Smart park navigation data analysis method and system based on multimodality

Through the comprehensive processing of generating navigation feature vectors and navigation decision-making networks, the problem of multimodal data processing in the smart park navigation system is solved, precise navigation and personalized services for complex environments are realized, and the flexibility of the system and user satisfaction are improved. Spend.

CN118794448BActive Publication Date: 2024-11-12GUANGZHOU ELECTRONICS TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411289139.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-14
Publication Date
2024-11-12
Estimated Expiration
2044-09-14

AI Technical Summary

Technical Problem

The existing smart park navigation system is difficult to effectively process multimodal data, ignores environmental information and user personalized needs, lacks flexibility and dynamic adjustment capabilities, and is difficult to deal with complex park environments and emergencies.

Method used

By generating navigation feature vectors, combining implicit perturbation features and navigation decision networks, comprehensive processing and real-time response to multimodal data is realized, forward and reverse delivery subnets are integrated to generate personalized navigation optimization decision results.

Benefits of technology

It improves the flexibility and accuracy of the navigation system, meets the needs of main navigation and subordinate navigation, and improves user satisfaction and navigation intelligence level.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118794448B_ABST
    Figure CN118794448B_ABST
Patent Text Reader

Abstract

The present application provides a method and system for analyzing navigation data of a smart campus based on multimodality. First, a first navigation feature vector and multiple second navigation feature vectors are generated through feature representation, which effectively integrates the main navigation requirements of the campus and the diversified subordinate navigation restriction requirements. Furthermore, by introducing implicit perturbation features and relying on the forward derivative transmission subnetwork, potential environmental change factors can be captured and integrated in real time, thereby improving the flexibility and adaptability of navigation decisions. In the reverse derivative transmission subnetwork, multiple parallel derivative transmission trajectories are generated by channel-by-channel fusion of navigation feature vectors, each of which is optimized for specific subordinate navigation requirements, thereby realizing the customization of personalized navigation solutions. Ultimately, the navigation optimization decision results generated based on these comprehensive features not only meet the unified main navigation restriction requirements, but also fully consider the differentiated subordinate navigation restrictions, significantly improving the accuracy of campus navigation and user satisfaction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the technical field of smart park technologies, and more specifically, to a multimodal-based smart park navigation data analysis method and system. Background Art

[0002] With the rapid development of smart cities, smart parks are an important part of urban intelligent construction. The intelligence level of their internal navigation systems directly affects the park's operational efficiency and user experience. Traditional park navigation systems mainly rely on single geographic location information or simple path planning algorithms, which are difficult to cope with complex and changing park environments and diverse user needs. Especially in large smart parks, it is necessary to consider not only the navigation of main roads and buildings, but also a variety of subordinate navigation restrictions such as visitor parking, employee restaurant locations, and cargo transportation channels. These requirements often involve multi-source and multi-type data (i.e., multi-modal data).

[0003] Most existing navigation data analysis methods are limited to processing single-mode data, such as using only GPS coordinate information for path planning, ignoring the rich environmental information in the park (such as weather changes, temporary activities, equipment failures, etc.) and the personalized navigation needs of users. In addition, these methods often lack flexibility and dynamic adjustment capabilities when dealing with complex park environments, and are difficult to cope with emergencies or changes in uncertain factors.

[0004] In order to overcome the above limitations and improve the intelligence level and user satisfaction of the smart park navigation system, a navigation data analysis method is needed that can comprehensively process multimodal data, consider various navigation restriction requirements, and respond to environmental changes in real time. Summary of the invention

[0005] In view of the above-mentioned problems, in combination with the first aspect of the present application, an embodiment of the present application provides a multi-modal smart park navigation data analysis method, the method comprising:

[0006] The main navigation multimodal data is characterized to generate a first navigation feature vector, and X second navigation feature vectors are generated according to the first navigation feature vector and X slave multimodal data; X is a positive integer, the main navigation multimodal data is multimodal acquisition data of main navigation restriction requirements in the smart park, the slave multimodal data is multimodal acquisition data of slave navigation restriction requirements in the smart park, and the slave navigation restriction requirements represented by the X slave multimodal data are all different; a second navigation feature vector is generated by the first navigation feature vector and one slave multimodal data;

[0007] Obtaining implicit disturbance features, loading the implicit disturbance features into a forward derivative transfer subnetwork in a target park navigation decision network, and using the forward derivative transfer subnetwork to output forward disturbance features;

[0008] The forward perturbation feature is loaded into the reverse derivative transfer subnetwork in the target park navigation decision network, and X derivative transfer trajectories in parallel relationship are generated by using the reverse derivative transfer subnetwork; the reverse derivative transfer subnetwork is used to fuse the first navigation feature vector in a first derivative fusion channel, and is used to fuse a second navigation feature vector for each derivative transfer trajectory in a second derivative fusion channel; the loading of the second derivative fusion channel includes the generation of the first derivative fusion channel;

[0009] Based on the forward disturbance feature, the first navigation feature vector fused in the first derivative fusion channel, and the X second navigation feature vectors fused in the second derivative fusion channel, a navigation optimization decision result corresponding to each of the X derivative transfer trajectories is generated; the main navigation restriction requirements in the X navigation optimization decision results are the same, and the subordinate navigation restriction requirements in the X navigation optimization decision results are different.

[0010] On the other hand, an embodiment of the present application also provides a smart campus service system, including a processor and a machine-readable storage medium, wherein the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0011] Based on the above aspects, the embodiment of the present application realizes accurate navigation support for complex campus environments by comprehensively processing the main navigation multimodal data and multiple subordinate multimodal data. Specifically, firstly, the first navigation feature vector and multiple second navigation feature vectors are generated through feature representation, which effectively integrates the main navigation requirements of the campus and the diversified subordinate navigation restriction requirements. Furthermore, by introducing implicit perturbation features and using the forward derivative transmission subnetwork, potential environmental change factors can be captured and integrated in real time, thereby improving the flexibility and adaptability of navigation decisions. In the reverse derivative transmission subnetwork, multiple parallel derivative transmission trajectories are generated by channel-by-channel fusion of navigation feature vectors, each of which is optimized for specific subordinate navigation requirements, thereby realizing the customization of personalized navigation solutions. Finally, the navigation optimization decision results generated based on these comprehensive features not only meet the unified main navigation restriction requirements, but also fully consider the differentiated subordinate navigation restrictions, significantly improving the accuracy and user satisfaction of campus navigation. As a result, the navigation intelligence level of the smart campus is greatly enhanced, providing users with a more convenient, efficient and personalized navigation service experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 It is a schematic diagram of the execution flow of the multimodal smart park navigation data analysis method provided in an embodiment of the present application.

[0013] Figure 2 It is a schematic diagram of the hardware architecture of the smart park service system provided in the embodiment of the present application. DETAILED DESCRIPTION

[0014] The present application will be described in detail below with reference to the accompanying drawings. Figure 1 This is a flow chart of a multimodal smart park navigation data analysis method provided in an embodiment of the present application. The multimodal smart park navigation data analysis method is introduced in detail below.

[0015] Step S110, feature representation of the main navigation multimodal data to generate a first navigation feature vector, and generate X second navigation feature vectors based on the first navigation feature vector and X subordinate multimodal data. X is a positive integer, the main navigation multimodal data is multimodal acquisition data of the main navigation restriction requirements in the smart park, the subordinate multimodal data is multimodal acquisition data of the subordinate navigation restriction requirements in the smart park, and the subordinate navigation restriction requirements represented by the X subordinate multimodal data are all different. A second navigation feature vector is generated by the first navigation feature vector and one subordinate multimodal data.

[0016] Specifically, in a smart park, primary navigation multimodal data refers to a multi-source, multi-type data set that has a decisive influence on the main navigation needs in the smart park. These primary navigation multimodal data usually include the traffic conditions of the main roads in the park, the location and opening hours of important buildings, the use of public facilities, etc., which are the basic information necessary for planning the main navigation paths in the park and formulating navigation strategies.

[0017] The first navigation feature vector refers to a vector generated by extracting features and learning representations of the main navigation multimodal data, which comprehensively reflects the key features and information of the main navigation multimodal data. The first navigation feature vector is the basis for subsequent navigation decision optimization and can fully characterize the core elements of the main navigation needs of the park.

[0018] Corresponding to the main navigation multimodal data, the subordinate multimodal data refers to the multi-source and multi-type data sets with other specific navigation restriction requirements in the smart park, in addition to the main navigation requirements. These subordinate multimodal data are usually targeted at specific user groups or specific scenarios, such as visitor parking requirements, employee restaurant location requirements, special channel requirements for cargo transportation, etc.

[0019] The second navigation feature vector is a vector generated based on the first navigation feature vector and specific subordinate multimodal data. Each second navigation feature vector targets a specific subordinate navigation requirement, integrates the main navigation information and the specific information of the requirement, and provides support for generating personalized navigation decision results.

[0020] That is, in this embodiment, the server is processing navigation data of a smart park, which has various navigation requirements, such as visitor navigation, employee navigation, cargo transportation navigation, etc.

[0021] The main navigation multimodal data can be multimodal data collected about navigation restriction requirements of the main roads, important buildings and public facilities in the park. For example, the server receives multimodal data about the traffic flow of the main roads in the park, road conditions (whether it is under construction, whether it is congested, etc.), opening hours of major buildings and access rules.

[0022] The server first represents the features of these main navigation multimodal data. It is assumed that the features are extracted by using a combination of convolutional neural network (CNN) and recurrent neural network (RNN) in deep learning. CNN is used to process image or spatial features, such as road layout images; RNN is used to process time series features, such as changes in traffic flow. After processing, a first navigation feature vector is generated that can comprehensively represent these main navigation information.

