Tourist behavior dynamic modeling method based on space-time big data

Abstract

The invention relates to the technical field of intelligent tourism management, and discloses a tourist behavior dynamic modeling method based on space-time big data, which comprises the steps of collecting and preprocessing multi-source positioning data, inputting the multi-source positioning data into a multi-source positioning fusion engine, and combining statistical optimization and deep learning model fusion to obtain a high-precision positioning model. Track reconstruction and time sequence aggregation are carried out on continuous position points of tourists based on a high-precision positioning model, a structured activity data graph is generated, activity data of the tourists at all positions are obtained, activity preference characteristics are extracted through an association rule mining and clustering algorithm, and a Markov chain and a spatio-temporal evolution model are combined to predict a tourist flow trend. And forming a flow prediction result, and generating a visual dynamic decision support by using the prediction result. According to the invention, visual and dynamic decision support is provided, the crowd congestion is relieved, and the operation safety of the scenic area is guaranteed.

Description

Technical Field

[0001] This invention relates to the field of intelligent tourism management technology, and more specifically, to a method for dynamic modeling of tourist behavior based on spatiotemporal big data. Background Technology

[0002] With the development of smart tourism and smart cities, tourists' behavioral patterns, activity preferences, and flow trends in scenic spots and urban cultural and tourism environments have become important bases for tourism management and service optimization. Traditional tourist behavior research mainly relies on questionnaires, manual observation, or single data collection methods, which have limited data sources and long update cycles, making it difficult to truly and comprehensively reflect the dynamic behavioral characteristics of tourists.

[0003] For example, invention patent CN111556445A, a scenic area management system based on real-time dynamic location information feedback of tourists, while capable of real-time uploading and displaying of badge locations and performing basic data analysis in the cloud, still has the following shortcomings: The data collection method is singular; badges mainly rely on wireless communication modules to transmit data. If there are network coverage blind spots, signal interference, or high-concurrency access scenarios within the scenic area, data delays or loss can easily occur, leading to untimely updates of tourist location information and affecting the real-time performance of monitoring and management. It primarily focuses on displaying location information and simple route analysis, lacking in-depth modeling and predictive capabilities of tourist behavior patterns, making it difficult to provide higher-level decision support for scenic area management, such as tourist flow trend prediction and regional congestion risk warning. Although the cloud can statistically analyze popular attractions and main routes, its analytical dimensions are relatively singular, failing to combine time, environment, and tourist group characteristics for multi-dimensional data mining. The cloud-based tourist number prediction algorithm mainly relies on data from the previous k days, holidays, and weather factors, ignoring the impact of real-time flow data and unexpected events, resulting in limited prediction accuracy and difficulty in dealing with sudden increases or decreases in tourist numbers.

[0004] Therefore, it is necessary to design a dynamic modeling method for tourist behavior based on spatiotemporal big data to solve the problems existing in the current technology. Summary of the Invention

[0005] In view of this, the present invention proposes a dynamic modeling method for tourist behavior based on spatiotemporal big data, aiming to solve the problems of frequent crowd congestion, poor tourist experience, and low level of scenic area safety management.

[0006] This invention proposes a method for dynamic modeling of tourist behavior based on spatiotemporal big data, comprising: collecting multi-source positioning data, including GPS location data, BeiDou positioning data, Bluetooth beacon data, and AI video recognition data; and preprocessing the multi-source positioning data;

[0007] The preprocessed multi-source positioning data is input into the multi-source positioning fusion engine, and the location results are fused based on statistical optimization methods and deep learning models to obtain a high-precision positioning model.

[0008] Based on the high-precision positioning model, the continuous location data of tourists are reconstructed and time-series aggregated to generate an activity data map. The activity data map is a structured spatiotemporal behavioral data set, and the activity data of tourists at each location is obtained.

[0009] Based on association rule mining and clustering algorithms, tourist activity preference features are extracted from the activity data. Based on the activity preference features, Markov chains and spatiotemporal evolution models are used to predict tourist flow trends and generate flow prediction results. Based on the flow prediction results, visualized dynamic decision support is generated.

[0010] Furthermore, the preprocessing of the multi-source positioning data includes:

[0011] Data cleaning involves identifying noise points and abnormal jumps in the positioning data and applying filtering techniques to eliminate positioning deviations caused by signal interference and equipment errors.

[0012] Time synchronization corrects clock offsets between different data sources and generates continuous time series through interpolation methods, ensuring that the multi-source positioning data are strictly aligned in the time dimension.

[0013] Coordinate normalization maps location data from heterogeneous coordinate systems to a common geographic coordinate system, making all location points spatially comparable.

[0014] Furthermore, when fusing the location results based on statistical optimization methods and deep learning models to obtain a high-precision positioning model, the following steps are included:

[0015] The statistical optimization method is a dynamic weight allocation mechanism that adaptively adjusts the weight coefficients of each data source based on the positioning reliability of each data source in the real-time environment. The positioning reliability is determined by analyzing the stability of historical data and the current signal quality. The weight coefficients of the data sources are dynamically updated as the environment changes. When a decrease in the signal quality of a data source is detected, its weight coefficient is automatically lowered; when an increase in the signal quality of a data source is detected, its weight coefficient is automatically raised.

[0016] The deep learning model employs a spatiotemporal feature joint extraction architecture, comprising a temporal encoding module and a spatial association module. The temporal encoding module utilizes a recurrent neural network structure to process continuous positional point sequences, capturing the evolutionary patterns of positional changes over time. The spatial association module analyzes the spatial adjacency relationships between positional points, quantifying the strength of movement associations between different regions. The temporal encoding module outputs a temporal feature vector, and the spatial association module outputs a spatial feature vector, which is input into an attention mechanism layer. The attention mechanism layer calculates the feature contribution weights of each feature vector and performs a weighted summation of the feature vectors based on these weights to generate a fused feature vector. The fused feature vector is then converted into positional coordinates via a fully connected layer.

[0017] Furthermore, when fusing location results based on statistical optimization methods and deep learning models to obtain a high-precision positioning model, the following steps are also included:

[0018] The statistical optimization method establishes a two-way data channel with the deep learning model. The preliminary fusion result output by the statistical optimization is used as one of the input features of the high-precision positioning model. The corrected result output by the high-precision positioning model is fed back to the statistical optimization stage to update the weights. The two-way data channel continuously exchanges positions and fuses intermediate results to form an iterative optimization loop.

[0019] Furthermore, when reconstructing the trajectory and aggregating the temporal data of continuous tourist locations, the following steps are included:

[0020] The trajectory reconstruction includes a trajectory compression stage and a semantic annotation stage;

[0021] The trajectory compression stage is based on a geometric feature simplification algorithm to identify key turning points in the position coordinates and remove redundant position points to generate a simplified movement path; the identification of key turning points is based on the path curvature change rate and the degree of abrupt change in direction.

[0022] The semantic annotation stage analyzes the clustering characteristics of location points and the distribution of dwell time, divides the effective dwell area, and associates it with the functional area type of the scenic area.

[0023] The trajectory reconstruction employs a dynamic verification mechanism to calculate the spatial deviation between the reconstructed path and the original position sequence. When the spatial deviation exceeds a preset threshold, a path correction process is triggered. The path correction process readjusts the positions of key turning points. The positions of key turning points and the dwell areas together constitute the structured trajectory skeleton of the activity data map, serving as the input data for the temporal aggregation.

[0024] Furthermore, when reconstructing and time-series aggregating the continuous location data of tourists, the process also includes:

[0025] The temporal aggregation includes spatiotemporal cube construction and behavioral feature quantification; the scenic area space is divided into uniform grid units, and the structured trajectory skeleton is aggregated in combination with discrete time windows to generate a three-dimensional spatiotemporal cube structure; the distribution of dwell time, visit frequency density and behavioral sequence patterns are statistically analyzed within the three-dimensional spatiotemporal cube structure; the periodic patterns and sudden change trends of tourist behavior are identified through cross-time window correlation analysis;

[0026] The temporal aggregation implements a dynamic granularity adjustment mechanism, which automatically adjusts the size of the spatial grid and the length of the time window according to the density of pedestrian traffic in the area; fine spatial grid division and short time window are used in densely populated areas, while coarse spatial grid division and long time window are used in sparsely populated areas.

[0027] Furthermore, when extracting tourists' activity preference features based on association rule mining and clustering algorithms and the activity data, the process includes:

[0028] Support and confidence thresholds are set to filter behavioral sequences. Association rules between tourist activities are mined based on the activity data. A time decay mechanism is introduced to assign dynamic weight coefficients to behavioral records. The time evolution trajectory of the activity sequence is analyzed to identify high-frequency co-occurrence patterns. False association rules are filtered in combination with environmental context information. The mining results are dynamically pruned to retain a subset of effective rules. High-frequency behavioral sequences are used as the initial seed point set for the clustering algorithm.

