through pupil edge detection under infrared light, with an accuracy of 0.1mm. The blink frequency is counted by detecting complete blinks where the eyelid covers more than 80% of the pupil area and lasts from 50ms to 400ms.

[0043] Gait data is collected through a ground-based plantar pressure sensor array and a top-mounted millimeter-wave radar, with a sampling frequency of 30Hz to 60Hz, preferably 50Hz. Instantaneous walking speed is calculated from the differential displacement of the plantar pressure center in consecutive frames, with a range of 0m / s to 2m / s and a resolution of 0.01m / s. Step frequency is calculated from the number of peak plantar pressure readings per unit time, ranging from 0 steps per minute to 180 steps per minute. The direction angle is calculated from the tangent of the pressure center trajectory or the motion vector of the radar point cloud, with a resolution of 1°, using the north of the exhibition hall as the 0-degree reference. Stride length is calculated from the displacement modulus of the pressure center between two adjacent steps, ranging from 0cm to 150cm. Stationary posture is determined by the combined variance of the plantar pressure center of gravity and the radar Doppler velocity; a stationary state is defined as when the variance of the center of gravity is below a threshold and the radar velocity is below 0.1m / s. Turning trajectory is obtained from the differential sequence of the direction angle at consecutive time points. The start / stop state is marked by the jump time when the speed increases from 0 m / s to above 0.3 m / s or decreases from above 0.3 m / s to 0 m / s, with a judgment window of 200 ms.

[0044] The eye-tracking camera and gait sensor are synchronized via the NTP protocol clock, with timestamp alignment accuracy within 1ms, ensuring that the eye-tracking and gait data form a spatiotemporally synchronized behavioral data pair at the same moment.

[0045] S12. Data preprocessing and spatiotemporal alignment unit, the specific working steps are as follows:

[0046] Kalman filtering is performed on the eye-tracking data. The state vector includes the gaze coordinates and velocity. The observation noise covariance is initialized to 0.01, and the process noise covariance is initialized to 0.001. Iterative updates remove noise from eye-tracking sampling jitter and head tremors. A finite impulse response low-pass filter is used on the gait behavior data. The order ranges from 16 to 32, with 24 being preferred. The cutoff frequency is 8Hz. This filter removes ground vibration and electromagnetic interference noise. Specifically: In the formula: For the filtered output sequence, the first... One sampling point; For the first First-order filter coefficients; For the input sequence, the first... One sampling point; This represents the filter order, ranging from 16 to 32, with 24 being preferred. Filter coefficients. The design adopts the Hamming window method to meet the requirements. Window length The cutoff frequency is 8 Hz, the passband ripple is 0.1 dB, and the stopband attenuation is 50 dB.

[0047] When removing invalid data frames, eye-tracking data frames are marked as invalid if the pupil recognition confidence is below 0.6 or the gaze point coordinates exceed the exhibition hall boundary; gait data frames are marked as invalid if the effective contact point ratio of the plantar pressure sensor array is below 30% or the millimeter-wave radar echo signal-to-noise ratio is below 10dB. Invalid frames are deleted directly.

[0048] Timestamp alignment is based on the eye-tracking system clock, and gait sensors are synchronized via the NTP protocol, achieving an alignment accuracy better than 1ms. Linear interpolation is used to map bimodal data to a unified time grid with a grid period of 10ms to 20ms, preferably 16.7ms, ensuring a strict correspondence between eye-tracking and gait states at the same physical moment. Specifically: In the formula: For a moment The interpolation result; For adjacent known data times; This corresponds to the known data value.

[0049] Eye focus features are extracted from aligned eye movement data, generated by weighted fusion of the mean fixation duration and pupil diameter change within a 1-3s sliding window, preferably 2s. Visual interest region features are generated by spatial clustering of fixation point coordinates using the DBSCAN density clustering algorithm, with a neighborhood radius ϵ ranging from 0.3m to 0.8m, preferably 0.5m, and a minimum number of cluster points minPts of 3 points. The cluster center coordinates and the percentage of fixation points within each cluster are extracted. Gait rhythm features are extracted from gait data, consisting of the mean gait frequency, stride standard deviation, and number of start-stop state transitions within a 3-5s sliding window, preferably 4s. In the formula: This is the gait rhythm feature vector; This is the normalized value of the step frequency mean; This represents the normalized value of the stride standard deviation; This is the normalized value for the number of start-stop switching operations. The travel trend characteristic is represented by a vector synthesized from the first-order difference of the instantaneous walking speed and the rate of change of the travel direction angle within a 3-5s window, preferably 4s. In the formula: This represents the feature vector of the travel trend; The normalized value of the first-order difference mean of the instantaneous walking speed; This is the normalized value of the mean rate of change of the direction angle.

[0050] S13. Cross-modal spatiotemporal coupled fingerprint construction unit, the specific working steps are as follows:

[0051] S131. Extract the saccade direction sequence of the user's saccade trajectory and the turning direction sequence of the turning trajectory within the current time window: The saccade direction sequence is calculated by the difference in coordinates of the start and end gaze points of adjacent saccade events. The arctangent of each direction vector is taken to obtain the angle value, ranging from 0° to 360°. The time window length is 5s to 15s, preferably 10s. The turning direction sequence is composed of continuous sampled values ​​of the walking direction angle in the gait data within the corresponding time window, with a sampling frequency of 50Hz. The preset time delay range is -500ms to +500ms, and the step interval is 20ms to 50ms, preferably 25ms.

[0052] The scan direction sequence is divided according to the number of sampling points corresponding to this offset. After translation, the coupling degree of the two sets of direction sequences under different time delay offsets is calculated using a cross-correlation function. The cross-correlation range is -1 to 1, with a higher absolute value indicating a higher coupling degree. Specifically: In the formula: For the first The directional angle of a scan event, in degrees; The coordinates of the starting point of this scanning event are in meters. The coordinates of the endpoint of the scanning event are in meters. Time delay offset The degree of cross-correlation coupling; To effectively scan the number of events; For the first step in the steering direction sequence Each sample value is in degrees. This is the time delay offset, in milliseconds (ms). The sampling period for the gait sensor is set to 20ms. Time delay offset The corresponding sampling point offset integer. After traversing all time delay offsets, the time delay coupling curve is obtained, and the peak of the curve corresponds to the maximum synchronization time delay.

[0053] S132. Extract the maximum coupling degree and its corresponding time delay offset from the time-delay coupling curve: The maximum coupling degree is the peak value of the absolute value of the cross-correlation function, ranging from 0 to 1, with a preferred value greater than 0.6 indicating effective coupling. The time delay offset is the abscissa value corresponding to the peak value, ranging from -500ms to +500ms. A negative value indicates that eye saccades lead gait turning, and a positive value indicates that eye saccades lag behind gait turning. Combine the maximum coupling degree and the time delay offset into a two-dimensional feature vector, which is the cross-modal spatiotemporal coupling fingerprint. This fingerprint uniquely represents the user's individual behavioral pattern in visual search and body turning coordination, where the maximum value reflects the strongest synchronization level between eye movement and gait in directional changes, and the time delay offset reflects the time lead or lag relationship between the two.

[0054] S133. Compare the generated cross-modal spatiotemporal coupled fingerprint with a pre-stored user behavior template library: Each known user in the template library stores at least 3 sets of historical coupled fingerprint samples, each set containing the maximum coupling value and the time delay offset. During comparison, Euclidean distance is used to calculate the distance between the current fingerprint and each template sample, i.e.: In the formula: For the current fingerprint and the first The first template Standardized Euclidean distance of each sample; These are the dimension weights for the maximum coupling value and the time delay offset, respectively, with values ​​ranging from 0.4 to 0.6, preferably 0.5; The normalized scaling factor for the time delay offset is the sum of the absolute values ​​of the preset time delay range, i.e., 1000 ms. This represents the maximum coupling level of the current fingerprint. For the first The first template The maximum coupling degree of each sample; This represents the time delay offset of the current fingerprint, in milliseconds. For the first The first template The time delay offset of each sample, in milliseconds.

[0055] The average matching distance is obtained by averaging the distances of multiple samples using the same template. In the formula: For the first The average matching distance of each template; For the first The number of historical samples for each template shall be no less than 3. For the current fingerprint and the first The first template The distance between each sample. The reciprocal of the average matching distance and normalized to the interval between 0 and 1 is taken as the degree of matching, i.e.: In the formula: For the first The degree of matching for each template, with a value range of 0 to 1; The total number of templates in the template library; the matching degree threshold is set to 0.7 to 0.9, with 0.8 being preferred. When the highest matching degree exceeds the threshold, the user identity identifier corresponding to the template is output as the cross-device consistent identity identifier of the current user, thereby achieving unique binding of user identity across camera scenarios.