[0023] Next, assume that there are X = 3 dependent multimodal data, which are about visitor parking requirements, employee restaurant location requirements, and special channel requirements for cargo transportation.

[0024] The subordinate multimodal data of visitor parking requirements may include the number of vacancies in each parking lot, parking fees, distance to the destination, etc. The server fuses the first navigation feature vector with the data of visitor parking requirements, and generates a second navigation feature vector for visitor parking through a specific fusion algorithm (such as weighted summation, feature splicing, etc.).

[0025] The subordinate multimodal data of the staff restaurant location requirements may include the restaurant's business hours, dish features, seat availability, etc. Similarly, the server fuses the first navigation feature vector with this data to obtain a second navigation feature vector of the staff restaurant location.

[0026] The subordinate multimodal data of the special channel requirements for cargo transportation may involve the opening time of the special channel, the types of cargo allowed to pass, and the size restrictions, etc. The server performs a fusion operation again to generate a second navigation feature vector for the special channel for cargo transportation.

[0027] Step S120, obtaining implicit disturbance features, loading the implicit disturbance features into the forward derivative transfer sub-network in the target park navigation decision network, and using the forward derivative transfer sub-network to output the forward disturbance features.

[0028] In detail, the implicit disturbance features refer to a set of features that are difficult to directly observe or predict but have a potential impact on navigation decisions. These implicit disturbance features may include uncertain factors such as weather changes, temporary activities in the park, and equipment failures, which need to be extracted and represented through methods such as data analysis and model prediction.

[0029] The forward derivative transmission subnetwork refers to a subnetwork in the target park navigation decision network that is responsible for forward transmission and processing of implicit disturbance features. The forward derivative transmission subnetwork fuses implicit disturbance features with other input features through a series of operations and transformations to generate forward disturbance features containing disturbance effects, providing input for subsequent navigation decision optimization.

[0030] Therefore, in this embodiment, the server first obtains the implicit disturbance feature. This can be achieved in the following way:

[0031] Assume that the server obtains a large amount of prior random perturbation data from an external data source. These prior random perturbation data may be simulation data about uncertain factors such as weather changes, temporary activities in the park, and equipment failures. The server loads these prior random perturbation data into the perturbation encoding subnetwork in the target park navigation decision network. The perturbation encoding subnetwork can be a module composed of a multi-layer fully connected neural network.

[0032] In the perturbation encoding subnetwork, the prior random perturbation data is first encoded, for example, converted into low-dimensional mean features and fluctuation features. The mean feature can represent the average impact of these perturbations, and the fluctuation feature reflects the magnitude and uncertainty of its changes. Then, the server randomly selects from these mean features and fluctuation features. For example, a combination of implicit mean features and implicit fluctuation features with a larger mean and smaller fluctuation is randomly selected.

[0033] Next, this implicit perturbation feature is loaded into the forward derivative transfer subnetwork in the target park navigation decision network. The forward derivative transfer subnetwork can be a deep neural network with multiple hidden layers. In this forward derivative transfer subnetwork, the server obtains the derivative transfer feature distribution, which describes the mode of data fusion and transmission in the network.

[0034] Assume that the forward derivative transfer subnetwork has Y=5 time series nodes. In each time series node, the server incorporates different random perturbation data into the implicit perturbation feature according to the characteristics of the node and the above-mentioned derivative transfer feature distribution.

[0035] For example, at the first time series node, a random disturbance data indicating a slight traffic congestion change is incorporated; at the second time series node, a random disturbance data indicating temporary construction in a certain area of ​​the park is incorporated; at the third time series node, a random disturbance data indicating a sudden change in weather (such as rainfall) is incorporated; at the fourth time series node, a random disturbance data indicating a sudden failure of an important device is incorporated; at the fifth time series node, a random disturbance data indicating an increase in the flow of people due to a large-scale event held in the park is incorporated. After processing at these five time series nodes, the server finally generates a forward disturbance feature.

[0036] Step S130, loading the forward disturbance feature into the reverse derivative transfer subnetwork in the target park navigation decision network, and using the reverse derivative transfer subnetwork to generate X derivative transfer trajectories in parallel. The reverse derivative transfer subnetwork is used to fuse the first navigation feature vector in the first derivative fusion channel, and to fuse a second navigation feature vector for each derivative transfer trajectory in the second derivative fusion channel. The loading of the second derivative fusion channel includes the generation of the first derivative fusion channel.

[0037] In detail, the reverse derivative transfer subnetwork refers to a subnetwork in the target park navigation decision network that is responsible for reverse transfer and processing features and generating multiple derivative transfer trajectories. The subnetwork generates multiple parallel derivative transfer trajectories by fusing the main navigation features, the subordinate navigation features and the forward perturbation features. Each trajectory targets a specific navigation demand scenario and provides support for generating personalized navigation decision results.

[0038] The derived transfer trajectory refers to multiple parallel navigation paths or strategy solutions generated in the reverse derived transfer subnetwork. Each derived transfer trajectory is based on the same main navigation requirement and different subordinate navigation requirements, and forms a personalized navigation path or strategy for a specific user group or scenario by integrating multiple feature information.

[0039] Therefore, in this embodiment, the server loads the above-generated forward disturbance features into the reverse derivative transfer subnetwork in the target park navigation decision network. The reverse derivative transfer subnetwork is used to process the reverse transfer and fusion of data. Assume that the reverse derivative transfer subnetwork includes a first derivative fusion channel and a second derivative fusion channel. In the first derivative fusion channel, the server integrates the first navigation feature vector into it. This may be achieved through a series of matrix operations and feature splicing operations to ensure that the first navigation feature vector can be effectively combined with subsequent processing steps. At the same time, in the second derivative fusion channel, a second navigation feature vector is fused for each derivative transfer trajectory of the parallel relationship.

[0040] Assume that there are X=3 parallel derivative transfer trajectories, corresponding to the visitor parking, employee restaurant location and special channel requirements for cargo transportation mentioned above.

[0041] For the derived delivery trajectory of visitor parking, the server loads the corresponding second navigation feature vector (including the fusion of visitor parking related features and the first navigation feature vector) into the second derived fusion channel. For the derived delivery trajectory of the employee restaurant location, the corresponding second navigation feature vector is also loaded into the channel. For the derived delivery trajectory of the special channel for cargo transportation, the same operation is performed.

[0042] In this way, feature fusion and trajectory generation in the reverse derivative transfer subnetwork are completed.

[0043] Step S140: Based on the forward disturbance feature, the first navigation feature vector fused in the first derivative fusion channel, and the X second navigation feature vectors fused in the second derivative fusion channel, a navigation optimization decision result corresponding to each of the X derivative transfer trajectories is generated. The main navigation restriction requirements in the X navigation optimization decision results are the same, and the subordinate navigation restriction requirements in the X navigation optimization decision results are all different.

[0044] In detail, the navigation optimization decision result is the final navigation decision result generated by the target park navigation decision network based on the forward disturbance feature, the first navigation feature vector and multiple second navigation feature vectors. These navigation optimization decision results are optimized for each derived transfer trajectory to ensure that while meeting the main navigation requirements, the subordinate navigation requirements and potential disturbance factors are fully considered to provide users with the best navigation path and strategy.

[0045] That is, in this embodiment, based on the previously fused forward disturbance features, the first navigation feature vector fused in the first derivative fusion channel, and the three second navigation feature vectors fused in the second derivative fusion channel, the server generates a navigation optimization decision result corresponding to each of the three derivative transfer trajectories.

[0046] Assume that the main navigation constraint requirements of these three navigation optimization decision results are all to reach the central square of the park from the park entrance, but the subordinate navigation constraint requirements are different, namely, visitors reach the central square after parking, employees reach the central square from the restaurant location, and goods reach the central square after being transported through a special channel.

[0047] The server first uses the reverse derivative transfer sub-network to fuse the first navigation feature vector in the first derivative fusion channel. Through a series of convolution, pooling and full connection operations, the forward perturbation features and the fused first navigation feature vector are compressed to generate the first decision factor feature. This first decision factor feature comprehensively considers the basic information of the main navigation and potential perturbation factors.

[0048] Then, the first decision factor feature is loaded into three derivative transfer trajectories respectively. In the second derivative fusion channel of each derivative transfer trajectory, the server performs dimensionality upscaling processing on the first decision factor feature and the corresponding second navigation feature vector.

[0049] Taking the derived transfer trajectory of visitor parking as an example, the server fuses and transforms the first decision factor feature with the second navigation feature vector of visitor parking through self-focus association processing and interactive focus association processing to generate the second decision factor feature for the visitor to reach the central square after parking. This feature fully considers the specific needs of visitor parking and the changes of the main navigation path under disturbance. For the derived transfer trajectories of the employee restaurant location and the special channel for cargo transportation, similar operations are performed to generate the corresponding second decision factor features respectively. Finally, the server generates the navigation optimization decision results based on these three second decision factor features.

[0050] For example, the decision result for visitors to arrive at the central square after parking may include the recommended parking lot, the specific route from the parking lot to the central square, the estimated arrival time, and take into account possible traffic congestion and temporary changes in road conditions; the decision result for employees to arrive at the central square from the restaurant location may provide the best route from the current restaurant location, avoiding construction areas and crowded areas; the decision result for goods to arrive at the central square after being transported through special channels may plan channels and routes that meet the restrictions on goods transportation, and estimate the transportation time and possible obstacles and solutions.