[0029] Furthermore, when extracting tourists' activity preference features based on association rule mining and clustering algorithms and the activity data, the method also includes:

[0030] The initial seed point set is divided into behavioral clusters based on its spatiotemporal aggregation characteristics. The silhouette coefficient is calculated to evaluate the quality of the clustering results. The clustering parameters are iteratively adjusted to optimize the cluster boundary division. The high-frequency behavioral sequences output by association rule mining are fused as constraints for cluster boundary adjustment. The clustering results are output, which include cluster center behavioral patterns, cluster dynamic evolution trajectories, and cluster feature vectors. The clustering results serve as the activity preference features.

[0031] Furthermore, based on the aforementioned activity preference characteristics, when using a Markov chain and spatiotemporal evolution model to predict tourist flow trends, the following steps are included:

[0032] The spatiotemporal evolution model integrates the state transition mechanism of the Markov chain with the dynamic response capability of the environment; it constructs a state space for the functional areas of the scenic area and defines the transfer probability matrix of tourists between areas; it introduces environmental factors to dynamically correct the transfer probability, including real-time visitor flow, area openness status, and the influence of external events; the transfer probability of the Markov chain is obtained through learning from historical behavioral data; the transfer probability matrix is ​​updated in real time in combination with the activity preference features; the spatiotemporal evolution model fine-tunes the transfer probability every time it receives new behavioral data; the prediction process generates a probability path for the distribution of tourists in future time periods, and the probability path serves as the flow prediction result.

[0033] Furthermore, when generating visualized dynamic decision support based on the flow prediction results, the process includes:

[0034] The dynamic decision support includes rendering a heat map of tourist density distribution and generating resource scheduling suggestions; a heat map of tourist density distribution is constructed based on the flow prediction results, and the degree of regional congestion is presented through color gradients;

[0035] The bottleneck areas exceeding the density threshold in the tourist density distribution heatmap are identified, and the dynamic decision support is generated in combination with the layout of scenic area facilities; the dynamic decision support includes diversion path planning schemes, service facility expansion schemes, and emergency response strategies.

[0036] Compared with existing technologies, the advantages of this invention are as follows: By integrating multi-source positioning data such as GPS, BeiDou, Bluetooth beacons, and AI video recognition, and combining a statistical optimization method with dynamic weight allocation and the spatiotemporal feature extraction capabilities of deep learning models, the invention achieves complementarity and optimization of different data sources in complex scenic environments, making tourist location estimation more accurate and robust, and avoiding the problem of accumulated bias caused by errors from a single data source. Through a dynamic weight adjustment mechanism and iterative optimization loop, the positioning fusion process can respond in real time to environmental changes and signal fluctuations, and can still output continuous and stable location data even when signals are lost or degraded, improving its applicability in complex terrain and high-density crowd scenarios. The trajectory reconstruction method not only simplifies trajectories and identifies regional functions through compression and semantic annotation, but also ensures the accuracy of trajectory reconstruction through dynamic verification and correction, thereby generating a structured activity data map, providing a solid foundation for subsequent behavior analysis and modeling. Through a spatiotemporal cube structure and a dynamic granular aggregation mechanism, the invention performs multi-dimensional statistics and modeling of tourist stay duration, visit frequency, and behavior patterns. Combined with association rule mining and clustering algorithms, it extracts tourist activity preference features, achieving refined expression of tourist behavior patterns and group differentiation identification. A predictive framework combining Markov chains and spatiotemporal evolution models not only characterizes the probability of tourist movement between functional areas but also incorporates environmental factors for dynamic correction in real time. This allows the prediction results to reflect the temporal characteristics of tourist flow and environmental responses, making them more consistent with real-world scenarios. By generating heatmaps of tourist density distribution and identifying bottleneck areas, and combining this with the layout of scenic area resources, the framework provides diversion schemes, facility upgrades, and emergency strategies. This offers scenic area managers visualized and dynamic decision support, alleviating congestion, improving the tourist experience, and ensuring the safety of scenic area operations. Attached Figure Description

[0037] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0038] Figure 1 A flowchart illustrating a method for dynamic modeling of tourist behavior based on spatiotemporal big data, as provided in an embodiment of the present invention.

[0039] Figure 2 The diagram shows the structure of a deep learning model for a dynamic modeling method of tourist behavior based on spatiotemporal big data, as provided in this embodiment of the invention. Detailed Implementation

[0040] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey its scope to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features described herein can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0041] For this, please refer to Figure 1 As shown, a dynamic modeling method for tourist behavior based on spatiotemporal big data is proposed, including:

[0042] Collect multi-source positioning data, including GPS location data, BeiDou positioning data, Bluetooth beacon data, and AI video recognition data; and preprocess the multi-source positioning data.

[0043] The preprocessed multi-source positioning data is input into the multi-source positioning fusion engine, and the location results are fused based on statistical optimization methods and deep learning models to obtain a high-precision positioning model.

[0044] Based on the high-precision positioning model, the continuous location data of tourists are reconstructed and time-series aggregated to generate an activity data map. The activity data map is a structured spatiotemporal behavioral data set, and the activity data of tourists at each location is obtained.

[0045] Based on association rule mining and clustering algorithms, tourist activity preference features are extracted from the activity data. Based on the activity preference features, Markov chains and spatiotemporal evolution models are used to predict tourist flow trends and generate flow prediction results. Based on the flow prediction results, visualized dynamic decision support is generated.

[0046] The system utilizes GPS, BeiDou, Bluetooth beacon, and AI video recognition data to address the issue of insufficient accuracy or signal loss in complex environments using single positioning methods, achieving 24 / 7 coverage of tourist locations. In the data preprocessing stage, the collected multi-source positioning data undergoes data cleaning, time synchronization, and coordinate normalization. After inputting the preprocessed data into the multi-source fusion engine, a joint modeling approach combining statistical optimization and deep learning is employed. The statistical optimization part adjusts weights based on real-time signal quality using a dynamic weight allocation mechanism. Initially, GPS and BeiDou each have a weight of 0.4, while Bluetooth and video each have a weight of 0.1. If the signal is obstructed, the weight of that source is automatically reduced. The deep learning part extracts spatiotemporal features and calculates feature contributions using an attention mechanism, ultimately obtaining high-precision positioning results with an error range of 1 to 3 meters. When performing trajectory reconstruction and temporal aggregation based on these results, a curvature threshold of 15° is set for trajectory compression to identify turning points and stops. The minimum dwell time for area identification is set at 120 seconds to ensure that the trajectory is both concise and retains behavioral characteristics. In the aggregation stage, the scenic area is divided into dynamic grids ranging from 10 to 50 meters, and a three-dimensional spatiotemporal cube is constructed by combining time windows of 5 to 30 minutes to statistically analyze dwell time, visit frequency, and behavioral sequence patterns. High-frequency combinations of tourist behaviors are identified by rule mining with a support threshold of 0.2 and a confidence threshold of 0.6, and preference features are formed by combining cluster analysis. In the trend prediction stage, Markov chains are used to model the state transitions between regions. The transition matrix is ​​learned from historical data of the past 30 days and updated every 5 minutes. At the same time, factors such as weather, holidays, and facility openings are introduced to dynamically adjust the transition probability. The output flow prediction results are visualized in the form of heat maps and scheduling suggestions to help managers identify bottleneck areas in advance and optimize resource allocation, thereby achieving a dual improvement in tourist experience and scenic area management efficiency.

[0047] The initial weights for the data sources are set to 0.4 for GPS and BeiDou, and 0.1 each for Bluetooth and video. When a signal quality degradation is detected, the weights are dynamically adjusted based on real-time signal indicators. The update follows the exponential smoothing formula wi(t+1)=αwi(t)+(1-α)Ri(t), where α is the smoothing factor and Ri(t) is the normalized signal quality. The curvature threshold for the trajectory compression stage is set to 15° by default. This is a fixed value in normal walking scenarios, but it is adaptively adjusted when the visitor's movement speed is too fast or too slow. For example, it is widened to 20° when the speed is less than 0.5m / s and narrowed to 10° when the speed is greater than 1.5m / s. The minimum dwell time for identifying the dwell area is set to 120 seconds. This value is a fixed empirical threshold to ensure that interference from short pauses can be eliminated. In spatiotemporal aggregation, the grid size and time window employ an adaptive mechanism, with a basic range of 10 to 50 meters and 5 to 30 minutes, respectively. Specific values ​​are dynamically adjusted based on real-time pedestrian density: high-density areas are divided into finer grids and use shorter time windows, while low-density areas are the opposite. In association rule mining, the support threshold of 0.2 and the confidence threshold of 0.6 are settable fixed parameters. The transition matrix in the Markov chain prediction phase is updated every 5 minutes; this parameter is a fixed value, while the transition probability is dynamically adjusted during the prediction process based on environmental factors such as weather, holidays, and facility opening status.