[0056] S14. Intent Conflict Resolution Unit, the specific working steps are as follows:

[0057] S141. Real-time reading of the user's eye gaze point's two-dimensional coordinates from the eye-tracking system at a frequency of 60Hz to 120Hz. The origin of the coordinate system is set to the intersection of the southwest corner of the exhibition hall and the ground, with the X-axis along the due east direction and the Y-axis along the due north direction, in meters. Simultaneously, the user's current walking direction angle is read from the gait acquisition system at a frequency of 50Hz, with due north as the reference point and the angle increasing clockwise, in degrees. The gaze point coordinates are mapped to a circular visual attention area in the exhibition hall's planar coordinate system, with a radius of 0.3m to 0.8m, preferably 0.5m.

[0058] The direction angle is converted into a pointing vector. Starting from the user's current foot position, a fan-shaped moving target area is formed by extending along this direction for 0.5m to 2m, preferably 1m. The fan angle is 30° to 60°, preferably 45°. Specifically: , In the formula: The coordinates of the center of the target area are in meters. The coordinates of the user's foot position, in meters; The length extended in the direction of travel; The direction of travel is the angle, expressed in degrees (°).

[0059] S142. Calculate the Euclidean distance between the center of the visual attention region and the center of the moving target region, i.e.: In the formula: The Euclidean distance between the center of the visual attention area and the center of the moving target area, in meters; The coordinates of the center of the visual focus area are in meters. The coordinates of the center of the target area are in meters. The preset spatial conflict threshold is 0.5m to 2m, preferably 1m; when the Euclidean distance exceeds this threshold, it is determined that the current time is in an intention conflict state.

[0060] In this state, the historical gaze duration and number of gazes for the current exhibit are extracted within the most recent 10-30 second sliding time window, with a 20-second window being preferred. The visual dominance intent weight is calculated by weighted fusion after normalizing the historical gaze duration and the number of historical gazes, i.e.: In the formula: The visually dominant intent weight, with a value range of 0 to 1; The duration of historical observation is measured in seconds (s). This refers to the number of times a view has been viewed in history. As a normalized reference for fixation duration, 20 seconds is used; The normalization benchmark for fixation counts is set at 20. The fixation duration weighting coefficient ranges from 0.4 to 0.6, with 0.5 being preferred. The fixation frequency weighting coefficient ranges from 0.05 to 0.15, with 0.1 being preferred.

[0061] S143. Extract the user's historical dwell time and average passage speed for the current exhibit within the most recent 10s to 30s sliding time window, preferably a 20s window. Historical dwell time is the cumulative time the user spends within a 2m radius of the exhibit. The weight of the mobile dominant intent is calculated by weighting and fusing the normalized historical dwell time and average passage speed, i.e.: In the formula: The weight represents the mobile-dominant intent, with a value range of 0 to 1; The duration of the historical record is expressed in seconds (s). The average speed is expressed in m / s. As a normalized baseline for dwell time, 20 seconds is used; The maximum passing speed is set to 2 m / s; The dwell time weighting coefficient ranges from 0.3 to 0.5, with 0.4 being preferred. The speed weighting coefficient ranges from 0.2 to 0.4, with 0.3 being preferred.

[0062] Dividing the visual dominant intent weight by the mobile dominant intent weight yields the conflict resolution coefficient, i.e.: In the formula: The coefficient for conflict resolution ranges from 0 to positive infinity. The visual dominant intent weight is set. The preset judgment threshold is between 0.8 and 1.2, with 1.0 being preferred. When the conflict resolution coefficient is greater than this threshold, the visual dominant intent is determined to be dominant, and the user is more likely to stop and observe; when it is less than this threshold, the mobile dominant intent is determined to be dominant, and the user is more likely to continue moving.

[0063] S15. Feature fusion coding unit, the specific working steps are as follows:

[0064] Eye-tracking attention characteristics are quantified as the mean duration of effective fixation within the current 2-second sliding window and the rate of change in pupil diameter. The mean fixation duration is in seconds, ranging from 0 to 20, and normalized to 0 to 1. The rate of change in pupil diameter is the ratio of the standard deviation to the mean of pupil diameter within the window, ranging from 0 to 0.5, and normalized to 0 to 1. The two values ​​are weighted and summed to obtain the scalar value of eye-tracking attention, i.e.: In the formula: As a scalar measure of eye-tracking attention; This is the normalized value of the mean fixation duration; This represents the standard deviation of pupil diameter, in mm. This represents the average pupil diameter in mm. The fusion weights are set to 0.6 and 0.4 respectively.

[0065] Visual interest region features are quantified as the exhibition hall planar coordinates of the spatial density cluster center of the gaze point at the current moment and the proportion of gaze points in that cluster, with coordinates in meters and proportion values ​​ranging from 0 to 1. Gait rhythm features are composed of the mean gait frequency, standard deviation of stride length, and number of start-stop transitions within a 4-second sliding window, concatenated. Gait frequency is measured in steps per minute, ranging from 0 to 180; stride length is measured in centimeters, ranging from 0 to 150; and the number of start-stop transitions ranges from 0 to 5. These three items are normalized to 0 to 1 and then concatenated into a three-dimensional vector.

[0066] The movement trend characteristic is composed of the concatenation of the first-order difference mean of instantaneous walking speed and the mean rate of change of the direction angle within the same 4-second window, with the speed difference unit being m / s. 2 The rate of change of orientation angle is in ° / s. Two terms, after Z-score standardization, are concatenated into a two-dimensional vector. The eye-tracking focus scalar, the three-dimensional vector of the visual interest region, the three-dimensional vector of gait rhythm, and the two-dimensional vector of walking trend are then concatenated into a nine-dimensional original feature vector, i.e.: In the formula: This is the original nine-dimensional feature vector; The coordinates of the cluster centers of the visual interest regions are in meters. This represents the percentage of fixations within a cluster, with a value ranging from 0 to 1. Defined as S12; then concatenated with the visual dominant intent weight, mobile dominant intent weight, and conflict resolution coefficient to form a twelve-dimensional fusion input, i.e.: In the formula: The input vector is a twelve-dimensional fusion vector; The definitions are the same as those in S142 and S143.

[0067] A fully connected neural network is used to weight and encode the twelve-dimensional fused input. The network contains a single hidden layer with 128 to 512 neurons, preferably 256. The weights are initialized with a Gaussian distribution with a mean of 0 and a standard deviation of 0.01. After ReLU activation and layer normalization, the output is a 128- to 512-dimensional user-consistent binding tour state feature vector, preferably 256 dimensions, representing the coupling relationship between the user's visual cognition and motion control state at the current moment. In the formula: Bind the browsing state feature vector to users for consistency; This is the weight matrix of the fully connected layer; Here, represents the bias vector of the fully connected layer; LayerNorm is the layer normalization operator; ReLU is the activation function. Number of output lines This represents the number of neurons in the hidden layer.

[0068] S2. Cognitive-Motion Conflict Analysis and Tour Behavior Prediction Module: This module performs antagonistic analysis of eye movement and gait timing to obtain a visual-motion conflict index, identifies inflection points as the points of intent to stay, calculates implicit interest transmission probabilities based on semantic association networks to generate implicit interest transmission paths, constructs a dynamic 3D user profile including a cognitive fatigue index, and generates behavior prediction results. Specifically, it includes:

[0069] S21. Timing prediction unit, the specific working steps are as follows:

[0070] Receive the user's consistent browsing state feature vector. This vector has a dimension of 256. With a time step of 20 frames as the input window and an interval of 100ms per frame, corresponding to 2 seconds of historical data, construct a time-series feature matrix, i.e.: In the formula: The input is the time series feature matrix; The time step is set to 20; The feature dimension is set to 256; input to the lightweight deep learning prediction model:

[0071] The model input layer receives the aforementioned temporal matrix. The temporal coding layer employs a single-layer bidirectional gated recurrent unit with a hidden layer dimension of 96. It performs forward and backward bidirectional encoding on the input temporal sequence to extract the temporal dependency between gait rhythm and eye movement shift, i.e.: In the formula: The output feature sequence of the bidirectional gated recurrent unit, dimension BiGRU is a bidirectional gated recurrent unit (BRU) computation. The attention layer employs a multi-head self-attention mechanism with four attention heads, weighting the output of the bidirectional gated recurrent unit to enhance the feature representation of key time steps, i.e.: In the formula: For multi-head self-attention, output feature sequence, dimension MultiHeadAttn is a multi-head self-attention operation with a set number of 4 attention heads.