[0051] Based on the above steps, the embodiment of the present application realizes accurate navigation support for complex campus environments by comprehensively processing the main navigation multimodal data and multiple subordinate multimodal data. Specifically, firstly, the first navigation feature vector and multiple second navigation feature vectors are generated through feature representation, which effectively integrates the main navigation requirements of the campus and the diversified subordinate navigation restriction requirements. Furthermore, by introducing implicit perturbation features and using the forward derivative transmission subnetwork, potential environmental change factors can be captured and integrated in real time, thereby improving the flexibility and adaptability of navigation decisions. In the reverse derivative transmission subnetwork, multiple parallel derivative transmission trajectories are generated by channel-by-channel fusion of navigation feature vectors, each of which is optimized for specific subordinate navigation requirements, thereby realizing the customization of personalized navigation solutions. Finally, the navigation optimization decision results generated based on these comprehensive features not only meet the unified main navigation restriction requirements, but also fully consider the differentiated subordinate navigation restrictions, significantly improving the accuracy and user satisfaction of campus navigation. As a result, the navigation intelligence level of the smart campus is greatly enhanced, providing users with a more convenient, efficient and personalized navigation service experience.

[0052] In a possible implementation, step S140 includes:

[0053] Step S141, using the reverse derivative transfer subnetwork to fuse the first navigation feature vector in the first derivative fusion channel, compressing the forward perturbation feature and the first navigation feature vector fused in the first derivative fusion channel, generating a first decision factor feature obtained by the first derivative fusion channel, and loading the first decision factor feature into the X derivative transfer trajectories respectively.

[0054] Step S142: fuse a second navigation feature vector for each derived transfer trajectory in the second derivative fusion channel, perform dimensionality up-conversion processing on the first decision factor feature and the second navigation feature vector fused in the second derivative fusion channel in each derived transfer trajectory, generate X second decision factor features obtained by the second derivative fusion channel, and generate navigation optimization decision results corresponding to each of the X derived transfer trajectories based on the X second decision factor features.

[0055] In this embodiment, it is considered that the smart park has a variety of complex navigation requirements, including visitor navigation, employee navigation, and cargo transportation navigation.

[0056] First, the server has completed the data preparation and feature extraction in the previous steps. In the first derivative fusion channel, the reverse derivative transfer subnetwork begins to fuse the first navigation feature vector. The first navigation feature vector integrates the key navigation restriction requirements of the main roads, important buildings and public facilities in the park, such as the real-time traffic flow of the main roads in the park, the opening hours of important buildings, etc.

[0057] The server compresses the forward disturbance features (including feature representations of uncertain factors such as weather changes, temporary activities in the park, equipment failures, etc.) and the first navigation feature vector fused in the first derivative fusion channel.

[0058] For example, the forward disturbance feature may be a high-dimensional vector containing many dimensions and complex information, such as the probability distribution of different weather conditions, the scale and impact range of temporary activities, etc. The first navigation feature vector may contain the mean traffic flow of different sections of the main road, the access rules of important buildings, etc.

[0059] Through a series of operations and feature screening, these features are compressed. Principal component analysis (PCA) or autoencoder techniques may be used to remove redundant information in the forward disturbance feature and the first navigation feature vector, extract the most critical common features, and generate the first decision factor feature. This first decision factor feature is a relatively low-dimensional vector that contains key information about the main navigation and potential disturbances. The generated first decision factor features are loaded into X derivative transfer trajectories respectively. Assuming X=3, these three derivative transfer trajectories correspond to the requirements of visitor parking, employee restaurant location, and special channel for cargo transportation.

[0060] For the derived transfer trajectory of the visitor parking, the server fuses the corresponding second navigation feature vector in the second derived fusion channel. This second navigation feature vector contains specific visitor parking demand information such as the number of vacancies in each parking lot, parking fees, and distance to the destination. The server processes the first decision factor feature and the second navigation feature vector of the visitor parking.

[0061] First, the first decision factor feature may be dimensionally expanded through a dimensionality-enhancing layer so that it can interact with the second navigation feature vector in the same dimensional space. Then, through the self-focus association process, the server self-associates and strengthens the expanded first decision factor feature to generate a more representative intermediate feature.

[0062] Next, in the second interactive focus association branch, the intermediate feature is interactively focused and associated with the second navigation feature vector of the visitor parking, for example, the correlation and similarity between the two are calculated or weighted fusion of the features is performed to generate an interactive focus association vector.

[0063] Based on this interactive focus correlation vector, the server further generates a second decision factor feature for the visitor parking derived transfer trajectory through a series of transformations and calculations. This second decision factor feature fully considers the basic situation of the main navigation, potential disturbances, and the specific needs of visitor parking. The number of decision fields is the same as that of the forward disturbance feature to ensure the integrity and accuracy of the information.

[0064] The server performs the same operation for the derived delivery trajectory of the employee restaurant location and the derived delivery trajectory of the special channel for cargo transportation.

[0065] For the trajectory of the staff restaurant location, the second navigation feature vector may include the restaurant's business hours, dish features, seat availability, etc. After performing dimensionality increase processing, self-focus association processing, and interactive focus association processing with the first decision factor feature, the second decision factor feature of the staff restaurant location trajectory is generated.

[0066] For the trajectory of the special channel for cargo transportation, the second navigation feature vector may involve the opening time of the special channel, the type and size restrictions of the cargo allowed to pass, etc. After the same processing flow, the second decision factor feature of the trajectory of the special channel for cargo transportation is generated.

[0067] Finally, the server generates corresponding navigation optimization decision results according to the three second decision factor characteristics.

[0068] For the decision result of the visitor parking and arriving at the central square, the server may comprehensively consider and recommend a parking lot that is close to the visitor's current location, with reasonable fees and more vacancies, plan a specific route from the parking lot to the central square that avoids congestion and construction areas, and estimate the arrival time based on the current traffic conditions and possible changes. At the same time, the server will also consider the impact of weather changes on parking and walking, such as providing visitors with a covered walking route on rainy days.

[0069] For the decision result of employees going from the restaurant to the central square, the server may choose the best route that avoids crowded areas and construction areas based on the current restaurant location of the employee. If the restaurant is far from the central square, it will also consider providing a rest point or transfer point in the middle. The decision result will also take into account the restaurant's business hours to ensure that employees have enough time to eat before going to the central square.

[0070] For the decision result that the goods arrive at the central square after being transported from the special channel, the server will plan a channel and route that meets the cargo transportation restrictions (such as cargo type, size, weight, etc.). Taking into account the opening time of the special channel and possible equipment failures and other disturbance factors, the transportation time is estimated, and solutions to possible obstacles are formulated in advance, such as alternative channels or temporary adjustments to the transportation method.

[0071] As a result, accurate and practical navigation optimization decision results that take into account various uncertain factors are generated for different navigation needs in smart parks, improving the quality and efficiency of park navigation services.

[0072] In a possible implementation, step S141 includes:

[0073] Step S1411, loading the forward perturbation feature into the reverse derivative transfer subnetwork. The reverse derivative transfer subnetwork includes a perturbation estimation unit. The perturbation estimation unit includes a dimension reduction subunit, a derivative transfer subunit and a dimension increase subunit. The loading of the derivative transfer subunit is the generation of the dimension reduction subunit, and the loading of the dimension increase subunit is the generation of the derivative transfer subunit. The dimension reduction subunit includes a first self-focusing association branch and a first interactive focusing association branch.

[0074] Step S1412: In the first self-focus association branch, the forward disturbance feature is subjected to self-focus association processing to generate a first self-focus association vector, and in the first interactive focus association branch, the first self-focus association vector and the fused first navigation feature vector are subjected to interactive focus association processing to generate a first decision factor feature obtained in the first derivative fusion channel. The number of decision fields of the first decision factor feature is less than the number of decision fields of the forward disturbance feature.

[0075] The method further includes: copying the first decision factor features according to the derivative transfer subunit, and distributing the copied X first decision factor features to the dimension-increasing subunit for outputting X second decision factor features in the second derivative fusion channel.

[0076] In this embodiment, after receiving the forward disturbance feature and the first navigation feature vector fused in the first derivative fusion channel, a key processing step is prepared to generate an accurate and effective navigation decision.

[0077] The server first loads the forward disturbance feature into the disturbance estimation unit of the reverse derivative transfer subnetwork. This forward disturbance feature is an extremely rich and complex data set. For example, the forward disturbance feature contains detailed information about weather changes, such as the probability distribution of heavy rain in different time periods, the intensity change trend of strong winds, and the range of temperature fluctuations; it also covers many details of temporary activities in the park, such as the specific scale of the event, the estimated number of participants, the demarcation of the activity area, and the changes in the flow of people and traffic that the event may bring; in addition, relevant information about equipment failures is also included, such as the fault type (short circuit, mechanical failure, etc.) of various types of equipment (such as elevators, monitoring equipment, lighting systems, etc.), the scope of the impact of the fault (whether it affects the passage of a certain area, whether it affects the use of specific facilities).

[0078] The perturbation estimation unit in this reverse derivative transfer subnetwork is a specially designed module consisting of a dimension reduction subunit, a derivative transfer subunit, and a dimension increase subunit. Each subunit has a unique and critical function, and they work together to achieve efficient feature processing and decision generation.

[0079] First, enter the first autofocus correlation branch of the dimension reduction subunit. In this first autofocus correlation branch, the server begins to perform in-depth autofocus correlation processing on the forward disturbance feature. The server carefully analyzes the intrinsic relationship and autocorrelation between each element in the forward disturbance feature through a series of mathematical operations and algorithmic logic.

[0080] Taking weather change information as an example, the server may find that the probability of heavy rain and the intensity of strong winds are correlated under certain conditions. When the probability of heavy rain is high, the intensity of strong winds tends to be relatively weak, and vice versa. Through this autocorrelation analysis, the server is able to extract these key intrinsic connections and integrate them into a more representative feature representation, thereby generating the first self-focused correlation vector.