[0048] Specifically, tourist locations are collected using multi-source data including GPS, BeiDou, Bluetooth beacons, and AI video recognition. This complementary data coverage extends to indoor and outdoor environments, as well as obstructed areas, enabling continuous and highly reliable location information acquisition. The preprocessing stage includes data cleaning, time synchronization, and coordinate normalization. Outliers in the GPS and BeiDou data with position changes exceeding 10 meters per second are removed, time synchronization deviations are controlled to the millisecond level, and all locations are uniformly mapped to the WGS84 coordinate system to ensure comparability and consistency across different data sources. The preprocessed data is then input into a multi-source fusion engine, where location results are fused using statistical optimization methods and deep learning models. The statistical optimization section employs a dynamic weight allocation mechanism, with initial weights set at GPS 0.4, BeiDou 0.4, Bluetooth beacon 0.1, and AI video 0.1, automatically adjusting weights based on real-time signal quality. The deep learning model utilizes a spatiotemporal joint feature extraction architecture. The temporal coding module uses a recurrent neural network to capture the temporal changes in location, while the spatial association module analyzes movement relationships between different areas and calculates feature contribution weights through an attention mechanism. The fused feature vector is then passed through a fully connected layer to output high-precision location coordinates, achieving an accuracy of up to 13 meters. Based on the high-precision positioning results, trajectory reconstruction and temporal aggregation are performed on the continuous locations of tourists: trajectory compression identifies key turning points using a curvature threshold of 15° and eliminates redundant points; the marking of stopping areas uses a minimum stay time of 120 seconds as a threshold, mapping the trajectory to functional areas of the scenic area to construct an activity data map; in the temporal aggregation stage, the scenic area is divided into a 1050-meter dynamic grid, and stay duration, visit frequency, and behavioral sequence patterns are statistically analyzed using time windows of 5–30 minutes. Simultaneously, the grid and window granularity are adjusted to adapt to changes in visitor density. Subsequently, association rule mining with a support threshold of 0.2 and a confidence threshold of 0.6, along with spatiotemporal clustering analysis, were used to extract tourist activity preference features. Based on these features, Markov chains and spatiotemporal evolution models were employed to predict tourist flow trends: a transition probability matrix was learned from historical data and updated every 5 minutes, dynamically adjusted in conjunction with real-time environmental factors such as visitor volume, scenic area opening status, and weather. Finally, the generated flow prediction results were visualized using heatmaps to display regional congestion levels, and suggestions for diversion route planning, facility expansion, and emergency response were output, enabling dynamic management of tourist behavior and optimization of scenic area resources.

[0049] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0050] In practical applications, multi-source positioning technology can be used to collect visitor location data in real time: GPS and BeiDou provide outdoor location information, Bluetooth beacons cover indoor exhibition halls, and AI video recognition monitors the flow of people at entrances and key nodes. Assuming a visitor enters the park at 9:00 AM, the preprocessing stage removes GPS positioning noise (points with sudden speed changes exceeding 3 m / s), timestamps are synchronized via NTP, and coordinates are uniformly mapped to WGS84. After processing by the fusion engine, the initial weights of the statistical optimization method are set to GPS 0.4, BeiDou 0.4, Bluetooth beacon 0.1, and AI video 0.1. These weights are dynamically adjusted based on real-time signal strength and HDOP accuracy. The deep learning model uses an RNN to capture the movement trend of visitors from the entrance to the roller coaster area, and the spatial association module identifies the spatial adjacency relationships between the amusement area, dining area, and rest area, generating a high-precision location sequence. Subsequently, the trajectory reconstruction phase identified key turning points (curvature change threshold of 15°) and dwell points (dwell time exceeding 120 seconds), and combined this with the labeling of functional areas within the park to form a structured trajectory skeleton. In the temporal aggregation phase, the park was divided into 30-meter grids and 15-minute time windows, and visitor dwell time, visit frequency, and behavioral patterns in the roller coaster area were statistically analyzed. Based on the activity data, association rule mining was performed with a support level of 0.2 and a confidence level of 0.6 to extract visitor preferences, such as high-frequency ride access patterns. This, combined with real-time visitor flow, area opening status, and weather information, used Markov chains and spatiotemporal evolution models to predict visitor flow trends over the next 30 minutes. A heat map was generated to show potential congestion in the roller coaster area, and suggestions for diversion routes and increased food and beverage area capacity were provided, enabling managers to allocate resources in advance to ensure visitor experience and park safety.

[0051] Through the above-described scheme, this application obtains high-quality, continuous tourist location information by collecting multi-source positioning data from GPS, BeiDou, Bluetooth beacons, and AI video recognition, and preprocessing the data. The preprocessed data is then input into a multi-source positioning fusion engine, where statistical optimization methods and deep learning models are combined to generate a high-precision positioning model with high accuracy and robustness, reducing positioning errors and dynamically adapting to changes in signal quality from different data sources. Trajectory reconstruction and temporal aggregation of continuous tourist locations generate structured activity data maps, reflecting tourists' dwell time and movement paths at various attractions, providing a reliable foundation for subsequent behavioral analysis. Tourist activity preference features are extracted through association rule mining and clustering algorithms, revealing high-frequency visit patterns and preferred areas, providing data support for personalized services and operational optimization. Markov chains and spatiotemporal evolution models are used to predict tourist flow trends, generating future tourist distribution probabilities in real time, and visualizing the results as dynamic decision support.

[0052] This application further proposes preprocessing the multi-source positioning data, including:

[0053] Data cleaning involves identifying noise points and abnormal jumps in the positioning data and applying filtering techniques to eliminate positioning deviations caused by signal interference and equipment errors.

[0054] Time synchronization corrects clock offsets between different data sources and generates continuous time series through interpolation methods, ensuring that the multi-source positioning data are strictly aligned in the time dimension.

[0055] Coordinate normalization maps location data from heterogeneous coordinate systems to a common geographic coordinate system, making all location points spatially comparable.

[0056] The process involves several key steps: Data cleaning removes noise and anomalous jumps caused by signal interference, obstruction, multipath effects, or equipment errors, ensuring the reliability of location information used in subsequent analyses. Smoothing based on Kalman filtering or moving average filtering marks and removes points with sudden speed or positional changes exceeding preset thresholds (e.g., instantaneous movement speed exceeding 5 m / s is considered anomaly). Time synchronization strictly aligns data from different devices in the time dimension, preventing trajectory misalignment caused by differences in device clocks or sampling frequencies. Timestamps from each data source are corrected using Network Time Protocol (NTP) or GPS time signals, and then continuous time series are generated through linear interpolation or spline interpolation. The preset time step can be 1 second or 2 seconds to ensure trajectory continuity. Coordinate normalization maps location information from different sources and coordinate systems (such as WGS-84, GCJ-02, or the local plane coordinate system) to a common geographic coordinate system, making each location point spatially comparable and superimposed. During the normalization process, affine transformation or projection transformation methods can be used, and reference points or benchmark matrices can be set for correction, thereby ensuring the accuracy of subsequent multi-source fusion and trajectory analysis.

[0057] Specifically, in the preprocessing of multi-source positioning data, the first step is data cleaning to remove noise points and anomalous jumps, ensuring data accuracy and reliability. The velocity, acceleration, and direction changes of consecutive location points are calculated. When a velocity exceeds a preset threshold (instantaneous velocity greater than 5 m / s) or a direction change exceeds a threshold (e.g., above 60°), the point is marked as an anomaly and smoothed or replaced using methods such as Kalman filtering, weighted moving average, or median filtering. Next, time synchronization is performed to eliminate timestamp discrepancies between different data sources, ensuring strict alignment of the data in the time dimension. The time of each data source is first calibrated using GPS or Network Time Protocol (NTP), and then a continuous time series is generated through linear interpolation or cubic spline interpolation. The preset interpolation interval is generally 1 second to balance data accuracy and computational efficiency. Finally, coordinate normalization is performed to map data from different coordinate systems (such as WGS-84, GCJ-02, or the local plane coordinate system) to a common geographic coordinate system, ensuring spatial comparability. The normalization process can use affine transformation or projection transformation methods, and adjust it with the scenic area entrance or preset benchmark as a reference, so that all location points correspond accurately in space.

[0058] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0059] In practical applications, GPS, BeiDou, Bluetooth beacons, and AI video recognition cameras installed in key passageways are deployed simultaneously. During a visit, GPS positioning may be affected by obstructions from tall buildings or tunnel environments, Bluetooth beacons may experience signal interference in dense crowds, and video recognition data may suffer from frame loss or recognition errors. During the data cleaning phase, a velocity threshold of 3 m / s and an acceleration threshold of 2 m / s² are set. 2 When consecutive location points exceed this range, they are automatically marked as outliers and smoothed using Kalman filtering. During the time synchronization phase, the clocks of all devices are first synchronized with the Network Time Protocol (NTP) server. Then, a continuous time series is generated using linear interpolation in 1-second increments to ensure strict temporal alignment of GPS, BeiDou, Bluetooth, and video data. In the coordinate normalization phase, GPS (WGS-84), BeiDou (CGCS2000), and Bluetooth planar coordinate data are uniformly mapped to the unified coordinate system of the scenic area's geographic information system through affine transformation. The offset is adjusted based on the reference point of the entrance plaza to ensure precise spatial correspondence among all location points.

[0060] Through the above technical solutions, this application eliminates noise points and abnormal jumps by cleaning data, which not only reduces false trajectories introduced by signal interference or equipment errors, but also improves trajectory smoothness, making the tourist movement path more realistic and credible. Time synchronization ensures that heterogeneous data sources such as GPS, Beidou, Bluetooth beacons and AI video recognition are strictly aligned in the time dimension, making it possible to compare and analyze the behavior of cross-source data at the same point in time, avoiding misidentification of activity patterns due to time misalignment. Coordinate normalization realizes the unification of heterogeneous coordinate systems, making all location points spatially comparable, thereby ensuring the accuracy of subsequent trajectory reconstruction, temporal aggregation and activity preference analysis.