[0072] The feature fusion layer flattens the attention output into a vector, maps it to 192 dimensions through a fully connected layer, uses ReLU as the activation function with a dropout rate of 0.2, and fuses high-order temporal representations. The multi-task output layer sets up three prediction branches in parallel.

[0073] The movement position prediction branch is a fully connected layer, outputting 2D coordinate values, corresponding to the x-axis and y-axis coordinates of the exhibition hall's planar coordinate system, respectively: In the formula: A two-dimensional coordinate vector for predicting the travel position, in meters; The branch weight matrix for location prediction; Output a flattened vector for the attention function; The location prediction branch bias vector predicts the user's spatial location within the next 2 to 5 seconds, with a preferred value of 3 seconds. The target exhibit attention probability branch is a fully connected layer, with an output dimension equal to the total number of exhibits in the exhibition hall. After normalization using the Softmax function, the attention probability distribution of each exhibit is obtained. The exhibit corresponding to the maximum probability is selected as the predicted target, i.e.: In the formula: This represents the probability distribution vector of attention for each exhibit. Predict the branch weight matrix for exhibits; The bias vector for the exhibit prediction branch is ; Softmax is the normalized exponential function. The expected dwell time branch is a fully connected layer, outputting a 1D regression value, which is constrained to be non-negative by ReLU activation, i.e.: In the formula: This is the predicted duration of stay, in seconds, ranging from 0 to 300. For duration prediction branch weight matrix; is the bias vector for the duration prediction branch; ReLU is the linear rectified activation function.

[0074] The model was trained using the Adam optimizer with an initial learning rate of 0.001 and a batch size of 48. The loss function was a weighted sum of the mean squared error of the travel position, the cross-entropy of the attention probability, and the mean squared error of the dwell time, i.e.: In the formula: This is the total loss function; For batch size, take 48; The first The actual and predicted values ​​of the travel position of each sample; It is the cross-entropy function; The first The true label and predicted distribution of the exhibit attention probability for each sample; The first The actual and predicted duration of stay for each sample; The loss weights are set at a ratio of 1:1:0.5.

[0075] S22. Antagonistic Relationship Analysis Unit, the specific steps are as follows:

[0076] S221. Using a preset-length sliding time window as the analysis unit, continuously extract spatiotemporally aligned eye-tracking and gait time-series data. The sliding time window length is 0.5s to 2s, preferably 1s, corresponding to 60 to 120 eye-tracking data sampling points and 30 to 60 gait data sampling points. The window step size is 0.1s to 0.5s, preferably 0.2s, to ensure overlap between adjacent windows. During continuous extraction, the window length is traced back from the current time point, and the eye-tracking and gait time-series data within that time period are extracted from the spatiotemporally aligned unified time grid to form an analysis unit. The eye-tracking data includes the spatial coordinates of the gaze point and the head position coordinates at each sampling time, and the gait data includes the instantaneous walking speed value at each sampling time. The window slides sequentially to form a continuous sequence of analysis units, providing a temporal basis for subsequent antagonistic relationship analysis.

[0077] S222. For the current sliding time window, read the user's head position coordinates and gaze point spatial coordinates at each sampling time within the window, calculate the difference between the two to obtain the offset vector, which contains x-direction and y-direction components, i.e.: In the formula: For the first The offset vector at each sampling time; For the first Spatial coordinates of the gaze point at each sampling time, in meters; For the first The coordinates of the user's head position at each sampling time point are given in meters. The offset vector sequence is obtained by traversing all sampling times within the window.

[0078] Calculate the standard deviation of the time series components in the x and y directions of this sequence, i.e.: , In the formula: These are the standard deviations of the time series of the x- and y-direction components of the offset vector, respectively, in meters. This represents the number of eye-tracking sampling points within the current sliding time window; Offset vectors The x- and y-direction components; These are the arithmetic mean of the components in the corresponding directions.

[0079] The offset distribution dispersion is calculated based on the spatial distribution characteristics of all offset vectors. The unit is m; the visual lock-on strength index is defined as In the formula: The standard deviation for the exhibition hall space is taken as 0.5m; The value ranges from 0 to 1, is dimensionless, and is considered to be effective when it is greater than 0.95. The higher the visual lock intensity index, the more concentrated the distribution of the user's gaze in space, and the stronger the stability of the visual attention system's lock on the external exhibit target, that is, the user is highly concentrating visual resources on a specific exhibit area.

[0080] S223. Synchronously read the instantaneous walking speed values ​​at each sampling moment within the current sliding time window, in m / s, to form a speed time series. Calculate the arithmetic mean and standard deviation of this series, i.e.: In the formula: The instantaneous walking speed within the window is the arithmetic mean, in m / s; The standard deviation of instantaneous walking speed within the window is expressed in m / s. This represents the number of gait sampling points within the current sliding time window. For the first The instantaneous walking speed values ​​at each sampling time point are expressed in m / s; the time series variation coefficient of the instantaneous walking speed is obtained by dividing the standard deviation by the arithmetic mean, i.e.: ,when season This indicates that there is currently no body forward drive, and this window is not included in the high-conflict state determination, where: The coefficient of variation is a dimensionless index of gait propulsion intensity, ranging from 0 to positive infinity. This coefficient of variation is used as the gait propulsion intensity index, with a preferred value greater than 0.3 indicating significant propulsion intensity.

[0081] The visual-motor conflict index is then calculated by dividing the visual lock-on intensity index by the gait propulsion intensity index, i.e.: In the formula: is the visual-motor conflict index, dimensionless, with a range from 0 to positive infinity; The visual lock intensity index is dimensionless; when When =0, let =0, this window is not included in the high-conflict state determination and subsequent inflection point identification sequence. When both the visual lock intensity index and the gait propulsion intensity index are at a high level, this ratio effectively quantifies the degree of antagonism between the visual attention system and the motion control system.

[0082] S224. The preset visual lock intensity high conflict determination threshold is set to 0.90 to 0.97, preferably 0.95; the preset gait advance intensity high conflict determination threshold is set to 0.2 to 0.5, preferably 0.3. When the visual lock intensity index of the current window exceeds its threshold and the gait advance intensity index exceeds its threshold, the window is determined to be in a high conflict state.

[0083] Arrange the time windows in chronological order and calculate the visual-motion conflict index difference between adjacent windows, i.e.: In the formula: For the first The window and the first Visual-motion conflict index difference of each window For the first The visual-motion conflict index of each window. The formal condition for identifying inflection point moments is: the existence of an index. This makes it possible to satisfy And for satisfy And the difference between the peak and the trough Those that meet the above conditions This is identified as the point at which the user's intention to stay engaged is triggered.

[0084] This inflection point is identified as the peak of user intent to stay. Based on the time difference between the inflection point and the current time, combined with historical dwell time statistics, the probability and duration of dwelling within the next 2 to 5 seconds are predicted, with 3 seconds being the optimal value.

[0085] The probability of stopping is calculated by superimposing a time decay factor on the peak value of the visual-motion conflict index at the inflection point after Sigmoid mapping, i.e.: In the formula: To predict the probability of stopping, the value range is 0 to 1; The peak value of the visual-motion conflict index at the inflection point; The sensitivity coefficient, with a value of 2 to 5, is preferred; As the criterion for determining the probability of stopping, we take 0.5; The time decay coefficient ranges from 0.1 / s to 0.5 / s, with 0.3 / s being preferred. The time difference between the current moment and the inflection point, in seconds. season The predicted dwell time at the inflection point is calculated by mapping the peak value to the baseline dwell time, i.e.: In the formula: Predict the dwell time at the S22 inflection point, in seconds; The base dwell time is set between 10 and 30 seconds, with 20 seconds being the preferred value.

[0086] The predicted dwell time at the inflection point is weighted and fused with the expected dwell time output by the S21 time series prediction unit to generate a comprehensive predicted dwell time, namely: In the formula: To comprehensively predict the length of stay; The expected dwell time output by the S21 deep learning model; The weighting factor is set to 0.3 to 0.5, with 0.4 being preferred. The predicted dwell time is used to match the content length of the S32 playback strategy adaptive unit.