[0081] For information about temporary events in the park, the server may find that there is a positive correlation between the scale of the event and the number of participants, and that there is also a certain correlation between the demarcation of the activity area and the traffic flow of the surrounding roads. By digging deeper into these autocorrelated information, the server can further streamline and optimize the forward disturbance features, making it more focused on key internal correlation patterns.

[0082] For equipment failure information, the server may find that certain types of equipment failures (such as elevator failures) tend to cause changes in the flow of people in a specific area (people choose to avoid the floor or area where the failed elevator is located). Through this self-focused association processing, the server is able to extract these potential patterns that have an important impact on navigation decisions.

[0083] After the above processing, the generated first self-focus correlation vector has a lower dimension than the original forward disturbance feature, but it highlights the key self-correlation information in the forward disturbance feature. It is no longer a simple pile of raw data, but the essence after careful refinement and screening.

[0084] Next, the server transmits the first self-focus correlation vector to the first interactive focus correlation branch of the dimension reduction subunit, in which the first self-focus correlation vector is interactively focused and correlated with the fused first navigation feature vector.

[0085] The first navigation feature vector also contains rich and critical information. For example, real-time traffic flow data on the main roads in the park, including detailed information such as traffic volume and speed of different sections at different time periods; the current opening status of important buildings, such as whether they are undergoing maintenance or whether access is restricted; the use of public facilities, such as the saturation of a rest area and the length of the queue in a public toilet.

[0086] The server uses complex calculations and logical judgments to deeply analyze the interaction between the first self-focusing association vector and the first navigation feature vector. For example, when there is a large-scale event in the park (which is reflected in the forward disturbance feature) and the traffic flow on a main road is large (which is part of the first navigation feature vector), the server will comprehensively consider these two factors and determine their joint impact on the navigation decision.

[0087] Assume that in a specific scenario, a large-scale job fair is being held in the park, which leads to an increase in the density of people in a certain area (forward disturbance feature), and at the same time, a nearby main road has a decrease in traffic due to construction (the first navigation feature vector). The server will use interactive focus association processing to comprehensively consider these two factors and determine whether the navigation route needs to be adjusted to guide people to avoid congested areas and choose a more suitable road to their destination.

[0088] Through this sophisticated interactive focus correlation processing, the server can fully fuse the potential uncertainty represented by the forward disturbance feature and the main navigation key information embodied by the first navigation feature vector to generate the first decision factor feature obtained in the first derivative fusion channel.

[0089] This first decision factor feature is a highly concentrated and refined information set that takes into account various complex factors. However, since the processing mainly extracts key common and related information and removes some relatively minor and redundant parts, the number of decision fields of the first decision factor feature is smaller than the number of decision fields of the initial forward disturbance feature. However, the information it contains is extremely critical and effective for the subsequent navigation decision generation.

[0090] After the generation of the first decision factor feature is completed, the server copies it according to the derived transfer subunit. Assume that copies need to be generated for X=5 different navigation demand scenarios, corresponding to visitor parking, employees going to different work areas, and cargo transportation choosing different routes.

[0091] For the visitor parking demand scenario, the server needs to consider factors such as the real-time change of parking spaces, the dynamic adjustment of parking fees, and the distance of the visitor's destination. Therefore, the server copies the first decision factor features to prepare for the fusion of subordinate multimodal data related to these specific visitor parking.

[0092] For the scenario where employees need to go to different work areas, the server needs to consider factors such as employee work schedules, access rights to different work areas, and the density of people in each area. Similarly, the copied first decision factor features will be combined with the subordinate multimodal data related to these specific employee navigations.

[0093] In the cargo transportation demand scenario, the server needs to consider many factors such as the type, size, weight of the cargo, the specifications of the transport vehicle, the carrying capacity of the road, the restrictions of special channels, etc. The copied first decision factor features will be fused with the subordinate multimodal data related to these cargo transportation in the future.

[0094] For each copy operation, the server ensures the accuracy and integrity of the data to ensure the reliability of subsequent processing.

[0095] After completing the copying, the server distributes the five copies of the first decision factor features to the dimension-increasing sub-units, which receive the copies and make preliminary preparations for generating the corresponding second decision factor features in the second derivative fusion channel.

[0096] Taking the second derivative fusion channel of visitor parking demand as an example, the copy of the first decision factor feature received by the dimension-upgrading subunit will further fuse and process the subordinate multimodal data related to visitor parking.

[0097] These dependent multimodal data may include details such as the real-time number of vacancies in various parking lots, changes in charging standards for different parking lots, and walking distances and route complexity between parking lots and visitor destinations.

[0098] The server first performs appropriate dimensionality upscaling on the feature copy of the first decision factor so that it can be effectively integrated and interacted with these rich subordinate multimodal data in the same dimensional space.

[0099] Then, through calculation and association operations, the server deeply integrates the upgraded copy of the first decision factor feature with the subordinate multimodal data of visitor parking. For example, if the first decision factor feature indicates that there is a certain amount of traffic congestion in the current park (due to temporary activities or weather reasons), and a parking lot is far from the visitor's destination, but it has more vacancies and lower fees, the server may consider these factors comprehensively, recommend this parking lot to the visitor, and plan a walking route that relatively avoids the congested area.

[0100] Through such fusion and processing, the server finally generates the second decision factor feature for the visitor's parking needs. This feature fully considers the forward disturbance feature, the first navigation feature vector and the specific subordinate multimodal data of the visitor's parking, and can provide visitors with accurate, efficient and uncertain navigation optimization decisions from selecting a parking lot to finally reaching the destination.

[0101] For scenarios where employees need to go to different work areas, the subordinate multimodal data may include information such as the current task schedule of each work area, whether there is urgent work that needs to be prioritized, and the personal preferences of employees (such as whether they prefer to take the stairs or the elevator).

[0102] The server will merge and process the copy of the first decision factor feature with the employee-specific subordinate multimodal data in the dimension-upgrading subunit. For example, if the first decision factor feature shows that a public facility in the park is under maintenance, causing some passages to be temporarily closed, and an employee needs to reach a high-rise work area as soon as possible, the server may take these factors into consideration and plan a route for the employee to avoid the maintenance area and reach the destination quickly, and recommend the use of an elevator to save time.

[0103] For cargo transportation demand scenarios, subordinate multimodal data may involve special transportation requirements for cargo (such as fragile goods requiring a smooth transportation route), fuel consumption of transportation vehicles, road slope and curve restrictions, and other information.

[0104] The server copies the first decision factor feature in the dimension-increasing subunit and fuses it with the subordinate multimodal data related to the transportation of these goods. For example, if the first decision factor feature indicates that the current weather is bad (such as strong winds and heavy rains), and the goods are of a type that is sensitive to the environment, the server may choose a sheltered and well-maintained transportation route, and adjust the transportation speed and vehicle configuration to ensure the safety and on-time arrival of the goods.

[0105] By finely processing the copy of the first decision factor feature and different subordinate multimodal data in the dimensionality-upgrading sub-unit, the server generates highly customized and optimized second decision factor features for each specific navigation demand scenario, laying a solid foundation for the ultimate generation of accurate and practical navigation optimization decision results.

[0106] During the entire processing process, the server always maintains efficient and accurate calculations, and continuously optimizes algorithms and models to cope with the increasingly complex and diverse navigation needs in smart parks, providing users with better quality and more reliable navigation services.

[0107] Then, for another navigation requirement scenario, such as an employee needs to go from the current location to the gym in the park, the server will repeat the above process again.

[0108] First, the server obtains relevant forward disturbance features, such as a fire drill being conducted in a certain area of ​​the park, which may cause some passages to be temporarily closed; as well as the first navigation feature vector, such as the opening hours of the gym and the flow of people on the main roads in the current park.

[0109] Then, it is processed in the dimension reduction subunit. In the first self-focus association branch, the forward disturbance feature is analyzed, and it is found that the area of ​​the fire drill intersects with the regular route to the gym, thereby generating the first self-focus association vector. In the first interactive focus association branch, the first self-focus association vector is interactively processed with the first navigation feature vector, and the first decision factor feature is generated by comprehensively considering the opening hours of the gym and the road traffic.

[0110] Next, the first decision factor feature is copied according to the derivative transfer subunit and distributed to the dimension-upgrading subunit. In the dimension-upgrading subunit, it is fused with the subordinate multimodal data of employees going to the gym, which may include employees' exercise habits (such as the length of time they like to warm up before exercise) and the current crowding level in the gym.

[0111] Through a series of calculations and processing, the second decision factor feature for employees going to the gym is generated. Based on this feature, the server may plan an optimal route for the employee that avoids the fire drill area, meets his exercise habits, and avoids crowding in the gym.

[0112] Consider the scenario of cargo transportation to a specific warehouse. The forward disturbance features may include sudden power failure areas in the park, failure of the lighting system on some roads, etc. The first navigation feature vector may cover the location of the warehouse, the regular use of the cargo transportation channel, etc.

[0113] After the first decision factor feature is generated in the dimension reduction subunit, it is copied and distributed to the dimension increase subunit. It is fused with the subordinate multimodal data of the goods transported to the warehouse, such as the urgency of the goods, the warehouse's receiving time limit, etc.

[0114] The resulting second decision factor features can help the server plan a safe, efficient route for cargo transportation that meets various constraints, such as avoiding power failure areas and poorly lit roads, while ensuring arrival within the warehouse receipt time.

[0115] For the scenario where visitors need to go to the exhibition hall in the park, the forward disturbance characteristics may include the crowd gathering caused by the commercial exhibition activities currently being held in the park, the temporarily added security measures, etc.; the first navigation feature vector may include the current exhibition content of the exhibition hall, the traffic capacity of the surrounding roads, etc.