[0061] In some of the solutions described above in this application, during the multi-source positioning data fusion process, the positioning accuracy may decrease due to the instability or abnormal fluctuations of signals from certain data sources.

[0062] This application further proposes a method for fusing location results based on statistical optimization methods and deep learning models to obtain a high-precision positioning model, including:

[0063] The statistical optimization method is a dynamic weight allocation mechanism that adaptively adjusts the weight coefficients of each data source based on the positioning reliability of each data source in the real-time environment. The positioning reliability is determined by analyzing the stability of historical data and the current signal quality. The weight coefficients of the data sources are dynamically updated as the environment changes. When a decrease in the signal quality of a data source is detected, its weight coefficient is automatically lowered; when an increase in the signal quality of a data source is detected, its weight coefficient is automatically raised.

[0064] The deep learning model employs a spatiotemporal feature joint extraction architecture, comprising a temporal encoding module and a spatial association module. The temporal encoding module utilizes a recurrent neural network structure to process continuous positional point sequences, capturing the evolutionary patterns of positional changes over time. The spatial association module analyzes the spatial adjacency relationships between positional points, quantifying the strength of movement associations between different regions. The temporal encoding module outputs a temporal feature vector, and the spatial association module outputs a spatial feature vector, which is input into an attention mechanism layer. The attention mechanism layer calculates the feature contribution weights of each feature vector and performs a weighted summation of the feature vectors based on these weights to generate a fused feature vector. The fused feature vector is then converted into positional coordinates via a fully connected layer.

[0065] The statistical optimization method employs a dynamic weight allocation mechanism to automatically adjust the weight coefficients of each data source based on its real-time positioning reliability. This ensures that data sources with high noise or unstable signals do not significantly impact the final positioning result. Positioning reliability is assessed through a comprehensive evaluation of historical data stability and current signal quality. Weights are automatically lowered when signal quality declines and increased when signal quality improves, achieving real-time adaptive optimization. The spatiotemporal feature joint extraction architecture of the deep learning model consists of a temporal encoding module and a spatial association module. The temporal encoding module uses a recurrent neural network to process continuous location sequence, capturing the patterns of tourist movement over time. The spatial association module analyzes the spatial adjacency relationships between location points, quantifying the strength of movement associations between different areas to ensure full utilization of spatial information. The temporal feature vectors output by the temporal encoding module and the spatial feature vectors output by the spatial association module are input into an attention mechanism layer. The attention mechanism automatically assigns weights based on the contribution of each feature vector to the final positioning result, generating a fused feature vector through weighted summation to highlight the most informative features, thereby further improving positioning accuracy. The fused feature vector is then converted into final location coordinates through a fully connected layer, achieving high-precision positioning output. The entire process effectively reduces errors from a single data source, balances the contributions of various heterogeneous data, and makes full use of temporal and spatial feature information through the joint collaboration of statistical optimization and deep learning. This ensures that the final positioning results are both stable and accurate, and is especially suitable for dynamic crowd positioning and trajectory analysis in complex scenarios.

[0066] Specifically, the statistical optimization method employs a dynamic weight allocation mechanism to evaluate and adjust the positioning reliability of various data sources in real-time environments. It analyzes historical data stability and current signal quality to determine the reliability index of each data source. For GPS signals, the weight is set to 0.7 in clear conditions, automatically decreasing to 0.4 in environments with multiple buildings obstructing the view. BeiDou signals are assigned a weight of 0.6 when the signal is strong and adjusted to 0.3 when the signal is weak. Bluetooth beacon data automatically adjusts its weight coefficient based on the Received Signal Strength Indicator (RSSI), typically set between 0.2 and 0.5. The weight for AI video recognition positioning is adjusted based on the target detection confidence level: 0.6 when the confidence level exceeds 0.8, and automatically adjusted to 0.3 when it falls below 0.5. This dynamic weight allocation mechanism adaptively reduces the impact of low-reliability data sources in situations of signal fluctuations or complex environments, while enhancing the contribution of high-reliability data sources, thereby reducing error accumulation. The deep learning model further optimizes positioning accuracy through a spatiotemporal feature joint extraction architecture. The temporal coding module utilizes a recurrent neural network (RNN) to process continuous location sequence, capturing the patterns of tourist movement trajectories over time. In cases of abrupt changes in path curvature or abnormal dwell times, the RNN can identify and adjust prediction biases. The spatial association module analyzes the spatial adjacency relationships between location points, quantifying the strength of movement associations between different areas, including the probability of tourist flow from the scenic area entrance to popular attractions, path selection preferences, and dwell density within areas. The temporal coding module outputs a temporal feature vector, while the spatial association module outputs a spatial feature vector. These vectors are input into the attention mechanism layer. The attention mechanism calculates the contribution weight of each feature vector to the final positioning result, performs a weighted sum based on these weights, and forms a fused feature vector, ensuring that important spatiotemporal information is given priority. The fused feature vector is then converted into precise location coordinates through a fully connected layer, achieving high-precision positioning output.

[0067] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0068] In practical applications, GPS, BeiDou, Bluetooth beacon, and AI video recognition data are collected simultaneously. GPS signals are weak in some areas due to obstruction by tall buildings or trees, while Bluetooth beacon coverage is more reliable in densely populated entrance areas. Video recognition can supplement positioning information in main channels. The statistical optimization method dynamically allocates weights based on historical data and current signal quality: for example, when the GPS signal strength is below 30dB, its weight is automatically reduced from 0.7 to 0.4, while the weight is increased to 0.6 when the Bluetooth beacon signal RSSI is greater than -65dBm; the weight is set to 0.7 when the video recognition confidence level is above 0.85. The temporal coding module in the deep learning model inputs continuous location points of tourists from the entrance to the roller coaster queue area into a recurrent neural network to capture patterns in tourist movement speed changes and abnormal dwell times. When abnormal increases in tourist dwell time in the queue area are detected, the model automatically adjusts the predicted path. The spatial association module analyzes the probability of tourist movement between different scenic areas, showing a higher frequency of movement from the children's area to the dining area, quantifying the strength of movement associations between different areas. Temporal and spatial feature vectors are weighted using an attention mechanism. During peak lunch hours, the spatial feature weight may be increased to 0.65 to emphasize traffic changes in hotspot areas. These are then fused to generate the final feature vector, which is converted into high-precision location coordinates via a fully connected layer. This allows for real-time output of the precise location of each visitor, maintaining an error of less than 2 meters even in areas with weak or obstructed signals. This achieves high-precision, continuous, and reliable real-time positioning, providing accurate data support for visitor traffic monitoring, flow control, and safety management.

[0069] Through the aforementioned technical solution, this application utilizes a dynamic weight allocation mechanism to adjust weights in real time based on the signal quality and historical stability of each data source, automatically reducing the impact of low-quality data sources while enhancing the contribution of high-reliability data sources. This reduces the interference of positioning errors and abnormal jumps on the overall results. The deep learning model captures the evolution of tourist positions over time through a temporal coding module and analyzes the strength of movement associations between different areas using a spatial correlation module. This not only accurately determines single-point locations but also understands continuous trajectories and regional behavioral patterns. The attention mechanism further optimizes feature fusion, assigning higher weights to spatiotemporal features at critical moments, ensuring the model maintains high accuracy in positioning even in complex environments.

[0070] This application further proposes that when fusing location results based on statistical optimization methods and deep learning models to obtain a high-precision positioning model, it also includes:

[0071] The statistical optimization method establishes a two-way data channel with the deep learning model. The preliminary fusion result output by the statistical optimization is used as one of the input features of the high-precision positioning model. The corrected result output by the high-precision positioning model is fed back to the statistical optimization stage to update the weights. The two-way data channel continuously exchanges positions and fuses intermediate results to form an iterative optimization loop.

[0072] The statistical optimization method generates preliminary fused location results based on the real-time reliability of each data source. These preliminary results not only provide initial weighted judgments of multi-source data but also serve as input features for the deep learning model, helping the model fully utilize the spatial and temporal information of each data source during training and inference. After processing the input features, the deep learning model outputs corrected, high-precision location results. These results are then fed back to the statistical optimization module to dynamically update the weight coefficients of each data source, making the weight allocation more aligned with real-time environmental changes. Through continuous information exchange via bidirectional data channels, an iterative optimization loop is formed. The statistical optimization and deep learning model serve as each other's input and output, mutually correcting deviations and gradually converging to a more accurate positioning result.