[0087] S23. Cognitive State Perception Unit, the specific working steps are as follows:

[0088] S231. Pupil diameter measurements (in mm) are read from the spatiotemporally aligned eye-tracking time-series data at sampling frequencies from 60Hz to 120Hz. The original pupil diameter sequence is filtered using a 3-point to 7-point moving average, preferably 5 points, to remove high-frequency measurement noise. The ratio of the standard deviation to the mean of the pupil diameter within the current 1-3s moving window is calculated as a quantitative representation of the rate of change of pupil dilation frequency, i.e.: , , In the formula: For the first The pupil diameter measurements at each sampling time, in mm; This represents the number of eye-tracking sampling points within the current 1-second to 3-second sliding window. This represents the average pupil diameter within the window, in mm. The standard deviation of the pupil diameter within the window is expressed in mm. The pupil dilation frequency change rate is dimensionless and ranges from 0 to 0.5. A value greater than 0.1 is considered significant pupil activity.

[0089] Simultaneously calculate the percentage change in pupil diameter within this window relative to the baseline mean of the previous 1 second, i.e.: In the formula: The average pupil diameter within the baseline window in the first 1 second is expressed in mm. The percentage change in pupil diameter is dimensionless, ranging from -20% to +20%. The rate of change in pupil dilation frequency is weighted and fused with the percentage change to output a visual cognitive arousal index, namely: In the formula: They are respectively and The value after normalization to 0 to 1; This is a visual cognitive arousal indicator, with a value range of 0 to 1. The higher the value, the stronger the user's visual cognitive resources are mobilized.

[0090] S232. Instantaneous walking velocity values ​​(in m / s) are read from the spatiotemporally aligned gait time series data at a sampling frequency of 50 Hz. The velocity sequence is smoothed by a sliding window of 0.5 s to 1 s, preferably 0.8 s, to eliminate instantaneous jitter. The difference between the average walking velocity within the current 2 s to 4 s sliding window and the average walking velocity within the baseline window of the previous 2 s to 4 s is calculated, and then divided by the baseline velocity to obtain the relative deceleration rate, i.e.: , , In the formula: The average walking speed within the current 2-4 s sliding window is expressed in m / s. This represents the number of gait sampling points in the current window. For the first in the window The instantaneous walking speed value at each sampling time, in m / s; The average walking speed during the baseline window from the first 2 seconds to the last 4 seconds is expressed in m / s. This represents the number of gait sampling points within the baseline window. For the first time within the baseline window The instantaneous walking speed value at each sampling time, in m / s; The relative deceleration rate is dimensionless and ranges from 0 to 1; when When <0.01m / s, let =0.

[0091] Simultaneously, the percentage of sampling points with speeds below 0.2 m / s within this window is calculated as the low-speed residence ratio, i.e.: , In the formula: The low-speed residence ratio ranges from 0% to 100%, with a preferred value greater than 30% indicating a tendency towards stillness. This represents the number of sampling points within the current window with a velocity lower than 0.2 m / s. This is an indicator function; it takes the value 1 if the condition is true, and 0 otherwise. The absolute value of the relative deceleration rate is then combined with the low-speed stagnation ratio according to weights to output the gait deceleration index, i.e.: In the formula: This is a gait deceleration index, with a value range of 0 to 1. The higher the index, the more significant the reduction in the user's forward drive.

[0092] S233. Input the visual cognitive arousal index and the gait deceleration index into the fusion layer, and sum them according to their weights to generate the cognitive intervention readiness index, i.e.: In the formula: The cognitive engagement readiness index ranges from 0 to 1. This index considers both cognitive resource mobilization and physical deceleration. A higher value indicates a greater probability that the user will transition from a walking browsing state to a deep cognitive processing state. As a visual cognitive arousal indicator As an indicator of gait deceleration, The fusion weights are set to values ​​of 0.5 to 0.6 and 0.4 to 0.5, respectively, with 0.55 and 0.45 being preferred.

[0093] A threshold of 0.6 to 0.8, preferably 0.7, is set for determining the state of deep cognitive intervention. When the cognitive intervention readiness index exceeds this threshold, the current moment is recorded as the start time of the time window, and the index is monitored for a sustained period of 2 to 5 seconds, preferably 3 seconds. If the index remains above the threshold, the period from the start time to the point where the index falls below the threshold is defined as the cognitive window period. The cognitive intervention readiness index value and the start and end times of the time window are output to guide the preloading of narrative content before the start of the cognitive window period.

[0094] S24. Interest Transmission Analysis Unit, the specific working steps are as follows:

[0095] S241. Read the network structure data from the exhibition hall exhibit semantic association network database. This network uses each exhibit as a node, and the semantic associations between exhibits as edges. The initial weight of the edge represents the association strength, with a value ranging from 0 to 1. The initial weight is pre-calculated and stored based on exhibit theme similarity, historical period association, and technological overlap. Use the target exhibit attention probability output by the time series prediction unit as the initial incentive value. This probability value is a normalized value between 0 and 1, and is loaded onto the node corresponding to the target exhibit in the exhibition hall exhibit semantic association network as the initial probability value of that node. The initial probability values ​​of all other non-target exhibit nodes are set to 0. After loading, the target exhibit node holds all the initial probability mass, and the network is in a state of waiting to spread, providing the starting conditions for subsequent probability propagation calculations.

[0096] S242. Starting with the target exhibit node as the probability diffusion starting point, an iterative probability propagation algorithm is used to perform diffusion. In each iteration, each node distributes its current probability value to all directly adjacent nodes according to the outgoing edge weight ratio. At the same time, a damping coefficient is introduced to retain a portion of the probability in the current node to prevent excessive probability loss, i.e.: In the formula: For the first During the first iteration The probability value of each node; The damping coefficient ranges from 0.15 to 0.35, with 0.25 being preferred. To point to the first The set of indexes of all predecessor nodes of a node; For the node To the node Edge weights; For the node to its neighboring nodes Edge weights.

[0097] The probability is propagated along multiple paths to indirectly related nodes of each order. After each iteration, the change in the L1 norm of the probability distribution of the entire network is calculated, i.e.: In the formula: For the first Wheel and the first The change in L1 norm between wheels; This represents the total number of nodes in the semantic association network of exhibits in the exhibition hall. When this change is less than the preset convergence threshold of 0.001 to 0.01, preferably 0.005, or when the number of iterations reaches the maximum limit of 50 to 200, preferably 100, the iteration is terminated, the probability distribution of the entire network is determined to have converged, and the final probability value of each node is obtained.

[0098] S243. The final probability value held by each node after the convergence iteration terminates is defined as the implicit interest transmission probability. This probability value quantifies the strength of interest energy transmission from the target exhibit to each associated exhibit, with a value range of 0 to 1. The preset transmission threshold is 0.02 to 0.15, preferably 0.05. Traverse all nodes and mark the exhibits corresponding to nodes with implicit interest transmission probabilities greater than this transmission threshold as implicitly associated exhibits. Sort all implicitly associated exhibits in descending order of implicit interest transmission probability, and select the top 3 to 10, preferably 5, to form the target exhibit prediction sequence together with the target exhibit. This sequence contains 1 direct target exhibit and multiple implicitly associated exhibits, arranged in descending order of probability value, and is used to guide the priority preloading of subsequent narrative content.

[0099] S244. For each implicitly related exhibit, trace back the probability propagation path from the target exhibit node to the implicitly related exhibit node during the probability diffusion process, and record the sequence of all intermediate nodes along the path. This sequence includes the starting target exhibit node, the indirect related nodes of each order along the way, and the ending implicitly related exhibit node. If multiple probability propagation paths reach the same implicitly related exhibit node, select the path with the largest probability product as the implicit interest transmission path for that node. Summarize the implicit interest transmission paths corresponding to all implicitly related exhibits and output them together with the target exhibit prediction sequence. The output implicit interest transmission paths represent the direction of interest transmission and the association chain between exhibits in the form of a node sequence, which is used to connect the related knowledge points of adjacent exhibits when arranging the narrative logic.

[0100] S25. Interest Measurement and Profile Building Unit, the specific work steps are as follows:

[0101] Five indicators were extracted for the current exhibit: gaze duration, pupillary variation, blink frequency, resting posture, and walking speed. Gaze duration was the cumulative value within a sliding window of 10 to 30 seconds prior to the current exhibit, in seconds, ranging from 0 to 60. Pupil variation was the standard deviation of pupil diameter within the window, in mm, ranging from 0 to 0.5. Blink frequency was the number of blinks per minute within the window, ranging from 0 to 30. Resting posture was determined by the variance of plantar pressure center of gravity, with a variance less than 0.01m. 2 The time is recorded as a stationary state, with a value of 0 or 1. The walking speed is the average value within the window, in meters per second, ranging from 0 to 2.