[0116] After being processed by the dimensionality reduction sub-unit and the dimensionality increase sub-unit, the second decision factor feature is generated by combining the subordinate multimodal data of visitors to the exhibition hall, such as the purpose of the visit (whether to participate in specific activities or to visit freely), the interest preference in the exhibition content, etc.

[0117] Based on this, the server plans a tour route that best suits the needs of visitors. For example, it prioritizes guiding visitors who are interested in specific activities to the activity area, while providing free visitors with a route that allows them to fully appreciate the exhibition and avoid crowds.

[0118] When processing the scenario where employees go to a training center within the park, the forward disturbance features may include ongoing greening construction near the training center, areas with unstable wireless network signals within the park, etc.; the first navigation feature vector may include the start time of the training course, the shortest distance between the employee's current location and the training center, etc.

[0119] After feature processing and fusion, the second decision factor feature that meets the needs of employees to go to the training center is generated. Based on this feature, the server may choose a route for employees to reach the training center as quickly as possible while reducing the impact of construction and network signals.

[0120] When goods need to be transported to a production workshop within the park, the forward disturbance characteristics may include changes in the temporary material storage area due to ongoing equipment maintenance in the workshop, temporary traffic control on roads around the workshop, etc.; the first navigation feature vector may include the urgency of the material demand in the production workshop, the carrying capacity of the transportation channel, etc.

[0121] Thus, a second decision factor feature that fits the needs of transporting goods to the production workshop is generated. The server can then plan an efficient transportation route that meets the production needs of the workshop and avoids temporary obstacles.

[0122] In short, in various navigation demand scenarios of smart parks, through the above feature processing and fusion process, decision factor features that adapt to different needs are continuously generated, providing accurate, efficient and reliable optimization decision results for each specific navigation task, thereby improving the traffic efficiency and user experience within the park.

[0123] Wherein, step S1412 includes:

[0124] Step S1412-1, calculate the first self-focus association vector and the retrieval index array to generate a retrieval vector, calculate the first self-focus association vector and the primary key index array to generate a primary key vector, calculate the first navigation feature vector and the numerical index array to generate a numerical vector.

[0125] Step S1412-2, generate self-focus weight information based on the retrieval vector and the primary key vector, compress the self-focus weight information according to the number of fields of the primary key vector, perform regularization conversion on the compressed self-focus weight information to generate regularization conversion features, and calculate the regularization conversion features with the numerical vector to generate the first decision factor features.

[0126] In the navigation decision-making process of the smart park, the server performs a series of complex and sophisticated calculations and processing operations in the first interactive focus association branch.

[0127] The server first obtains the first self-focus correlation vector, which is obtained by self-focus correlation processing of the forward disturbance feature and contains key information about the potential disturbance. For example, it may reflect the slippery degree of the road surface in a certain area of ​​the park due to sudden weather, or the changing trend of the flow of people in a specific section of the road due to temporary activities.

[0128] At the same time, the server also obtains the first navigation feature vector fused in the first derivative fusion channel, which covers important information of the main navigation of the park, such as the capacity of the main roads, the location and accessibility of important buildings, etc.

[0129] The server calculates the first self-focus correlation vector and the retrieval index array. Assume that the retrieval index array contains a series of indicators related to potential disturbance characteristics, such as the impact weight of different weather conditions on road safety, the impact degree of different changes in passenger flow on traffic efficiency, etc. Through this calculation, the server generates a retrieval vector.

[0130] For example, if the first self-focus correlation vector indicates that the current weather is light rain and the passenger flow on a certain road section has slightly increased, after calculation with the retrieval index array, the generated retrieval vector may represent the quantitative value of the degree of reduction in safety of the road section and the degree of impact on traffic efficiency.

[0131] Next, the server calculates the first self-focused correlation vector and the primary key indicator array to generate a primary key vector. The primary key indicator array may include some key indicators that can identify the main disturbance factors, such as the threshold at which a specific weather condition becomes a dominant influencing factor, the change in passenger flow reaches a critical value that affects road selection, etc.

[0132] The calculated primary key vector can clarify which factors play a dominant role in the current first self-focus correlation vector. For example, if the calculation results show that the impact of light rain on the road is the main factor in the current situation, the primary key vector will highlight this information.

[0133] At the same time, the server calculates the first navigation feature vector and the numerical indicator array to generate a numerical vector. The numerical indicator array may include specific numerical indicators related to the main navigation feature, such as the traffic capacity value corresponding to the width of the road, the navigation priority value corresponding to the distance to important buildings, etc.

[0134] Through this calculation, the generated numerical vector can convert the information in the first navigation feature vector into a specific quantifiable and comparable numerical form.

[0135] Based on the generated search vector and key vector, the server generates self-focus weight information, which reflects the relative importance and influence of different factors in the current navigation decision.

[0136] For example, if the retrieval vector indicates that a road section has a high degree of safety degradation, and the primary key vector indicates that weather is the main influencing factor, the self-focused weight information may give a higher weight to weather-related factors to highlight their importance in the decision.

[0137] Next, the server compresses the self-focus weight information according to the number of fields in the primary key vector. Assuming that the number of fields in the primary key vector is large, in order to improve computing efficiency and reduce data redundancy, the server will compress the self-focus weight information according to certain algorithm rules.

[0138] For example, through principal component analysis or feature selection methods, the most critical and representative weight information is retained, and some relatively minor and less relevant parts are removed.

[0139] The server performs regular transformation on the compressed autofocus weight information, which may include mapping the weight value to a specific interval or performing standardization to ensure that the weight information has a uniform measurement and comparability.

[0140] The generated regularized transformation features are more in line with the requirements of subsequent calculations and decision-making, and have better stability and versatility.

[0141] Finally, the server calculates the regularized transformation features and the previously generated numerical vectors. Through complex mathematical operations such as multiplication and addition, the numerical quantization information of the main navigation features and the processed autofocus weight information are comprehensively considered.

[0142] For example, if the numerical vector indicates that a road has a strong capacity, but the regularized conversion feature shows that its weight should be appropriately reduced due to weather reasons, the final calculation result will combine these two factors to generate a more accurate and balanced first decision factor feature.

[0143] This first decision factor feature can fully reflect the complex interactive relationship between the forward disturbance feature and the first navigation feature vector, providing a key basis for subsequent navigation decisions.

[0144] Assume that in another specific scenario, the first self-focus association vector indicates that strong sunlight in a certain period of time in the park causes obstructed vision in some areas, and the first navigation feature vector includes the number of bends and slope information of different roads leading to multiple destinations.

[0145] Calculation and processing are performed according to the above steps. The retrieval vector generated by the retrieval index array calculation may reflect the specific impact of the obstructed view on driving safety; the primary key vector generated by the primary key index array calculation may determine the dominant role of the strong sunlight factor in the current situation; the numerical vector generated by the numerical index array calculation converts the number of bends and slope information of the road into quantifiable values.

[0146] The self-focus weight information generated based on the retrieval vector and the primary key vector will focus on the weight of the factor of strong sunlight; after compression and regularization conversion, it is calculated with the numerical vector. The final generated first decision factor feature can fully weigh the impact of strong sunlight on the basis of considering the road conditions itself, providing a more accurate basis for navigation decisions.

[0147] When processing the navigation needs of employees from the office area to the meeting room, the first self-focus association vector may include the frequency of use of the elevator in the area during the current time period, and the first navigation feature vector includes the walking distance of different routes and the number of public areas passed through.

[0148] Through a series of calculations, the server generates a search vector that reflects the impact of elevator use on traffic efficiency, a key vector that determines the dominant factor, and a numerical vector that quantifies the walking distance and the number of public areas. It then generates reasonable self-focusing weight information, which is processed and calculated with the numerical vector to obtain the first decision factor feature, and plans an optimal route for employees that takes into account both the busy situation of the elevator and the comprehensive walking conditions.

[0149] For the scenario where goods are transported from a warehouse to a production workshop, the first self-focus association vector may be temporary traffic control information on a certain section of the road during transportation, and the first navigation feature vector includes the weight of the goods and the carrying capacity of the vehicle.

[0150] The server also executes the above calculation process to generate corresponding vectors and features, and finally obtains the first decision factor feature that comprehensively considers traffic control and cargo transportation conditions, thereby planning a suitable transportation route.

[0151] As a result, accurate and effective first decision factor characteristics can be generated in various complex navigation demand scenarios, providing solid support for navigation optimization decisions in smart parks.

[0152] In a possible implementation, the X derived transfer trajectories include a target derived transfer trajectory. Step S142 includes:

[0153] Step S1421, loading the first decision factor feature into the dimension-raising subunit. The dimension-raising subunit includes a second self-focusing association branch and a second interactive focusing association branch corresponding to the target derived transfer trajectory.

[0154] Step S1422: In the second self-focus correlation branch, perform self-focus correlation processing on the first decision factor feature to generate a second self-focus correlation vector.

[0155] Step S1423, in the second interactive focus association branch, interactive focus association processing is performed on the second navigation feature vector fused with the second self-focus association vector and the target derived transfer trajectory to generate an interactive focus association vector, and a second decision factor feature obtained by generating the target derived transfer trajectory based on the interactive focus association vector. The number of decision fields of the second decision factor feature is equal to the number of decision fields of the forward disturbance feature.

[0156] For example, suppose there are X=3 derived transfer trajectories, which are the trajectory of visitors going to the park attractions after parking, the trajectory of employees going from the office area to the cafeteria, and the trajectory of cargo transport vehicles going to the warehouse. Here, the target derived transfer trajectory of visitors going to the park attractions after parking is taken as an example for detailed description.