[0073] At each time step, a statistical optimization method performs preliminary fusion of multi-source positioning data using the current data source weights to obtain a preliminary location estimate and its uncertainty. This result, along with the signal quality indicators, residual characteristics, and short-term statistical features of each data source, is input into a deep learning model. The deep learning model outputs a location correction and uncertainty estimate based on temporal coding and spatial correlation analysis. The correction is added to the preliminary estimate to obtain the corrected location result, which is then weighted and fused with the statistical results to generate the final location output. Simultaneously, the deviation between the corrected location and the statistical results is fed back to the statistical optimization module to update the weight coefficients of each data source: when a data source is detected to have an increased residual or a Mahalanobis distance exceeding a set threshold within the current window, its weight is automatically reduced; conversely, when its signal quality and stability improve, its weight is gradually increased. The entire process is repeated within a set number of iterations until the change in the location estimate is lower than a preset convergence threshold, thereby achieving bidirectional linkage correction between statistical optimization and the deep learning model, ensuring the accuracy and stability of the final positioning result.

[0074] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0075] In practical applications, the statistical optimization method initially calculates tourist locations based on the real-time reliability of each data source. When GPS signals weaken due to obstruction in a certain forest area, its weight automatically decreases, while the weights of Bluetooth beacon and video recognition data are correspondingly increased. This preliminary result is input into a deep learning model. The model analyzes the continuous sequence of tourist locations through a temporal coding module and calculates the probability of tourists moving between different areas of the scenic area using a spatial correlation module. Simultaneously, an attention mechanism calculates the importance of each feature, generating a corrected location result. This corrected result is then fed back into the statistical optimization stage to adjust the weights of each data source. The GPS weight is readjusted to reflect signal recovery. Through iterative optimization loops via bidirectional data channels, the final high-precision positioning result accurately reflects the actual trajectory of tourists within the scenic area. Even in environments with complex signals or partial obstruction, it maintains positioning accuracy and continuity, providing a reliable foundation for subsequent tourist behavior analysis and traffic prediction.

[0076] Through the above technical solution, this application establishes a bidirectional data channel between the statistical optimization method and the deep learning model, enabling iterative optimization loops. This allows the initial fused positioning results to be continuously corrected by the deep learning model and fed back to update the statistical optimization weights, thereby improving positioning accuracy and stability. This mechanism can effectively address signal fluctuations, occlusion interference, and abnormal jumps in multi-source data, making the final generated high-precision location coordinates more reliable and continuous.

[0077] This application further proposes methods for trajectory reconstruction and temporal aggregation of continuous location data of tourists, including:

[0078] The trajectory reconstruction includes a trajectory compression stage and a semantic annotation stage;

[0079] The trajectory compression stage is based on a geometric feature simplification algorithm to identify key turning points in the position coordinates and remove redundant position points to generate a simplified movement path; the identification of key turning points is based on the path curvature change rate and the degree of abrupt change in direction.

[0080] The semantic annotation stage analyzes the clustering characteristics of location points and the distribution of dwell time, divides the effective dwell area, and associates it with the functional area type of the scenic area.

[0081] The trajectory reconstruction implements a dynamic verification mechanism, which calculates the spatial deviation between the reconstructed path and the original position sequence. When the spatial deviation exceeds a preset threshold, a path correction process is triggered. The path correction process readjusts the positions of key turning points. The positions of key turning points and the dwell areas together constitute the structured trajectory skeleton of the activity data map, which serves as the input data for the temporal aggregation.

[0082] In the trajectory compression stage, a geometric feature simplification algorithm is applied to process the original position sequence. This algorithm identifies key turning points based on the rate of change of path curvature and the degree of abrupt changes in direction. When a tourist's path shows a directional change greater than 30° or a curvature change exceeding a preset threshold (e.g., 0.05 / m), it is marked as a key point, and other redundant position points are removed, thus generating a simplified movement path that retains the main movement trajectory while reducing the amount of data. Subsequently, in the semantic annotation stage, the clustering characteristics and dwell time of position points are analyzed to divide effective dwell areas. Tourists who stay in the main square of the scenic area for more than 5 minutes are marked as "sightseeing dwell area," and those who stay in the catering area for more than 10 minutes are marked as "dining area," and these are associated with the functional type of the scenic area. The trajectory reconstruction also introduces a dynamic verification mechanism, which calculates the spatial deviation between the simplified path and the original position sequence in real time. When the deviation exceeds a preset threshold (e.g., 5 meters), a path correction process is triggered to readjust the positions of key turning points, ensuring that the reconstructed trajectory is consistent with the actual movement height. Ultimately, key turning points and areas of stay form a structured trajectory skeleton, providing accurate input for time-series aggregation and generating precise activity data maps for analyzing tourist behavior patterns, stay preferences, and flow patterns.

[0083] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0084] In practical applications, tourists move along multiple routes after entering from the entrance, and their continuous GPS and Bluetooth beacon data are processed for trajectory compression. Using a geometric feature simplification algorithm, the algorithm identifies key turning points for tourists at locations such as observation decks, roller coaster entrances, and dining areas, while eliminating redundant points generated when tourists walk along straight paths, generating simplified movement paths. In the semantic annotation stage, the time tourists spend in these areas is analyzed. It was found that tourists spend longer periods at popular attractions such as roller coasters, and shorter periods in pedestrian walkways, thus labeling these points as "sightseeing areas," "rest areas," or "dining areas." During trajectory reconstruction, a dynamic verification mechanism continuously calculates the spatial deviation between the simplified trajectory and the original position sequence. When the deviation exceeds 5 meters, the positions of key turning points are automatically adjusted to ensure accuracy.

[0085] Through the above technical solutions, the simplified movement path generated after removing redundant location points in the trajectory compression stage not only reduces data storage and computation costs but also avoids analysis errors caused by continuous data noise. The semantic annotation stage associates the stopping areas with the scenic area's functional type, ensuring that each trajectory not only contains spatial location information but also reflects tourist behavior characteristics, thus forming an interpretable, structured trajectory skeleton. The dynamic verification mechanism ensures that the deviation between the reconstructed path and the original location sequence is within an acceptable range, avoiding data distortion.

[0086] This application further proposes that when performing trajectory reconstruction and temporal aggregation on continuous location data of tourists, it also includes:

[0087] The temporal aggregation includes spatiotemporal cube construction and behavioral feature quantification; the scenic area space is divided into uniform grid units, and the structured trajectory skeleton is aggregated in combination with discrete time windows to generate a three-dimensional spatiotemporal cube structure; the distribution of dwell time, visit frequency density and behavioral sequence patterns are statistically analyzed within the three-dimensional spatiotemporal cube structure; the periodic patterns and sudden change trends of tourist behavior are identified through cross-time window correlation analysis;

[0088] The temporal aggregation implements a dynamic granularity adjustment mechanism, which automatically adjusts the size of the spatial grid and the length of the time window according to the density of pedestrian traffic in the area; fine spatial grid division and short time window are used in densely populated areas, while coarse spatial grid division and long time window are used in sparsely populated areas.

[0089] In the temporal aggregation of continuous visitor location data, a spatiotemporal cube was constructed to divide the scenic area into uniform grid units, each initially set at 10 meters × 10 meters. Combined with discrete time windows, initially set at 5 minutes, the structured trajectory skeleton was aggregated spatially and temporally to form a three-dimensional spatiotemporal cube structure. Within this cube, the distribution of dwell time, visit frequency density, and behavioral sequence patterns within each grid unit were statistically analyzed to identify visitor behavior patterns of continuous stays exceeding 15 minutes in the exhibition area or the frequency of brief stays in the rest area. Through cross-time window correlation analysis, the periodic patterns of visitor behavior could be revealed: morning peaks are mainly concentrated at the entrance and popular attractions, midday crowds concentrate in the dining area, and sudden changes in trends, such as crowd gatherings caused by temporary performances or unexpected events, were also observed. Temporal aggregation also implements a dynamic granularity adjustment mechanism: in densely populated areas, such as popular tourist attractions or squares, the spatial grid is automatically divided into 5m×5m sections and the time window is shortened to 2 minutes to capture fine-grained behavioral changes; in sparsely populated areas, such as forest trails or remote scenic spots, the spatial grid can be expanded to 20m×20m sections and the time window can be extended to 10 minutes to reduce computation and retain key behavioral information.

[0090] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0091] In practical applications, when performing time-series aggregation of continuous visitor location data, the area is divided into uniform grid cells, initially set at 10 meters × 10 meters, and the time is divided into discrete 5-minute windows. The structured trajectory skeleton of visitors is aggregated into a three-dimensional spatiotemporal cube within this space-time grid. Each cube cell records the visitor's dwell time, visit frequency, and behavioral sequence patterns within that time period. Within the 9:00–9:05 AM time window, a popular roller coaster attraction grid cell recorded 10 visitors continuously staying and queuing; this behavioral pattern is recorded as high-density short-term stays. In contrast, grid cells in remote garden areas may only record 1–2 visitors briefly staying every 10 minutes. Through cross-time window analysis, periodic patterns of visitors are identified, such as morning peaks concentrated at the entrance and popular rides, midday crowds concentrated in the dining area, and trends related to sudden events, such as sudden outdoor performances causing visitor gatherings in specific areas. The dynamic granularity adjustment mechanism is effective in application: in the roller coaster queue area, where crowds are dense, the spatial grid is reduced to 5m x 5m and the time window is shortened to 2 minutes to precisely capture queue movement; while in gardens or forest trails where crowds are sparse, the grid is expanded to 20m x 20m and the time window is extended to 10 minutes to reduce computational load while maintaining the integrity of behavioral information. This method accurately reflects the dwell and movement behavior of visitors in different areas, providing a reliable basis for traffic control and service optimization.