[0102] After normalizing the five indicators to the 0-1 range, a weighted sum is calculated to obtain the comprehensive interest score, i.e.: In the formula: The score is a comprehensive interest score, with a value range of 0 to... This is the normalized value for gaze duration; This represents the normalized value of pupillary changes; This is the normalized value for blink frequency; This represents the normalized value for the dwell posture. The normalized value is the walking speed. A high interest threshold of 0.7 to 0.9, preferably 0.8, is set; a medium interest threshold of 0.3 to 0.5, preferably 0.4, is set. A score greater than 0.8 indicates high interest, between 0.4 and 0.8 indicates medium interest, and less than 0.4 indicates low interest.

[0103] Based on cross-device consistent identity identification, historical user behavior data is retrieved, and preference tags are generated by statistically analyzing the historical average interest levels for each exhibit category, with 5 to 10 categories. The tour pace tag is generated by clustering users' average walking speed and average dwell time, categorized into three types: fast, balanced, and slow. The cognitive level tag is determined by the historical distribution mean of the cognitive intervention readiness index, categorized into three levels: shallow, medium, and deep.

[0104] The cognitive fatigue index is calculated by combining the duration of continuous sightseeing with the increasing trend of blinking frequency, namely: In the formula: The cognitive fatigue index ranges from 0 to 1, with a value exceeding 0.7 indicating fatigue. The duration of the tour is continuous, in minutes. The starting time for fatigue accumulation is set between 20 and 40 minutes, with 30 minutes being preferred. The blink rate within the current 1-minute sliding window, in times / min; The baseline blink rate during the initial stage of the tour is expressed in blinks per minute. The maximum blinking frequency under fatigue conditions is determined by taking the upper limit of the blinking frequency statistics of the user group after continuous browsing for more than 30 minutes in history. The preset constant is 25 times / min to 35 times / min, with 30 times / min being preferred. For joint weighting, values ​​of 0.6 and 0.4 are used respectively. Preference tags, rhythm tags, cognitive level tags, and cognitive fatigue index are combined to form a dynamic 3D user profile. The final output behavior prediction results include target exhibit identifiers, dwell probability values, predicted dwell time, visual-motor conflict index, target exhibit prediction sequence, implicit interest transmission path, interest level, and the dynamic 3D user profile.

[0105] S3. Personalized Narrative Adaptive Scheduling and Closed-Loop Optimization Module: This module adjusts the playback start time based on inflection points and conflict indices, reorganizes narrative logic according to implicit interest transmission paths, triggers cognitive deload mode when the cognitive fatigue index exceeds a threshold, and constructs a group implicit interest transmission network to iteratively optimize the prediction model. Specifically, it includes:

[0106] S31. Content Matching and Preloading Unit, the specific working steps are as follows:

[0107] After receiving the behavior prediction results, the target exhibit prediction sequence is analyzed, and the preloading priority order is determined by sorting the exhibits in the sequence according to the probability of implicit interest transmission from high to low. The corresponding narrative content for each exhibit is retrieved from a pre-generated multi-level narrative resource library, which contains four types of content for each exhibit: basic popular science version, in-depth detailed version, interesting story version, and professional academic version.

[0108] When the probability of stopping exceeds a preset threshold of 0.6 to 0.9, preferably 0.75, the preloading process is triggered. The pre-playback start timing correction is calculated based on the cognitive window start time corresponding to the cognitive intervention readiness index. When the cognitive window start time is less than 3 to 8 seconds from the current time, preferably 5 seconds, content decoding is initiated ahead of schedule. Simultaneously, the arrival time is calculated by combining the predicted movement position with the real-time spatial distance between the user and the target exhibit. , ,when season , Take 10 seconds; if the real-time spatial distance at this time Then pre-playback is triggered directly, where: The real-time spatial distance between the user and the target exhibit, in meters; The coordinates of the user's current showroom layout, in meters. , The coordinates of the target exhibit's location are in meters. The estimated arrival time is in seconds. The user's current average walking speed is expressed in m / s. When the real-time spatial distance is less than 1m to 3m, preferably 2m, and the estimated arrival time is less than 1s to 3s, preferably 2s, the advance timing of the start is further adjusted.

[0109] Within the first preset time window of 0.5s to 2s, preferably within 1s, after the inflection point is determined, the preloading and preplay of the narrative content corresponding to the current exhibit is initiated to ensure that the content is buffered before the user arrives. At the same time, based on the implicit interest transmission path, within the pre-visit interval of 1m to 5m, preferably 3m, before the user arrives at the direct target exhibit, the narrative content of implicitly related exhibits is preloaded into the terminal cache queue in descending order of transmission probability to achieve seamless narrative connection.

[0110] S32. Playback strategy adaptive unit, the specific working steps are as follows:

[0111] The narrative length and depth are matched based on the predicted dwell time and level of interest. A basic science version is used for a predicted dwell time of less than 30 seconds; an in-depth explanation or an engaging story version is used for 30 to 90 seconds; and a professional academic version is used for more than 90 seconds. For a high level of interest, the in-depth explanation version is prioritized; for a medium level, the basic science version is used; and for a low level, a concise summary version is used.

[0112] The narrative content type and presentation format are selected based on preference tags, rhythm tags, and cognitive level tags in the dynamic 3D user profile. Preference tags pointing to technology prioritize technical analysis content, while those pointing to humanities prioritize storytelling content. For fast-paced content, short videos are chosen; for slower-paced content, audio narration is selected. For deeper cognitive levels, the density of technical terminology is increased; for shallower levels, visual elements are added. The content entry rhythm is jointly characterized by the duration of the narrative prelude and the fading rate. The basic prelude duration is 3s to 8s, preferably 5s, and the fading rate is a linear volume gain slope of 0.2 / s to 0.5 / s. When the visual-motor conflict index is greater than 0.8, the prelude duration is extended to 8s to 15s, preferably 10s, and the fading rate is reduced to 0.1 / s to 0.2 / s to achieve a smooth entry through a gradual transition from blur to clarity. When the conflict index is less than 0.3, the prelude duration is shortened to 0s to 2s, preferably 1s, and the fading rate is increased to 0.5 / s to 1.0 / s to directly enter the core content.

[0113] The start playback time and content insertion rhythm are dynamically adjusted based on the real-time value of the visual-motion conflict index. When the index is greater than 0.8, the start playback time is delayed by 0.5s to 2s, preferably 1s, and gradually inserted after the user's visual lock is stable. When the index is between 0.3 and 0.8, the content is inserted at the normal time. When the index is less than 0.3, playback is initiated 0.5s earlier.

[0114] Dynamically adjust playback speed, switching speed, and display duration based on user gait rhythm, eye movement focus, and cognitive fatigue index:

[0115] When the step frequency is greater than 120 steps per minute, the speech rate is increased to 1.2 to 1.5 times the original speech rate, preferably 1.3 times, and the switching speed is accelerated by 30% to 50%. The speech rate adjustment formula is as follows: In the formula: To adjust the playback speed; Original speech rate; The user's current cadence, in steps per minute; The baseline value for normal cadence is 100 steps per minute; The upper limit of the step frequency is set at 180 steps per minute. This is the step frequency adjustment coefficient, ranging from 0.2 to 0.5, preferably 0.3; Eye-tracking focus, with a value range of 0 to 1; The focus adjustment coefficient ranges from 0.1 to 0.3, with 0.2 being preferred. The switching speed adjustment formula is as follows: In the formula: To adjust the content switching speed; Based on switching speed; The switching speed adjustment coefficient is set to a value between 0.3 and 0.5, preferably 0.4. The user's current cadence, in steps per minute; This is the normal step frequency baseline value; This is the baseline value for the upper limit of the step frequency. And it sets the final speech rate... In the formula: To achieve the lowest effective speaking speed, take .