[0157] The server first loads the first decision factor feature into the dimension-upgrading subunit. This first decision factor feature is generated through a series of complex processing and fusion. It integrates the key information of the forward disturbance feature and the first navigation feature vector, and has been compressed and screened. It is a relatively refined feature that contains important navigation decision clues.

[0158] The dimension-increasing subunit includes a second self-focusing association branch and a second interactive focusing association branch corresponding to the target-derived transfer trajectory.

[0159] In the second self-focus association branch, the server performs self-focus association processing on the first decision factor features that are loaded. For example, the first decision factor features may contain information such as the traffic congestion of the main roads in the current park and the impact of possible weather changes on walking. Through self-focus association processing, the server will deeply analyze the inherent correlation and common patterns between these information.

[0160] Assuming that the first decision factor feature indicates that road congestion is mainly concentrated in the east of the campus, and weather changes may cause slippery roads in the west, through self-focused correlation processing, the server will generate a second self-focused correlation vector that highlights these key correlations. This vector is more focused on the internal features and patterns that are directly related to the target-derived transfer trajectory.

[0161] Next, in the second interactive focus association branch, the server performs interactive focus association processing on the second self-focus association vector just generated and the second navigation feature vector fused with the target-derived transfer trajectory.

[0162] For the target-derived transfer trajectory of visitors heading to the park attractions after parking, the fused second navigation feature vector may contain information such as the opening hours of each attraction, the current visitor flow, and the walking distance between attractions.

[0163] The server will comprehensively consider the general navigation clues represented by the second self-focus association vector and the scenic spot related information unique to the second navigation feature vector. For example, if the second self-focus association vector indicates that the current road condition is poor, and the tourist flow of a certain scenic spot is large and the distance is far, the server will generate an interactive focus association vector through complex calculations and association analysis.

[0164] This interactive focus correlation vector fully integrates the road conditions, scenic spots and the mutual influence between them. Based on this interactive focus correlation vector, the server further processes and transforms it to generate the second decision factor feature obtained by the target derived transfer trajectory.

[0165] Importantly, the number of decision fields in this second decision factor feature is equal to the number of decision fields in the initial forward perturbation feature. This means that it contains rich and comprehensive information, and can provide visitors with detailed, accurate, and complex navigation decision suggestions from the parking lot to the park attractions.

[0166] For example, it may include recommending specific walking routes, avoiding congested and slippery roads, selecting attractions with relatively small tourist traffic and suitable opening hours, estimating the time required to reach each attraction, and making dynamic adjustments based on real-time changes.

[0167] For the target-derived transfer trajectory of employees from the office area to the cafeteria, the second navigation feature vector may contain information such as the queue situation at different windows of the cafeteria, the currently launched special dishes, and the employees' dietary preferences.

[0168] The server also performs processing in the dimension-raising subunit. In the second self-focus association branch, the first decision factor feature is processed to generate a second self-focus association vector reflecting the road from the office area to the cafeteria. Then, in the second interactive focus association branch, it is interactively focused and associated with the second navigation feature vector related to the cafeteria to generate an interactive focus association vector for employees going to the cafeteria, and finally the second decision factor feature is obtained.

[0169] This feature might recommend windows with shorter queues for employees, select appropriate dishes based on employee dietary preferences, and plan a route that gets employees there quickly and avoids areas of possible congestion.

[0170] For the target derived delivery trajectory of the cargo transport vehicle heading to the warehouse, the second navigation feature vector may involve information such as the storage space of the warehouse, the loading and unloading time of the cargo, and the restriction conditions of the transport vehicle.

[0171] The server generates the second decision factor features for cargo transportation through processing of the dimension-upgrading sub-units. This may include selecting the optimal transportation route, avoiding construction areas around the warehouse, arranging the storage order of the goods according to the warehouse storage space, and estimating the time required for the entire transportation and loading and unloading process.

[0172] Therefore, through careful processing and analysis of each target-derived transfer trajectory, accurate, comprehensive and adaptable second decision factor characteristics are generated in the upgraded sub-unit, providing high-quality decision support for various navigation needs within the smart park and improving the traffic efficiency and user experience within the park.

[0173] In a possible implementation, the reverse derivative transfer subnetwork includes Y disturbance estimation units, where Y is a positive integer. Each disturbance estimation unit is used to fuse the first navigation feature vector in a first derivative fusion channel, and to fuse a second navigation feature vector for each derivative transfer trajectory in a second derivative fusion channel. The Y disturbance estimation units include a disturbance estimation unit, where a is a positive integer not greater than Y.

[0174] Step S140 may further include:

[0175] Step A110, loading the feature to be processed into the disturbance estimation unit, using the disturbance estimation unit to compress the feature to be processed and the first navigation feature vector fused in the first derivative fusion channel, generating X first decision factor features obtained by the disturbance estimation unit, and loading the X first decision factor features obtained by the disturbance estimation unit into the X derivative transfer trajectories obtained by the disturbance estimation unit. One first decision factor feature is loaded into one derivative transfer trajectory. If a is 1, the feature to be processed is the forward disturbance feature, and the X first decision factor features are all the same feature. If a is greater than 1, the feature to be processed includes X second decision factor features obtained by the disturbance estimation unit, and one first decision factor feature is determined by a second decision factor feature obtained by the disturbance estimation unit and the first navigation feature vector, the disturbance estimation unit is the previous disturbance estimation unit of the disturbance estimation unit among the Y disturbance estimation units, and the generation of the disturbance estimation unit is the loading of the disturbance estimation unit.

[0176] Step A120: In the second derivative fusion channel, a second navigation feature vector is fused for each derived transfer trajectory of the disturbance estimation unit respectively; in each derived transfer trajectory, the first decision factor feature obtained by the disturbance estimation unit and the fused second navigation feature vector are respectively subjected to dimensionality increase processing to generate X second decision factor features obtained by the disturbance estimation unit.

[0177] Step A130: if a is Y, then generating navigation optimization decision results corresponding to the X derived transfer trajectories of the disturbance estimation unit respectively according to the X second decision factor features obtained by the disturbance estimation unit.

[0178] In this embodiment, in the navigation decision system of the smart park, the server is responsible for processing complex navigation data and generating optimized decision results. Assuming that the reverse derivative transfer subnetwork includes Y=5 disturbance estimation units, the processing process of each unit is viewed in turn.

[0179] For the first disturbance estimation unit (a=1), the server loads the forward disturbance feature as a feature to be processed into this unit. The forward disturbance feature may include sudden severe weather information in the park, temporary equipment maintenance notices, etc. At the same time, the first navigation feature vector is also introduced, which covers basic navigation information such as the layout of the main roads in the park and the location of important places.

[0180] The disturbance estimation unit compresses the forward disturbance feature and the first navigation feature vector. For example, through a series of mathematical operations and feature screening, some relatively minor or repeated information is removed, and the most critical and interrelated parts of the two are extracted, thereby generating X=3 identical first decision factor features. These three first decision factor features are loaded into the processing flows corresponding to the three derived transfer trajectories of visitor parking, employee restaurant location, and special channel for cargo transportation.

[0181] Then, in the second derivative fusion channel, the corresponding second navigation feature vector is fused for each derivative transfer trajectory. For the visitor parking trajectory, the second navigation feature vector may include the real-time number of vacancies and charging standards of each parking lot; for the employee restaurant location trajectory, it may include the restaurant's current queue situation, special dishes, etc.; for the special channel for cargo transportation, it may include the temporary traffic restrictions of the channel, the busyness of the cargo loading and unloading point, etc.

[0182] Then, in each derived transfer trajectory, the first decision factor feature and the fused second navigation feature vector are processed for dimensionality increase. By increasing the dimension and richness of the features and taking various factors into account more comprehensively, three second decision factor features are generated.

[0183] For the second disturbance estimation unit (a=2), the features to be processed are no longer the forward disturbance features, but the three second decision factor features obtained by the first disturbance estimation unit. These features have integrated certain navigation information and disturbance factors.

[0184] The disturbance estimation unit combines a second decision factor feature obtained by the previous unit with the first navigation feature vector, performs new compression processing, and generates different first decision factor features for each derived transfer trajectory.

[0185] Similarly, in the second derivative fusion channel, the corresponding second navigation feature vector is fused for each trajectory, and the newly generated first decision factor feature and second navigation feature vector are subjected to dimensionality upgrading processing to generate a new second decision factor feature.

[0186] This processing is carried out in the third and fourth disturbance estimation units respectively.

[0187] When it reaches the fifth disturbance estimation unit (a=5), it has undergone multiple feature fusions and processing to obtain a more accurate and comprehensive second decision factor feature.

[0188] Based on these three final second decision factor features, the server generates a specific navigation optimization decision result for each derived transfer trajectory.

[0189] Taking the visitor parking trajectory as an example, if the final second decision factor feature indicates that the current traffic in the park is relatively congested and the parking spaces near the visitor's destination are tight, the server may recommend a parking lot that is a little farther away but has relatively smooth traffic and vacancies, and plan a route to avoid congested sections, while estimating the arrival time and possible situations.

[0190] For the trajectory of the employee restaurant location, if the second decision factor feature shows that the current restaurant has a long queue and the employees have tight working hours, the server may recommend other nearby restaurants with less customer traffic and plan the fastest route to get there.

[0191] For the trajectory of a special channel for cargo transportation, if the second decision factor feature indicates that the channel is limited but the cargo transportation time is urgent, the server may re-plan an alternative route to ensure that the cargo is delivered on time, while considering transportation costs and safety factors.

[0192] Through the sequential processing and feature fusion of these five disturbance estimation units, the server can fully consider various complex factors and uncertainties, provide accurate, practical and personalized optimization decision results for different navigation needs in the smart park, and improve the traffic efficiency and service quality within the park.