[0092] Through the aforementioned technical solution, this application quantifies tourist behavior characteristics, including dwell time distribution, visit frequency density, and behavioral sequence patterns, by temporally aggregating continuous location data and constructing a three-dimensional spatiotemporal cube. Simultaneously, cross-time window analysis captures periodic patterns and sudden trends, enabling managers to have a comprehensive understanding of visitor flow dynamics. The dynamic granularity adjustment mechanism further enhances the accuracy and efficiency of data analysis: in densely populated areas, fine grids and short time windows accurately characterize peak queue lengths, congestion in hotspot areas, and short-term behavioral changes; while in sparsely populated areas, coarse grids and long time windows reduce computational redundancy while preserving the overall flow trend.

[0093] This application further proposes a method for extracting tourist activity preference features based on association rule mining and clustering algorithms and the activity data, including:

[0094] Support and confidence thresholds are set to filter behavioral sequences. Association rules between tourist activities are mined based on the activity data. A time decay mechanism is introduced to assign dynamic weight coefficients to behavioral records. The time evolution trajectory of the activity sequence is analyzed to identify high-frequency co-occurrence patterns. False association rules are filtered in combination with environmental context information. The mining results are dynamically pruned to retain a subset of effective rules. High-frequency behavioral sequences are used as the initial seed point set for the clustering algorithm.

[0095] In the process of extracting tourist activity preference features based on association rule mining and clustering algorithms, a support threshold (0.05, meaning the activity sequence appears in at least 5% of all tourists) and a confidence threshold (0.6, meaning that given the preceding behavior, the subsequent behavior has at least a 60% probability of occurring) are first set. These thresholds are used to filter out initially effective behavior sequences to ensure the statistical reliability of the mining results. A time decay mechanism is introduced into the activity records, giving higher weight to recent behaviors and lower weight to historical behaviors, so that the evolutionary features of the activity sequences can dynamically reflect the latest preferences of tourists. Tourists who frequently visit a certain exhibition area recently have higher weights and can preferentially influence rule generation. During the mining process, environmental context information (such as scenic area opening hours, special holiday activities, etc.) is combined to filter out possible pseudo-association rules to avoid misjudgments caused by accidental co-occurrence. A dynamic pruning strategy is used to remove low-value or redundant rules, retaining high-frequency behavior sequences as the initial seed points for the clustering algorithm to ensure that subsequent clustering can statistically reflect the true preference patterns of tourists.

[0096] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0097] In practical applications, visitor activity data revealed that some visitors frequently visited the roller coaster area between 9:00 AM and 11:00 AM, and then moved to the food and beverage area between 11:00 AM and 12:00 PM. With a support threshold of 0.05 and a confidence threshold of 0.6, the high-frequency behavior sequence "roller coaster → food and beverage area" was selected. Introducing a time decay mechanism, access records from the past week received higher weight, while records from a month ago received lower weight, making the extracted rules more closely reflect current visitor preferences. Simultaneously, by incorporating environmental context information such as holiday activities and scenic area opening hours, abnormal access data caused by temporarily closed scenic areas was filtered out, avoiding false rules. After dynamic pruning, high-frequency behavior sequences such as "roller coaster → food and beverage area" and "children's playground → gift shop" were retained. These sequences were used as initial seed points for clustering algorithms, forming visitor activity preference clusters, thus more accurately reflecting the behavioral characteristics of different types of visitors and providing foundational data for flow prediction and personalized recommendations.

[0098] Through the above technical solutions, this application effectively filters out true high-frequency behavior sequences by setting support and confidence thresholds, eliminating low-frequency and occasional behaviors, thus ensuring the accuracy of activity preference features. Introducing a time decay mechanism automatically assigns higher weight to recent behaviors, enabling the mining results to reflect tourists' latest preferences in a timely manner and adapt to changing behavioral trends. Combining environmental context information to filter out false association rules and retaining a subset of effective rules through dynamic pruning makes the initial seed points of the clustering algorithm more accurate, improving the stability and interpretability of subsequent clustering results.

[0099] This application further proposes that, when extracting tourist activity preference features based on association rule mining and clustering algorithms and the activity data, it also includes:

[0100] The initial seed point set is divided into behavioral clusters based on its spatiotemporal aggregation characteristics. The silhouette coefficient is calculated to evaluate the quality of the clustering results. The clustering parameters are iteratively adjusted to optimize the cluster boundary division. The high-frequency behavioral sequences output by association rule mining are fused as constraints for cluster boundary adjustment. The clustering results are output, which include cluster center behavioral patterns, cluster dynamic evolution trajectories, and cluster feature vectors. The clustering results serve as the activity preference features.

[0101] The initial seed point set is divided into several behavioral clusters based on their spatiotemporal aggregation characteristics. Density or distance is calculated based on the spatial coordinates and timestamp of each behavioral point, and behavioral points that are close in both time and location are grouped into the same cluster. The compactness and separation of each cluster are evaluated by calculating the silhouette coefficient to quantify the rationality of the clustering results. To optimize the cluster boundary partitioning, an iterative adjustment mechanism is implemented: clustering parameters, such as the number of clusters, the nearest neighbor radius, or the density threshold, are dynamically adjusted according to changes in the silhouette coefficient (the number of clusters can be initially set to 5, and the nearest neighbor radius can be set to 50 meters based on the total area of ​​the scenic area and the average tourist density). The high-frequency behavioral sequences obtained from association rule mining are used as constraints for cluster boundary adjustment to ensure that the high-frequency behavioral patterns within each cluster are as complete and continuous as possible. The final output clustering results include the central behavioral pattern of each cluster, the dynamic evolution trajectory of the cluster, and the cluster feature vector.

[0102] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0103] In practical applications, when analyzing tourist activity data, the initial set of high-frequency behavior points (e.g., "amusement facility A → dining area B → rest area C") is first clustered in the spatiotemporal dimension to form several behavior clusters. Each cluster represents a group of tourists with similar paths and stay patterns. For example, cluster 1 mainly consists of tourists who ride roller coasters and carousels in the morning, while cluster 2 consists of tourists who stay longer in the dining area at noon. The density and separation of each cluster are evaluated by calculating the silhouette coefficient. If the silhouette coefficient of a cluster is lower than 0.5, the clustering parameters, such as the number of clusters or the spatial proximity radius (which can be set from 50 meters to 150 meters, adjusted according to the density of the scenic area), are iteratively adjusted. At the same time, the high-frequency behavior sequences mined by association rules constrain the cluster boundaries to make the behavior sequences within the cluster as complete and continuous as possible. The output clustering results include cluster center behavior patterns (such as "entertainment first → dining then resting"), cluster dynamic evolution trajectory (showing the path the group moves over time), and cluster feature vectors (quantifying dwell time, visit frequency, and intensity of behavioral preferences). This information can be directly used as tourist activity preference features, providing data support for scenic area management, personalized recommendations, and visitor flow optimization.

[0104] Through the above technical solution, this application improves the accuracy and representativeness of tourist activity preference characteristics by clustering initial seed points into behavioral clusters in the spatiotemporal dimension and combining the silhouette coefficient to evaluate the quality of the clustering results. Iteratively adjusting the clustering parameters and fusing high-frequency behavioral sequences as boundary constraints can reduce inter-cluster overlap and noise interference, making the cluster center behavior patterns, cluster dynamic evolution trajectories, and cluster feature vectors more realistically reflect the behavioral patterns of tourists.

[0105] This application further proposes a method for predicting tourist flow trends using Markov chains and spatiotemporal evolution models based on the aforementioned activity preference characteristics, including:

[0106] The spatiotemporal evolution model integrates the state transition mechanism of the Markov chain with the dynamic response capability of the environment; it constructs a state space for the functional areas of the scenic area and defines the transfer probability matrix of tourists between areas; it introduces environmental factors to dynamically correct the transfer probability, including real-time visitor flow, area openness status, and the influence of external events; the transfer probability of the Markov chain is obtained through learning from historical behavioral data; the transfer probability matrix is ​​updated in real time in combination with the activity preference features; the spatiotemporal evolution model fine-tunes the transfer probability every time it receives new behavioral data; the prediction process generates a probability path for the distribution of tourists in future time periods, and the probability path serves as the flow prediction result.

[0107] In the process of predicting tourist flow trends based on activity preference characteristics, a state space of functional areas within the scenic area is constructed, with each area considered as a state node. Historical tourist behavior data is used to learn the transition probability matrix between areas. If historical data indicates that the transition probability from the scenic area entrance to the themed amusement area is 0.35, and the transition probability from the amusement area to the dining area is 0.25, then the initial matrix includes these values. The Markov chain model iteratively updates the state transition probabilities and dynamically corrects them using real-time environmental factors: reducing the transition probability of some paths during periods of high visitor traffic, and adjusting the probability values ​​accordingly when an area is temporarily closed or an external event (such as a sudden weather event) occurs. Simultaneously, activity preference characteristics are introduced, and the behavioral patterns of different tourist groups (such as frequent visits to amusement facilities and prolonged stays in dining areas) are weighted and incorporated into the transition matrix, making the predictions more aligned with individual preferences. The spatiotemporal evolution model fine-tunes the probability matrix in real time upon receiving new behavioral data, generating probabilistic paths for tourists in each area within future time periods. The model can predict that within the next hour, 40% of tourists may move from the entrance to the amusement area, 30% may go to the dining area, and 30% will be distributed in other rest areas. These probability paths constitute the flow prediction results.