[0116] When eye-tracking focus is greater than 0.7, the speech rate is reduced to 0.8 to 1.0 times, preferably 0.9 times, and the presentation duration is extended by 20% to 40%. The formula for adjusting the presentation duration is as follows: In the formula: The adjusted display duration is in seconds (s). The basic display duration, in seconds; The focus duration adjustment coefficient ranges from 0.2 to 0.4, with 0.3 being preferred. For eye-tracking focus; The fatigue duration adjustment coefficient ranges from 0.4 to 0.6, with 0.5 being preferred. Cognitive fatigue index

[0117] When the cognitive fatigue index exceeds the threshold of 0.7 to 0.9, preferably 0.8, the cognitive deload mode is triggered, automatically downgrading in-depth content to a basic science popularization version, and reducing the display time by 40% to 60%, preferably 50%.

[0118] S33. Narrative Logic Arrangement Unit, the specific work steps are as follows:

[0119] The system reads the user's current exhibition hall location coordinates and historical path in real time, using a sliding time window of 5 to 15 seconds, preferably 10 seconds, to analyze the curvature changes and angular offset of the user's browsing trajectory. When two or more exhibits show a consecutive jump in interest level from medium or low to high, where: In the formula: This refers to the number of times the interest level increases; The number of most recently viewed exhibits, take 3 to 5; For the first The interest level of each exhibit is quantified as follows: 1 for low, 2 for medium, and 3 for high; II is the indicator function; or when an exhibit with an implicit interest transmission probability of more than 0.05 is actually viewed by the user for more than 1 second, it is determined that an interest shift trend has formed.

[0120] The narrative order of subsequent exhibits is dynamically adjusted based on interest shift trends. The top 3 to 5 (preferably 4) implicitly related exhibits within a 3m to 8m radius (ideally 5m) of the user's current location, located on the implicit interest transmission path, are inserted into the subsequent positions of the current narrative sequence. If the user's actual direction of travel deviates from the original planned path by more than 45° for more than 3 seconds, the narrative order is regenerated starting from the current location, i.e.: In the formula: This represents the angular deviation between the actual direction of travel and the original planned path, expressed in degrees. The actual direction angle of travel for the user, in degrees; This represents the original planned path direction angle, in degrees. The deviation duration is expressed in seconds.

[0121] Based on the semantic association types between nodes along the implicit interest transmission path, the system automatically extracts related knowledge points between adjacent exhibits. Each exhibit is limited to extracting 1 to 3 related knowledge points, with 2 being the preferred option. The semantic association types include five categories: shared historical period, technological inheritance, material association, regional culture, and figures / events. Each semantic association type in the pre-stored exhibit knowledge graph corresponds to 1 to 3 related knowledge point templates. When the edge attribute between two adjacent nodes on the implicit interest transmission path belongs to one of the above semantic association types, the system automatically calls the corresponding related knowledge point template, outputting the pre-stored knowledge points of the two exhibits along that association dimension as related knowledge points. These related knowledge points are embedded in the transition sections of the narrative content of adjacent exhibits, presented through voice prompts or graphic hints, achieving a smooth connection of narrative logic and knowledge linking between exhibits.

[0122] The coherence of the generated narrative chain is evaluated by calculating the mean absolute value of the difference in interest level between adjacent exhibits in the chain, i.e.: In the formula: This is an index for narrative chain coherence, with a value range of 0 to 1. To cover the number of exhibits in the narrative chain, ; For the first in the chain The interest level of each exhibit is quantified as follows: 1 for low, 2 for medium, and 3 for high. When the index is higher than 0.7 and the narrative chain covers no less than 3 exhibits, it is judged as a coherent and effective personalized tour narrative chain and output to the terminal for display.

[0123] S34. Multi-terminal display and feedback collection unit, the specific working steps are as follows:

[0124] The exhibition hall is equipped with five types of devices: a large exhibition screen, a standing touchscreen, AR glasses, a mobile app, and a voice-guided terminal. The system calculates the Euclidean distance between the user's current location on the exhibition hall floor plan and the spatial coordinates of the terminal device, i.e.: In the formula: For users and the The Euclidean distance between the terminals is expressed in meters. For the first The exhibition hall coordinates of each terminal, in meters; the terminal closest to the user's line of sight within a 30° to 60° sector area is selected as the primary output device, preferably 45°; where: In the formula: This represents the angular offset of the terminal relative to the user's line of sight, in degrees. , This represents the user's current eye gaze direction angle, in degrees. When the user is more than 2m to 5m away from the target terminal, preferably 3m, the system automatically switches to the personal terminal.

[0125] Before output, content format is converted based on the terminal screen resolution and rhythm tags in the user profile. Large screens output high-definition video and mixed text / image content at a resolution of 1920x1080 or higher. Touchscreens output interactive text / images and short videos. AR glasses output 3D annotations and audio narration overlaid on exhibits. Mobile app outputs lightweight text / images and audio streams. Voice narration terminals output pure audio streams with a sampling rate of 44.1kHz.

[0126] Real-time collection of user content browsing feedback data, including content dwell time, number of interactive operations, content switching frequency, and number of repeated views. Content dwell time is obtained by recording the continuous duration of the user's gaze on the terminal screen. A valid gaze is defined as the gaze point falling within the terminal display area and lasting for more than 200ms. The criteria for a valid gaze are as follows: and At this time, In the formula: This refers to the duration of a single content session, measured in milliseconds (ms). These are the times when the gaze enters and leaves the terminal display area, respectively, in milliseconds; The coordinates of the user's gaze point; For the first The display area of ​​each terminal. Interactive operation count: Statistics on the number of valid triggers of user touch clicks, voice commands, or gesture interactions; a single operation response time of less than 500ms is considered valid. Content switching frequency: Calculation of the ratio of the number of times the user actively or passively switches content to the total browsing time, i.e.: In the formula: Content switching frequency, measured in times per minute; This represents the total number of times content was switched during the tour. Total viewing time, in seconds. Repeat viewing count records the number of times a user fully plays or partially replays the same narrative content.

[0127] S35. Closed-loop iterative optimization unit, the specific working steps are as follows:

[0128] S351. Read the structured browsing feedback dataset that has been aggregated to the central processing unit by the multi-terminal display and feedback collection unit in S34. The dataset includes content dwell time, number of interactive operations, content switching frequency and number of repeated views. The collection and judgment criteria for each data item are the same as in S34.

[0129] S352. Using user browsing feedback data as a sample, construct an exhibit-to-interest transfer matrix. For any two exhibits, count the number of times a user's line of sight or walking trajectory enters the vicinity of the second exhibit within 5 to 30 seconds (ideally 10 seconds) after leaving the first exhibit; this is recorded as the positive transfer frequency. Divide the positive transfer frequency by the total number of viewers of the first exhibit to obtain the interest transfer probability, i.e.: In the formula: To the exhibits To the exhibits The probability of interest shift, ranging from 0 to 1. For positive transition frequency, For exhibits Total number of visitors.

[0130] The current edge weights in the semantic association network of exhibits in the exhibition hall are incrementally accumulated based on the product of the interest transition probability and the reinforcement learning rate, i.e.: In the formula: For the updated node Pointing to node Edge weights; The corresponding edge weights before the update; To enhance the learning rate, a value between 0.01 and 0.1 is selected, with 0.05 being preferred. After incremental updates, the edge weights of the entire network are normalized so that the sum of the outgoing edge weights of each node is 1, i.e.: In the formula: To normalize the weights, For nodes The set of all successor node indices; generating a group implicit interest transmission network.

[0131] S353. Based on the updated edge weights in the swarm implicit interest transmission network, recalculate the interest transmission probability. Use the mean square error between the recalculated implicit interest transmission probability of each node and the normalized value of the node's actual access frequency in the real browsing feedback data as the first loss signal, i.e.: In the formula: This is the first loss signal; This represents the total number of nodes in the semantic association network of the exhibits in the exhibition hall. For exhibits recalculated based on the updated edge weights The probability of latent interest transmission is defined as in S243; For exhibits The proportion of exhibits where users actually stay for more than 3 seconds out of the total number of times they stay at an exhibit. The error threshold is set between 0.001 and 0.01, preferably 0.005; if the error exceeds the threshold, the edge weights are fine-tuned in the direction of gradient descent, with a step size of 0.001 to 0.01, preferably 0.005.

[0132] Simultaneously, the exhibit corresponding to the exhibit with the highest probability of attention output by the time-series prediction unit is compared with the actual target exhibits where users actually stayed for more than 3 seconds, and the cross-entropy loss is calculated as the second loss signal, i.e.: In the formula: This is the second loss signal; This represents the total number of exhibits in the exhibition hall. For exhibits The true target label is: 1 for exhibits where users actually stay for more than 3 seconds, and 0 for the rest; Exhibits output by the timing prediction unit The probability of attracting attention to the target exhibits is calculated. The model parameters are then backpropagated to a lightweight deep learning prediction model using the Adam optimizer, with a learning rate ranging from 0.0001 to 0.001, preferably 0.0005, and the model parameters are updated accordingly.