[0193] In a possible implementation, step S142 further includes: loading the X second decision factor features into a mapping subnetwork of the target park navigation decision network, in the mapping subnetwork, mapping and outputting the implicit perturbation features respectively according to the X second decision factor features, generating X mapping knowledge point data, and generating X navigation optimization decision results based on the X mapping knowledge point data.

[0194] In the navigation optimization decision-making process of the smart park, the server undertakes the important task of generating accurate navigation results based on the characteristics of the second decision factor. Assume that there are X=3 derivative transfer trajectories, corresponding to visitor parking, employee restaurant location and special channel for cargo transportation.

[0195] The server obtains the three second decision factor features and prepares to load them into the mapping subnetwork of the target campus navigation decision network.

[0196] The first is the second decision factor feature of visitor parking. This feature includes information about the visitor's current location, parking lot information (such as the number of vacancies, distance, fees, etc.), as well as comprehensive factors such as traffic conditions in the park and possible weather impacts. The server loads it into the mapping subnetwork.

[0197] In the mapping subnetwork, the implicit perturbation feature is mapped and output based on the second decision factor feature. The implicit perturbation feature may contain some potential influencing factors that are not clearly expressed. For example, a temporary event to be held in a certain area of ​​the park may attract more people, thereby affecting the traffic conditions of the surrounding roads. Through the mapping process, a mapping knowledge point data related to visitor parking is generated.

[0198] This mapping knowledge point data may specifically indicate the predicted availability of a particular parking lot in the future, as well as a detailed plan for the best route from the visitor's current location to the parking lot. Based on this mapping knowledge point data, the server further generates navigation optimization decision results for visitor parking. For example, the server may decide to recommend visitors to go to parking lot A, which is about 500 meters away from their current location, because according to current and predicted conditions, the parking lot not only has vacancies, but the route to it is not expected to be severely congested in the next 30 minutes. At the same time, the server will also provide detailed walking instructions to inform visitors of the shortest path from parking lot A to their destination and estimate the time required for the entire trip.

[0199] Next is the second decision factor feature of the employee restaurant location. This feature integrates the current location of the employee, the real-time status of the restaurant (such as the number of people in line, seat availability, food supply, etc.), the employee's work schedule, and the distribution of people in the park. The server inputs it into the mapping subnetwork.

[0200] The mapping subnetwork maps the implicit perturbation feature based on this feature and generates corresponding mapping knowledge point data. This may include the flow trend of people in the restaurant after the lunch rush, or information that a restaurant may adjust its menu due to a temporary shortage of ingredients. Based on these mapping knowledge point data, the server generates navigation optimization decision results for the location of the employee restaurant. For example, the server may recommend that employees go to restaurant B, which is closer to their work area, because the number of people queuing in restaurant B is relatively small at this time, and the dishes that employees like are available. At the same time, the server plans a walking route that avoids crowded areas to ensure that employees can quickly reach the restaurant and enjoy their meal time.

[0201] Finally, the second decision factor feature of the special channel for cargo transportation covers the type, size, weight of the cargo, the performance parameters of the transport vehicle, the real-time traffic conditions of the special channel (such as whether there is construction, traffic restrictions, etc.), and the distribution of other transportation tasks in the park. The server loads it into the mapping subnetwork.

[0202] The mapping subnetwork maps the implicit disturbance feature based on this feature to generate predictions about the traffic changes in the special channel in the future, or other potential factors that may affect the transportation of goods. Based on the generated mapping knowledge point data, the server generates navigation optimization decision results for special channels for cargo transportation. For example, the server may plan the transport vehicle to pass through special channel C, because according to the forecast, in the next 2 hours, although some sections of the channel are narrow, it can meet the traffic requirements of the transport vehicle, and the traffic flow is relatively small, which helps to reduce transportation time and reduce the risk of damage to goods. At the same time, the server will also provide an alternative route plan to prevent the channel from being accidentally closed during transportation.

[0203] By processing these three second decision factor features separately in the mapping sub-network, the server generates accurate and practical navigation optimization decision results for each derived transfer trajectory that take into account various dynamic changes and potential impacts, effectively improving the quality and efficiency of different types of navigation services in the smart park.

[0204] At another specific moment, such as when a large-scale event is being held in the park, the situation is different.

[0205] For the second decision factor feature of visitor parking, due to the influence of the event, the vacancy situation of surrounding parking lots changes frequently, and more control measures have appeared on the roads. When processed in the mapping subnetwork, the generated mapping knowledge point data will focus more on the dynamics of parking lots around the event area and the changes in temporary traffic rules. Based on this, the navigation optimization decision result generated by the server may be to guide visitors to parking lot D, which is slightly farther away from the event area but has smoother traffic, and provide detailed detour instructions to avoid congested sections caused by the event.

[0206] For the second decision factor feature of the employee restaurant location, due to the influx of a large number of visitors, the restaurant's queues and food supply have changed. The mapping knowledge point data generated by the mapping subnetwork will reflect the impact of these changes on employees' dining choices. The navigation optimization decision result generated by the server may be to recommend employees to go to restaurant E, which is usually deserted but less affected during the event, and adjust the walking route to avoid the peak flow of people in the event area.

[0207] For the second decision factor feature of the special channel for cargo transportation, the traffic pressure of the special channel increases due to the increase in personnel and vehicles brought by the event. The mapping knowledge point data generated by the mapping subnetwork will contain the real-time congestion level of the channel and possible extension of the restriction information. The navigation optimization decision result generated by the server may be to adjust the transportation time, choose to use the special channel during the non-peak period after the event, or choose other temporarily open backup channels for cargo transportation.

[0208] Therefore, by accurately processing and analyzing the characteristics of the second decision factor in the mapping subnetwork, it is possible to generate the most optimized decision results for various navigation needs according to the ever-changing actual situation in the park, thereby ensuring the normal operation of the park and the convenient experience of users.

[0209] In a possible implementation, step S110 includes:

[0210] Step S111, loading the main navigation multimodal data into the target park navigation decision network. The target park navigation decision network includes a feature extraction subnetwork.

[0211] Step S112: disassemble the main navigation multimodal data according to the feature extraction subnetwork to generate a navigation node data set corresponding to the main navigation multimodal data. A navigation node data in the navigation node data set refers to the smallest navigation data segment generated after disassembling the smart park navigation data.

[0212] Step S113: extracting features from each navigation node data in the navigation node data set to generate a first navigation feature vector.

[0213] In the navigation optimization processing of the smart park, the server first conducts in-depth analysis and feature extraction of the main navigation multimodal data to generate the key first navigation feature vector.

[0214] The server loads the collected main navigation multimodal data into the target park navigation decision network. These main navigation multimodal data come from a wide range of sources and in various forms, such as real-time traffic flow data from the park road monitoring system, park road and building layout information provided by the geographic information system, and the opening hours and access rules of various important places.

[0215] The target park navigation decision network includes a special feature extraction sub-network, which has powerful data analysis and processing capabilities.

[0216] According to the function of the feature extraction sub-network, the server disassembles the main navigation multimodal data. For example, the road system of the park is disassembled into navigation nodes such as various road sections and intersections. Assuming there is a main road in the park, the server will disassemble it into multiple sections. The starting and end points, road width, allowed driving speed, etc. of each section are regarded as independent navigation node data. At the same time, for important buildings, such as office buildings, R&D centers, etc., their entrances, exits and connection points with surrounding roads are also disassembled into navigation nodes.

[0217] Through such a disassembly operation, a navigation node data set corresponding to the main navigation multimodal data is generated. Each navigation node data in this set is the smallest navigation data fragment generated after the smart park navigation data is disassembled.

[0218] Next, the server extracts features from each navigation node data in the navigation node data set.

[0219] For the navigation node data of a specific road section, the server may extract features such as average traffic volume, traffic volume during peak hours, road slope, whether there are traffic lights, etc. For the navigation node data of intersections, the features that may be extracted include turning restrictions at intersections, traffic volume from different directions, pedestrian crossing frequency, etc. For the navigation node data of important buildings, the features that may be extracted include the use nature of the building (such as office, production, warehousing, etc.), peak hours for people entering and leaving, whether there are special safety inspection requirements, etc.

[0220] By extracting and quantifying these rich and diverse features, the server converts each navigation node data into a feature vector with a specific numerical representation.

[0221] Then, the server further integrates and processes the feature vectors of all navigation node data, and may use weighted average, principal component analysis and other methods to merge these feature vectors into a unified first navigation feature vector that can fully reflect the key features of the main navigation multimodal data.

[0222] For example, in a large science and technology park, there is a main road that runs through the park. The server breaks down this main road into multiple road segment navigation nodes and extracts features such as traffic volume and road conditions for each road segment. At the same time, for the R&D buildings in the park, the server extracts features such as the entrance and exit locations and peak hours of personnel flow.

[0223] For a certain section of the main road, the server extracts that the average hourly traffic volume is 500 vehicles, the peak traffic volume reaches 800 vehicles, the road slope is 5%, there are traffic lights and the green light time accounts for 60%. These features are converted into corresponding values ​​to form the feature vector of the navigation node of the section.

[0224] For the navigation node of the R&D building, it is extracted that 9:00 to 10:00 a.m. and 5:00 to 6:00 p.m. every day are the peak hours for personnel to enter and exit. There are 4 entrances and exits, 2 of which are mainly used for employees to enter and 2 are mainly used for employees to leave. This information is also quantified into feature vectors.

[0225] Combining the feature vectors of all these navigation nodes, the server generates a first navigation feature vector that can fully reflect the main navigation situation of the park after complex calculations and integration. This vector contains key information such as the traffic capacity of roads and the flow of people in important buildings, providing a solid foundation for subsequent navigation decision optimization.