[0108] The process of dynamically correcting the transfer probability of environmental factors involves quantifying the real-time collected environmental factors. For example, real-time pedestrian flow is divided into three levels (low, medium, and high) according to a preset interval and mapped to the corresponding correction coefficient λflow. The open state of a region is converted into a binary variable. If a region is closed, its relevant transfer probability is directly decayed to a value close to zero and mapped to the correction coefficient λopen. For external events (such as sudden weather changes or temporary activities), a corresponding correction coefficient λevent is assigned according to the event type and scope of impact. The above correction coefficients are weighted and fused with the original Markov transfer probability Pij to obtain the corrected transfer probability Pij′ = Pij × λflow × λopen × λevent. Normalization is performed before each round of prediction to ensure that the matrix row sum is 1. The process of updating the transfer probability matrix in real time involves counting the actual number of times tourists transfer from one area to another when receiving new behavioral data. A sliding time window mechanism (e.g., 5 minutes or 10 minutes) is used to dynamically count the latest transfer frequency and update it in combination with historical transfer probabilities. The update formula is Pij(t+1)=βPij(t)+(1-β)P^ij(t), where P^ij(t) is the observed transition frequency within the current time window, and β is the forgetting factor used to balance historical experience with real-time data. When a significant deviation is detected between short-term tourist flow patterns and long-term historical distributions, the response to new trends is accelerated by reducing the value of β. Through the above environmental factor correction and real-time update mechanism, the spatiotemporal evolution model can maintain the accuracy and sensitivity of the transition probability matrix in dynamic environments.

[0109] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0110] In practical applications, tourist flow prediction is divided into several functional areas, including an entrance area, amusement park area, dining area, rest area, and exit area, with each area serving as a state node in a Markov chain. By analyzing tourist behavior data from the past month, a preliminary transition probability matrix is ​​constructed, with a transition probability of 0.4 from the entrance area to the amusement park area, 0.2 from the entrance area to the dining area, and 0.3 from the amusement park area to the dining area, etc. During the prediction process, real-time monitoring of visitor flow, opening status, and environmental factors such as unforeseen events (e.g., rain or facility maintenance) is conducted. When excessive visitor flow is detected in the amusement park area, the transition probability from the entrance area or other areas to that area is dynamically reduced; if the dining area is temporarily closed, the relevant probability is adjusted to zero. Simultaneously, the activity preference characteristics of different tourist groups are incorporated into the model; the transition probability from the entrance area to the amusement park area increases accordingly for teenagers who prefer amusement park rides. Whenever new real-time location data is received, the spatiotemporal evolution model fine-tunes the transition matrix, generating probability paths for tourist distribution in each area for the next hour. The forecast results show that in the next hour, approximately 50% of visitors will enter the amusement park area from the entrance area, 25% will head to the food and beverage area, 15% will stay in the rest area, and the remaining 10% may go directly to the exit area. These probabilistic paths provide dynamic decision-making support for the park's operation and management, and can be used to optimize visitor flow guidance, facility scheduling, and safety management.

[0111] Through the aforementioned technical solution, this application uses Markov chains and a spatiotemporal evolution model to predict tourist flow trends based on activity preference characteristics, thereby improving the scientific and refined level of scenic area management. By constructing a state space for functional areas of the scenic area and defining a transition probability matrix between areas, the model can quantify the probability of tourist flow between different areas. Introducing environmental factors such as real-time visitor flow, area openness status, and unexpected events allows for dynamic correction of the transition probability, enabling rapid response to abnormal situations. Combining activity preference characteristics with real-time updates to the transition matrix enables the model to accurately reflect the behavioral preferences and spatial selection patterns of different tourist groups. Each time new behavioral data is received, the transition probability can be fine-tuned, thereby generating probabilistic paths for tourist distribution in future time periods, providing the scenic area with quantitative flow prediction results.

[0112] This application further proposes that when generating visualized dynamic decision support based on the flow prediction results, it includes:

[0113] The dynamic decision support includes rendering a heat map of tourist density distribution and generating resource scheduling suggestions; a heat map of tourist density distribution is constructed based on the flow prediction results, and the degree of regional congestion is presented through color gradients;

[0114] The bottleneck areas exceeding the density threshold in the heat map of tourist density distribution are identified, and dynamic decision support is generated in combination with the layout of scenic area facilities; the dynamic decision support includes diversion path planning schemes, service facility expansion schemes and emergency response strategies.

[0115] The system generates visualized dynamic decision support based on the flow prediction results, enabling real-time optimization and precise scheduling of scenic area management. It maps the probability paths of tourist distribution in different areas over future periods onto the scenic area map and constructs a heat map of tourist density distribution. Different density levels are represented by color gradients: green for below 200 people / hectare, yellow for 200–500 people / hectare, orange for 500–800 people / hectare, and red for above 800 people / hectare, visually demonstrating the degree of congestion in each area. Further analysis of the heat map automatically identifies bottleneck areas exceeding a preset density threshold (e.g., 800 people / hectare) and generates dynamic decision support based on the layout of scenic area facilities, passageway capacity, and safety requirements. This decision support includes: planning diversion routes for high-density areas, guiding tourists to detour through low-density passageways; generating service facility enhancement plans, such as temporarily opening rest areas, adding drinking water points or sanitation facilities; and formulating emergency response strategies, including early warning broadcasts, security personnel deployment, and emergency evacuation routes. Through these visualized and intelligent decision outputs, managers can monitor the dynamic flow of people in the scenic area in real time, intervene in potential congestion in advance, optimize the visitor experience, and improve the overall operational safety and efficiency of the scenic area.

[0116] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0117] In practical applications, predicted future visitor flow is mapped onto a scenic area map to generate a real-time visitor density distribution heatmap. The visitor flow in each grid cell is mapped to a color gradient: green represents areas with fewer than 200 people per cell, yellow represents 200–500 people per cell, orange represents 500–800 people per cell, and red represents high-density areas with more than 800 people per cell. Red high-density areas are automatically identified as bottlenecks, such as the main entrance, plazas in front of popular attractions, or pedestrian walkways. Combined with the layout of scenic area facilities, dynamic decision support is generated: for congested areas at the main entrance, diversion routes are planned to guide some visitors to enter the scenic area via bypasses or secondary passages; for popular attractions, service facility enhancement plans are proposed, such as adding temporary rest areas, mobile drinking fountains, and sanitation facilities; simultaneously, emergency response strategies are developed, such as initiating broadcast alerts, dispatching security personnel to guide crowds, and opening backup passages to ensure safety when visitor flow continuously exceeds thresholds. Through this visualized dynamic decision support, scenic area managers can intuitively grasp visitor flow distribution, take timely control measures, and improve visitor experience and safety management.

[0118] Through the aforementioned technical solutions, this application generates visualized dynamic decision support based on flow prediction results, improving the scientific nature and response efficiency of scenic area management. By constructing a heat map of visitor density distribution, managers can intuitively observe the congestion level in each area and identify potential bottlenecks in advance, such as entrances, popular attractions, and pedestrian walkways. In areas exceeding the density threshold, a diversion path planning scheme is automatically generated based on the layout of scenic area facilities, guiding visitors to distribute themselves rationally and reducing the risk of congestion. Simultaneously, a service facility enhancement plan is proposed, adding temporary rest areas, mobile drinking fountains, and sanitation facilities, improving visitor comfort and satisfaction. For sudden peak visitor flows, the dynamic decision support provides emergency response strategies, including dispatching security personnel, activating broadcast guidance, and opening backup passages.

[0119] In summary, by integrating multi-source positioning data such as GPS, BeiDou, Bluetooth beacons, and AI video recognition, and combining a statistical optimization method with dynamic weight allocation and the spatiotemporal feature extraction capabilities of a deep learning model, the complementary and optimized use of different data sources in complex scenic environments is achieved. This makes tourist location estimation more accurate and robust, avoiding the problem of accumulated bias caused by errors from a single data source. Through a dynamic weight adjustment mechanism and iterative optimization loop, the positioning fusion process can respond in real time to environmental changes and signal fluctuations, still outputting continuous and stable location data even when signals are lost or degraded, improving its applicability in complex terrain and high-density crowd scenarios. The trajectory reconstruction method not only simplifies trajectories and identifies regional functions through compression and semantic annotation, but also ensures the accuracy of trajectory reconstruction through dynamic verification and correction, thereby generating a structured activity data map, providing a solid foundation for subsequent behavior analysis and modeling. Through a spatiotemporal cube structure and a dynamic granular aggregation mechanism, multi-dimensional statistics and modeling of tourist stay duration, visit frequency, and behavior patterns are performed. Combined with association rule mining and clustering algorithms to extract tourist activity preference features, a refined expression of tourist behavior patterns and group differentiation identification are achieved. A predictive framework combining Markov chains and spatiotemporal evolution models not only characterizes the probability of tourist movement between functional areas but also incorporates environmental factors for dynamic correction in real time. This allows the prediction results to reflect the temporal characteristics of tourist flow and environmental responses, making them more consistent with real-world scenarios. By generating heatmaps of tourist density distribution and identifying bottleneck areas, and combining this with the layout of scenic area resources, the framework provides diversion schemes, facility upgrades, and emergency strategies. This offers scenic area managers visualized and dynamic decision support, alleviating congestion, improving the tourist experience, and ensuring the safety of scenic area operations.