[0133] S354. Analyze user feedback on various narrative content types under different cognitive fatigue index ranges. Divide the cognitive fatigue index from 0 to 1 into 3 to 5 equally wide intervals, preferably 4 intervals, with each interval width of 0.25. Within each interval, calculate the weighted score of the average content dwell time and number of interactive operations for four types of content: basic science popularization, in-depth explanation, entertaining story, and professional academic content. In the formula: For the first The feedback effect score of a certain type of narrative content within a cognitive fatigue index range; the higher the score, the better the feedback effect. This is the normalized value of the average content dwell time for this type of content within this interval; This is the normalized value of the average number of interactive operations for this type of content within this interval; For the first Normalized value of the average number of repeated views of this type of content within a given interval; For the first The normalized value of the average content switching frequency of this type of content within a given interval.

[0134] The difference between the score of each content category in its respective interval and the preset benchmark score is used as the reward signal. The benchmark score is the average score of all content types in that interval. In the formula: For the first Reward signals for this type of narrative content within a given interval; For the first The average feedback score for all content types within each interval. The matching priority weights for various content types in the multi-level narrative resource library are adjusted based on reward signals, i.e.: In the formula: The adjusted content is in the [section / section]. The matching priority weight for each interval; Weights before adjustment; The weight adjustment ratio ranges from 5% to 20%, with 10% being preferred; sgn is the sign function, set to 1 for a positive reward and -1 for a negative reward. The final weights are then set... In the formula: The lowest effective priority weight is set to 0.1. Iterative optimization is performed every 24 to 72 hours, with 48 hours being the preferred time.

[0135] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention is limited to these examples; within the framework of the invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of the different aspects of the invention as described above, which are not provided in detail for the sake of brevity.

[0136] This invention is intended to cover all such substitutions, modifications, and variations that fall within the broad scope of this specification. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A personalized narrative adaptive push system for exhibition halls based on eye tracking and gait prediction, characterized in that, include: The dual-modal perception data fusion and tour state feature extraction module is used to simultaneously collect eye tracking and gait behavior data, and construct a cross-modal spatiotemporal coupled fingerprint of eye saccades-gait turning to generate a consistent identity identifier; Used to construct a visual-mobile intent conflict resolution matrix, output conflict resolution coefficients, and generate a consistent binding tour state feature vector; The cognitive-motor conflict analysis and sightseeing behavior prediction module is used to analyze the antagonism between eye movement and gait timing to obtain the visual-motor conflict index and identify the inflection point as the point of intent to stay. It is used to calculate the implicit interest transmission probability based on semantic association network to generate implicit interest transmission path, construct dynamic three-dimensional user profile with cognitive fatigue index, and generate behavior prediction results; The personalized narrative adaptive scheduling and closed-loop optimization module is used to adjust the playback start time based on the inflection point and conflict index. It is used to reorganize narrative logic based on implicit interest transmission paths and trigger cognitive deload mode when the cognitive fatigue index exceeds the threshold. Used to construct a group implicit interest transmission network for back-to-back iterative optimization of the prediction model.

2. The personalized narrative adaptive push system for exhibition halls based on eye tracking and gait prediction as described in claim 1, characterized in that, The dual-modal perception data fusion and tour state feature extraction module specifically includes: The data synchronization acquisition unit is used to synchronously acquire user eye tracking data and gait behavior data. The eye tracking data includes the spatial coordinates of the fixation point, the duration of fixation, the saccade trajectory, the change in pupil diameter, and the blink frequency. The gait behavior data includes instantaneous walking speed, cadence, walking direction angle, stride length, dwelling posture, turning trajectory, and start / stop status. The data preprocessing and spatiotemporal alignment unit is used to filter and denoise eye-tracking data and gait behavior data, remove invalid frames and unify timestamp alignment, and extract eye-tracking focus features, visual interest region features, gait rhythm features and walking trend features. A cross-modal spatiotemporal coupled fingerprint construction unit is used to analyze the coupling relationship between the scanning trajectory and the turning trajectory within a preset time delay range, extract the user's scanning-turning coupling mode features, and generate a user's cross-device consistent identity identifier. The intent conflict resolution unit is used to calculate the weights of the visual dominant intent and the mobile dominant intent based on historical behavior patterns and the current browsing state when the visual attention area and the mobile target area are inconsistent, and outputs the conflict resolution coefficient. The feature fusion encoding unit is used to weight and fuse eye-tracking focus features, visual interest region features, gait rhythm features, and walking trend features, and combine them with visual dominant intent weights, mobile dominant intent weights, and conflict resolution coefficients to encode and generate a user consistency binding tour state feature vector.

3. The personalized narrative adaptive push system for exhibition halls based on eye tracking and gait prediction according to claim 2, characterized in that, The specific working steps of the cross-modal spatiotemporal coupled fingerprint construction unit are as follows: Extract the scanning direction sequence and turning direction sequence within the current time window, and analyze the synchronicity and time delay relationship of the directional changes of the two sets of sequences within a preset time delay range to obtain the degree of coupling under different time delay offsets, i.e.: In the formula: Time delay offset The degree of cross-correlation coupling; To effectively scan the number of events; For the first step in the steering direction sequence Each sample value is in degrees. Time delay offset The corresponding sampling point offset integer; Extract the maximum value of the coupling degree and its corresponding time delay offset to form the feature vector of the cross-modal spatiotemporal coupled fingerprint; The cross-modal spatiotemporal coupled fingerprint is compared and matched with the user behavior template library, that is: In the formula: For the first The degree of matching of each template, This represents the total number of templates in the template library. For the first The average matching distance of each template; the user identity corresponding to the template with the highest matching degree is output as the cross-device consistent identity of the current user.

4. The personalized narrative adaptive push system for exhibition halls based on eye tracking and gait prediction according to claim 3, characterized in that, The specific working steps of the intent conflict resolution unit are as follows: Real-time acquisition of the user's eye gaze point coordinates and travel direction angle, which are then mapped to the exhibition hall's planar coordinate system to determine the visual attention area and the moving target area; Calculate the spatial distance between the visual attention area and the moving target area. ,Right now: In the formula: The coordinates of the center of the visual focus area. The coordinates of the center of the target area are used; when the spatial conflict threshold is exceeded, it is determined to be an intent conflict state, and the visual dominant intent weight is calculated based on the user's historical gaze duration and number of gazes on the current exhibit. ,Right now: In the formula: For the duration of historical observation, For the number of times historical attention, As a reference for normalized fixation duration, As a reference for normalized fixation counts, This is the fixation duration weighting coefficient. This is the weighting coefficient for the number of fixations; The weight of mobile dominant intent is calculated based on the user's historical dwell time and passage speed at the current exhibit. ,Right now: In the formula: For the length of time to stay in history, The average traffic speed, As a benchmark for normalized stay duration, For maximum traffic speed, This is the weighting coefficient for the duration of stay. The traffic speed weighting coefficient is used; the ratio of the visually dominant intent weighting to the mobile-dominant intent weighting is used as the conflict resolution coefficient. Output, i.e.: The dominant relationship is determined based on the comparison between the conflict resolution coefficient and the preset threshold.

5. The personalized narrative adaptive push system for exhibition halls based on eye tracking and gait prediction according to claim 4, characterized in that, The cognitive-motor conflict analysis and sightseeing behavior prediction module specifically includes: The time-series prediction unit is used to receive the user's consistent bound tour state feature vector and input it into a lightweight deep learning prediction model to predict the user's future travel position, the probability of focusing on target exhibits and the expected stay time within a preset time period. The antagonistic relationship analysis unit is used to perform antagonistic relationship analysis on eye movement time series data and gait time series data. It calculates the visual lock intensity index based on the time series standard deviation of the fixation point offset vector, extracts the time series coefficient of variation of instantaneous walking speed as the gait propulsion intensity index, calculates the ratio of the two as the visual-motor conflict index, and identifies the inflection point when it changes from rising to falling in order to predict the probability of stopping and the duration of dwell. The cognitive state perception unit is used to construct a cognitive intervention readiness index based on the pupil dilation frequency change rate and gait deceleration characteristics, and to predict the time window for the user to enter a deep cognitive intervention state. The interest transmission analysis unit is used to calculate the implicit interest transmission probability between related exhibits based on the semantic association network of exhibits in the exhibition hall, and to generate the target exhibit prediction sequence and implicit interest transmission path; The interest measurement and profile building unit is used to measure the user's interest in exhibits into three levels by combining gaze duration, pupil change, blink frequency, dwelling posture and walking speed. Based on the user's historical behavior data, interest preferences, browsing pace, content acceptance habits and cognitive intervention readiness index associated with the cross-device consistent identity, it constructs a dynamic three-dimensional user profile including preference tags, pace tags, cognitive level tags and cognitive fatigue index, and outputs behavior prediction results.

6. The personalized narrative adaptive push system for exhibition halls based on eye tracking and gait prediction according to claim 5, characterized in that, The specific working steps of the antagonistic relationship analysis unit are as follows: Using a preset length sliding time window as the analysis unit, the spatiotemporally aligned eye movement time series data and gait time series data are continuously extracted; Extract the offset vector of the gaze point spatial coordinates relative to the user's head position at each sampling moment within the current sliding time window, and calculate the time series standard deviation based on its spatial distribution characteristics as an indicator of visual lock-on strength. ,Right now: , In the formula: These are the time series standard deviations of the x- and y-direction components of the offset vector, respectively. The dispersion of the offset distribution, The standard deviation is used as a reference for the exhibition hall space. Instantaneous walking speed values ​​at each sampling time within the window are extracted synchronously. The coefficient of variation of the time series is calculated based on its variation characteristics as a gait propulsion intensity index. The visual-motor conflict index is calculated based on the visual lock-on intensity index and the gait propulsion intensity index. ,Right now: , In the formula: As an indicator of gait propulsion intensity, , These are the arithmetic mean and standard deviation of the instantaneous walking speed within the window, respectively; When both the visual lock intensity index and the gait propulsion intensity index exceed the preset high conflict threshold, the current window is judged as a high conflict state. The visual-motion conflict index of each time window is arranged in chronological order, and the inflection point when it changes from rising to falling is identified as the burst point of the user's intention to stay. Based on the time difference between the inflection point and the current time, the probability of the user stopping and the duration of stay within the preset time period in the future are predicted.

7. The personalized narrative adaptive push system for exhibition halls based on eye tracking and gait prediction according to claim 6, characterized in that, The specific working steps of the cognitive state perception unit are as follows: Pupil diameter measurements were extracted from spatiotemporally aligned eye-tracking data, and visual cognitive arousal indices were calculated based on the trend of pupil dilation and constriction frequency changes. ,Right now: In the formula: They are respectively and The normalized value, The rate of change of pupil dilation and constriction frequency. , , These represent the mean and standard deviation of the pupil diameter within the window, respectively. This represents the percentage change in pupil diameter. Instantaneous walking speed values ​​are extracted from spatiotemporally aligned gait time-series data, and gait deceleration indices are calculated based on walking speed decline characteristics. ,Right now: In the formula: The relative deceleration rate, , , These represent the average walking speed within the current sliding window and the baseline window, respectively. The low-speed retention ratio; The cognitive intervention readiness index is generated by integrating the visual cognitive arousal index and the gait deceleration index. ,Right now: In the formula: To integrate weights; based on the changing trend of the cognitive intervention readiness index over time, predict the time window for the user to enter a deep cognitive intervention state, and output the cognitive intervention readiness index and the start and end times of the time window as the cognitive state perception result.

8. The personalized narrative adaptive push system for exhibition halls based on eye tracking and gait prediction according to claim 7, characterized in that, The specific working steps of the interest transmission analysis unit are as follows: Read the network structure data from the semantic association network of exhibits in the exhibition hall, with each exhibit as a node, semantic association as an edge, and each edge having an initial weight. Load the target exhibit attention probability as the initial incentive to the node corresponding to the target exhibit. Starting with the target exhibit node, probability diffusion is performed along the semantic association network edges according to the initial weights, so that the interest probability is transmitted to adjacent nodes and indirect association nodes of all orders, that is: In the formula: For the first During the first iteration The probability value of each node. The damping coefficient is... To point to the first The set of all predecessor node indices of a node. For the node To the node edge weights, For the node to its neighboring nodes The edge weights are determined; the iteration continues until the probability distribution of the entire network converges. The probability values ​​of each node after convergence are used as the implicit interest transmission probability. The exhibits corresponding to the nodes whose probability values ​​exceed the preset transmission threshold are marked as implicitly associated exhibits, and together with the target exhibits, they form the target exhibit prediction sequence. The complete path from the target exhibit node to each implicitly related exhibit node along the probability diffusion path is recorded as the implicit interest transmission path, and the target exhibit prediction sequence and the implicit interest transmission path are output.

9. The personalized narrative adaptive push system for exhibition halls based on eye tracking and gait prediction according to claim 1, characterized in that, The personalized narrative adaptive scheduling and closed-loop optimization module specifically includes: The content matching and preloading unit is used to receive the behavior prediction result, determine the preloading priority based on the target exhibit prediction sequence, and when the dwell probability value exceeds the preset threshold, correct the pre-play start time based on the cognitive window period, the predicted travel position and the real-time spatial distance, start the preloading and pre-play of the current exhibit narrative content in the first preset time window after the inflection point, and preload the narrative content of implicitly related exhibits in the pre-visit interval based on the implicit interest transmission path. The playback strategy adaptive unit is used to match narrative content of corresponding length and depth based on predicted dwell time and interest level. It selects suitable narrative content types and display formats based on preference tags, rhythm tags and cognitive level tags in dynamic 3D user profile. It dynamically adjusts the start playback time and entry rhythm of the pushed content based on visual-motor conflict index. It dynamically adjusts the playback speed, switching speed and display duration based on gait rhythm, eye movement focus and cognitive fatigue index. When the cognitive fatigue index exceeds the preset threshold, it triggers cognitive deload mode. The narrative logic arrangement unit is used to dynamically adjust the narrative order and content association logic of exhibits according to the user's browsing trajectory and interest shift trend, and automatically connect the related knowledge points of adjacent exhibits according to the implicit interest transmission path to form a coherent personalized tour narrative chain. The multi-terminal display and feedback collection unit is used to output adapted narrative content to various terminal devices for display and to collect user content browsing feedback data in real time. The closed-loop iterative optimization unit is used to construct a group implicit interest transmission network, iteratively optimize the edge weights of the semantic association network of exhibits in the exhibition hall based on multi-user browsing feedback data, and iteratively optimize the lightweight deep learning prediction model and content matching strategy in reverse.

10. The personalized narrative adaptive push system for exhibition halls based on eye tracking and gait prediction according to claim 9, characterized in that, The specific working steps of the closed-loop iterative optimization unit are as follows: Real-time data collection of user browsing feedback on displayed content is conducted through various terminal devices, including content dwell time, number of interactive operations, content switching frequency, and number of repeated views. Using user browsing feedback data as a sample, we statistically analyzed the interest transfer patterns between exhibit pairs, namely: In the formula: To the exhibits To the exhibits The probability of interest shift, For positive transition frequency, For exhibits The total number of visitors; as the basis for incremental edge weights, the current edge weights of each edge in the semantic association network of exhibits in the exhibition hall are positively reinforced and updated to generate a group implicit interest transmission network; Based on the updated edge weights in the aforementioned latent interest transmission network, the interest transmission probability is recalculated, and the deviation between the recalculated latent interest transmission probability and the actual interest transfer pattern is used as the first loss signal. Fine-tuning and correcting the edge weights, where: In the formula: This represents the total number of nodes in the semantic association network of the exhibits in the exhibition hall. For exhibits recalculated based on the updated edge weights The probability of latent interest transmission. For exhibits The proportion of exhibits whose actual user dwell time exceeds 3 seconds out of the total number of exhibit dwell times; the mismatch between the probability of attention to a target exhibit and the actual target exhibit viewed by users is used as a second loss signal. Backpropagation is performed to the lightweight deep learning prediction model to update the model parameters, where: In the formula: The total number of exhibits in the exhibition hall. For exhibits The true target label, Exhibits output by the timing prediction unit The probability of attracting attention to the target exhibits; The user feedback effects of various narrative contents under different cognitive fatigue index ranges are statistically analyzed and used as reward signals for the content matching strategy. Based on the reward signals, the matching priority weights of various narrative contents in the multi-level narrative resource library are adjusted, and the content matching strategy is iteratively optimized.

Images