[0120] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program goods. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program goods embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0121] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program goods according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0122] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0123] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0124] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1.A method for dynamic modeling of tourist behavior based on spatiotemporal big data, characterized in that, The application relates to a method for predicting tourist flow trend based on multi-source positioning data. The method comprises the following steps: Collecting multi-source positioning data, wherein the multi-source positioning data comprises GPS position data, Beidou positioning data, Bluetooth beacon data and AI video recognition data; and preprocessing the multi-source positioning data; Inputting the preprocessed multi-source positioning data into a multi-source positioning fusion engine, and fusing position results based on a statistical optimization method and a deep learning model to obtain a high-precision positioning model; Based on the high-precision positioning model, reconstructing a trajectory and aggregating time sequences of continuous position point data of tourists to generate an activity data graph, wherein the activity data graph is a structured spatio-temporal behavior data set, and activity data of the tourists at various positions is obtained; Based on an association rule mining and clustering algorithm and according to the activity data, extracting activity preference features of the tourists; 2.The tourist behavior dynamic modeling method based on spatiotemporal big data according to claim 1, wherein, Based on the activity preference features, predicting a tourist flow trend by using a Markov chain and a spatio-temporal evolution model, and generating a flow prediction result, and generating a visual dynamic decision support according to the flow prediction result. When preprocessing the multi-source positioning data, the following steps are included: Data cleaning, wherein noise points and abnormal jumps in the positioning data are identified, and filtering technology is applied to eliminate positioning deviations caused by signal interference and equipment errors; Time synchronization, wherein clock offsets between different data sources are corrected, and a continuous time sequence is generated by using an interpolation method, so that the multi-source positioning data are strictly aligned in the time dimension; 3.The tourist behavior dynamic modeling method based on spatiotemporal big data according to claim 2, wherein, Coordinate normalization, wherein position data in heterogeneous coordinate systems are uniformly mapped to a common geographic coordinate system, so that all position points have spatial comparability. When fusing the position results based on the statistical optimization method and the deep learning model to obtain the high-precision positioning model, the following steps are included: The statistical optimization method is a dynamic weight distribution mechanism, which adaptively adjusts data source weight coefficients according to positioning reliability of each data source in a real-time environment; the positioning reliability is determined by comprehensively analyzing historical data stability and current signal quality; the data source weight coefficients are dynamically updated with the change of the environment, and the data source weight coefficients are automatically lowered when it is detected that the signal quality of the data source decreases; the data source weight coefficients are automatically increased when it is detected that the signal quality of the data source increases; 4.The tourist behavior dynamic modeling method based on spatiotemporal big data according to claim 3, wherein, The deep learning model adopts a spatio-temporal feature joint extraction architecture, which comprises a time sequence coding module and a spatial correlation module; the time sequence coding module processes a continuous position point sequence by using a recurrent neural network structure, and captures the evolution law of the position change with time; the spatial correlation module analyzes the spatial adjacency relationship between position points, and quantifies the moving correlation strength between different regions; the time sequence coding module outputs a time sequence feature vector, and the spatial correlation module outputs a spatial feature vector, which are input into an attention mechanism layer; the attention mechanism layer calculates feature contribution weights of the feature vectors, and generates a fusion feature vector by weighted sum of the feature vectors according to the feature contribution weights; the fusion feature vector is converted into a position coordinate through a full connection layer. When fusing the position results based on the statistical optimization method and the deep learning model to obtain the high-precision positioning model, the following steps are also included: The statistical optimization method and the deep learning model establish a bidirectional data channel, the preliminary fusion result output by the statistical optimization is taken as one of the input features of the high-precision positioning model, and the correction result output by the high-precision positioning model is fed back to the statistical optimization link for updating the weight; the bidirectional data channel continuously exchanges the intermediate results of position fusion, forming an iterative optimization cycle. 5.The tourist behavior dynamic modeling method based on spatiotemporal big data according to claim 4, wherein, When reconstructing the trajectory of the continuous position point data of the tourists and performing time sequence aggregation, the following are included: The trajectory reconstruction includes a trajectory compression stage and a semantic labeling stage; The trajectory compression stage simplifies the algorithm based on geometric features, identifies key turning points in the position coordinates and eliminates redundant position points, and generates a simplified moving path; the identification of the key turning points is based on the path curvature change rate and the direction mutation degree; The semantic labeling stage analyzes the position point aggregation characteristics and the stay duration distribution, divides the effective stay area and associates the function area type of the scenic area; The trajectory reconstruction adopts a dynamic verification mechanism to calculate the spatial deviation of the reconstructed path and the original position sequence, and triggers a path correction process when the spatial deviation exceeds a preset threshold; the path correction process adjusts the position of the key turning points; the position of the key turning points and the stay area jointly constitute the structured trajectory skeleton of the activity data graph, which is used as the input data of the time sequence aggregation. 6.The tourist behavior dynamic modeling method based on spatiotemporal big data according to claim 5, wherein, When reconstructing the trajectory of the continuous position point data of the tourists and performing time sequence aggregation, the following are also included: The time sequence aggregation includes the construction of a spatio-temporal cube and the quantification of behavior characteristics; the structured trajectory skeleton is aggregated to generate a three-dimensional spatio-temporal cube structure by dividing the scenic area space into uniform grid units and combining discrete time windows; the stay duration distribution, visit frequency density and behavior sequence pattern are statistically analyzed within the three-dimensional spatio-temporal cube structure; the periodicity of the tourist behavior and the sudden change trend are identified through cross-time window association analysis; The time sequence aggregation implements a dynamic granularity adjustment mechanism to automatically adjust the spatial grid size and the time window length according to the regional passenger flow density; fine spatial grid division and short time window are used in areas with dense passenger flow, and coarse spatial grid division and long time window are used in areas with sparse passenger flow. 7.The tourist behavior dynamic modeling method based on spatiotemporal big data of claim 6, wherein, When extracting the activity preference features of the tourists based on association rule mining and clustering algorithm and according to the activity data, the following are included: The support threshold and the confidence threshold are set to filter the behavior sequence, the association rules between the activities of the tourists are mined according to the activity data, a time decay mechanism is introduced to assign dynamic weight coefficients to the behavior records, the time evolution trajectory of the activity sequence is analyzed to identify high-frequency co-occurrence patterns, the environmental context information is combined to filter false association rules, the mining results are dynamically pruned to retain an effective rule subset, and the high-frequency behavior sequence is taken as the initial seed point set of the clustering algorithm. 8.The tourist behavior dynamic modeling method based on spatiotemporal big data of claim 7, wherein, When extracting the activity preference features of the tourists based on association rule mining and clustering algorithm and according to the activity data, the following are also included: The initial seed point set is divided into clusters according to the aggregation characteristics in the space-time dimension, the contour coefficient is calculated to evaluate the quality of the clustering results, the cluster boundary division is optimized by iteratively adjusting the clustering parameters, the high-frequency behavior sequence output by the association rule mining is fused as a constraint condition for adjusting the cluster boundary, and the clustering results are output, which include the cluster center behavior mode, the cluster dynamic evolution trajectory and the cluster feature vector, and the clustering results are used as the activity preference feature. 9.The tourist behavior dynamic modeling method based on spatiotemporal big data of claim 8, wherein, Based on the activity preference feature, the Markov chain and the space-time evolution model are used to predict the tourist flow trend, including: The space-time evolution model integrates the state transition mechanism of the Markov chain and the dynamic response ability of the environment; the state space of the functional area of the scenic spot is constructed, and the transfer probability matrix of tourists between areas is defined; the dynamic correction of the transfer probability is introduced by introducing environmental factors, including real-time passenger flow, area opening state and external event influence; the transfer probability of the Markov chain is obtained by learning from historical behavior data; the transfer probability matrix is updated in real time combined with the activity preference feature; the space-time evolution model adjusts the transfer probability every time new behavior data is received; the prediction process generates the probability path of the future period tourist distribution, which is used as the flow prediction result. 10.The tourist behavior dynamic modeling method based on spatiotemporal big data of claim 9, wherein, According to the flow prediction result, the visual dynamic decision support is generated, including: The dynamic decision support includes tourist density distribution heat map rendering and resource scheduling suggestion generation; the tourist density distribution heat map is constructed based on the flow prediction result, and the color gradient is used to present the area congestion degree; The bottleneck area of the super-density threshold in the tourist density distribution heat map is identified, and the dynamic decision support is generated combined with the layout of the scenic spot facilities; the dynamic decision support includes the shunt path planning scheme, the service facility supplement scheme and the emergency response strategy.

Images