[0226] In another scenario, suppose that infrastructure construction is underway in the park and some roads are temporarily closed. When processing the main navigation multimodal data, the server will disassemble these temporarily closed road sections as special navigation nodes and extract their closure time, expected opening time, impact on surrounding road traffic and other features.

[0227] For the road sections that were temporarily closed due to construction, the server extracted that the closure time started from this Monday and was expected to be opened next Friday, which had an increase of about 30% in the traffic volume of the two main roads in the surrounding area. These features were also converted into corresponding values ​​to form the feature vector of this special navigation node.

[0228] When comprehensively generating the first navigation feature vector, the server will fully consider the changes caused by the construction, so that the generated first navigation feature vector can accurately reflect the actual situation of the current park navigation, thereby providing users with more accurate and practical navigation decision suggestions, such as recommending alternative routes to avoid construction areas, or adjusting the time schedule for traveling to certain important places.

[0229] Therefore, through the detailed disassembly, feature extraction and integration of the main navigation multimodal data, the first navigation feature vector with important value is generated, which provides a key prerequisite for realizing efficient navigation decisions in smart parks.

[0230] In a possible implementation, step S120 includes:

[0231] Step S121, obtaining a priori random disturbance data, loading the a priori random disturbance data into the disturbance coding subnetwork in the target park navigation decision network, performing disturbance coding on the a priori random disturbance data according to the disturbance coding subnetwork, and generating the mean characteristics and fluctuation characteristics of the a priori random disturbance data.

[0232] Step S122, randomly selecting from the mean features and fluctuation features of the prior random disturbance data to generate implicit mean features and implicit fluctuation features, and generating implicit disturbance features based on the implicit mean features and the implicit fluctuation features.

[0233] In this embodiment, during the navigation decision-making process of the smart park, the server needs to obtain implicit disturbance features to cope with various uncertainties and changes.

[0234] The server first obtains the prior random disturbance data. The sources of these prior random disturbance data are wide and diverse, and may come from historical park operation data, simulated emergency data, or relevant information obtained from the external environment. For example, it may include records of water accumulation on park roads due to bad weather in the past, statistics on crowd gathering and traffic jams caused by large-scale events held temporarily in the park, and the frequency and impact range of sudden equipment failures.

[0235] After obtaining these prior random perturbation data, the server loads them into the perturbation encoding subnetwork in the target park navigation decision network. This perturbation encoding subnetwork is a module specially designed to process and analyze this type of uncertain data.

[0236] According to the function of the perturbation encoding subnetwork, the server performs perturbation encoding on the prior random perturbation data. For example, for a set of data describing the impact of severe weather, the degree to which rainfall of different intensities slows down road speeds may be analyzed, and the mean and fluctuation of this degree of slowdown may be calculated. Assume that in light rainfall, the road speed decreases by an average of 10%, with a fluctuation range of 5% to 15%; in moderate rainfall, the average decrease is 20%, with a fluctuation range of 10% to 30%; in heavy rainfall, the average decrease is 30%, with a fluctuation range of 20% to 40%. Through such analysis, the mean characteristics and fluctuation characteristics of the prior random perturbation data are generated.

[0237] The mean characteristic reflects the average impact under various disturbance conditions, while the fluctuation characteristic shows the variation and uncertainty of this impact.

[0238] The server randomly selects the mean and fluctuation features of the generated prior random perturbation data. This selection is not completely random, but based on a certain probability distribution or specific rules. For example, if the current weather forecast shows a certain possibility of rainfall, the server may be more inclined to select mean and fluctuation features related to rainfall.

[0239] Assume that the server randomly selects the implicit mean feature for moderate rainfall conditions, i.e., an average 20% reduction in road speed, and the corresponding implicit fluctuation feature, i.e., a fluctuation range between 10% and 30%.

[0240] Based on these selected implicit mean features and implicit fluctuation features, the server generates implicit disturbance features, which will serve as an important reference factor for subsequent navigation decision optimization.

[0241] For example, on a weekday morning, the server obtains a priori random perturbation data including the peak flow of people in the park at different time periods. Through the perturbation coding subnetwork, the mean and fluctuation characteristics of the impact of the peak flow of people on road congestion are analyzed.

[0242] Assume that between 10am and 11am, the average congestion impact of the peak flow of people on a main road is a 30% increase in traffic volume, with a fluctuation range of 20% to 40%. The server randomly selects the implicit mean feature (30% increase in traffic volume) and implicit fluctuation feature (20% to 40% fluctuation range) for this time period.

[0243] Then, implicit disturbance features are generated based on these features. Based on this implicit disturbance feature, when optimizing navigation decisions, the server will consider recommending routes that avoid this time period for users who may pass through this road, or plan alternative routes in advance to deal with possible congestion.

[0244] In another scenario, the server obtains a priori random disturbance data about the maintenance and upgrade schedule of equipment in the park. Through the disturbance coding subnetwork, the mean and fluctuation characteristics of the impact of equipment maintenance on surrounding roads and regional traffic are analyzed.

[0245] For example, when equipment maintenance is performed in a certain area, the average impact on the traffic capacity of nearby roads is a 15% decrease, with a fluctuation range of 10% to 20%. The server randomly selects this set of features as implicit mean features and implicit fluctuation features, and generates implicit disturbance features.

[0246] When planning navigation routes for users, the server will consider avoiding areas where equipment maintenance is ongoing, or provide special travel guides during maintenance periods to ensure that users can reach their destinations smoothly while reducing the inconvenience and delays caused by equipment maintenance.

[0247] Therefore, by obtaining a priori random disturbance data and performing a series of operations such as disturbance encoding, random selection and feature generation, we can obtain implicit disturbance features that can reflect various potential uncertainties and changes, providing more comprehensive and flexible support for the navigation decision-making of smart parks to cope with various complex and changing situations.

[0248] In a possible implementation, step S120 further includes:

[0249] In the forward derived transfer subnetwork, a derived transfer feature distribution is obtained, and according to the forward derived transfer subnetwork, the implicit perturbation feature is respectively integrated into the random perturbation data in Y time series nodes to generate a forward perturbation feature. Y is a positive integer, the derived transfer feature distribution is a feature distribution after data fusion, the random perturbation data is randomly selected and generated according to the derived transfer feature distribution, the random perturbation data added in the Y time series nodes are different, and the time series nodes represent the forward time series depth of data fusion.

[0250] In this embodiment, in the navigation decision optimization process of the smart park, the server is responsible for loading the implicit disturbance features into the forward derivative transfer subnetwork in the navigation decision network of the target park and generating the forward disturbance features.

[0251] The server first loads the acquired implicit perturbation features into the forward derivative transfer subnetwork. This forward derivative transfer subnetwork has a complex internal structure and function, which can deeply process and transform the input features.

[0252] In the forward derivative transmission subnetwork, the server obtains the derivative transmission feature distribution. This distribution is formed based on the fusion and analysis of a large amount of historical data and real-time monitoring data, which reflects the characteristic performance of disturbances of different types and degrees after data fusion.

[0253] Assume that in the navigation system of this smart park, the derived transmission feature distribution includes the impact of weather changes, park activities, equipment maintenance and other factors on navigation. For example, in terms of weather changes, it may include the impact of different intensities of rainfall, strong winds, high temperatures, etc. on road conditions and pedestrian flow; in terms of park activities, it may include the impact of the scale, time and location of meetings, exhibitions, celebrations and other activities on surrounding traffic; in terms of equipment maintenance, it may include the restriction of various facilities maintenance areas and time on the route.

[0254] Next, the server integrates the random perturbation data into the implicit perturbation features in Y time series nodes according to the forward derivative transmission sub-network to generate forward perturbation features. Assume that Y=5, which represents five different time stages or processing steps.

[0255] At the first time series node, the server randomly selects and generates random perturbation data related to the current processing stage based on the derived transfer feature distribution. For example, if it is the morning of a weekday, the random perturbation data may be quantitative data about the road congestion caused by the rush hour of employees that usually occurs during this time period. Then, this random perturbation data is integrated into the implicit perturbation feature for processing.

[0256] At the second time series node, new random perturbation data is randomly selected again based on the derived transfer feature distribution. Assume that this time the selected data is about the restriction of traffic on nearby roads due to a small-scale construction in a certain area of ​​the park. It is integrated into the features processed by the first time series node to further enrich and adjust the feature representation.

[0257] At the third time series node, the randomly selected random disturbance data may be related data about a key channel being temporarily shut down due to a sudden equipment failure. Similarly, it is integrated into the features processed by the first two time series nodes to continuously update and optimize the expression of the features.

[0258] At the fourth time series node, the selected random disturbance data may be the forecast data of the impact of a short but heavy rainfall on the drainage and traffic capacity of some roads in the park that afternoon. It is integrated and processed to make the features closer to real-time and future changes.

[0259] At the fifth time series node, the last selected random disturbance data may be the evaluation data about an important enterprise in the park that is about to hold a large-scale product launch conference, which may attract a large number of external visitors and vehicles and put pressure on the surrounding traffic. The final feature processing is integrated and completed to generate the forward disturbance feature.

[0260] It should be noted that the random perturbation data added in these five time series nodes are different, and each time series node represents the forward time series depth of data fusion. In other words, as the time series nodes advance, the server's processing of features is getting closer to the current and future actual conditions, and the factors considered are becoming more comprehensive and specific.

[0261] For example, in a specific scenario, the random disturbance data integrated into the first time series node by the server is the peak data on vehicles entering and exiting the main entrances and exits of the park between 9 and 10 a.m.; the second time series node is the data on the increase in pedestrian flow near a restaurant in the park around 10 a.m.; the third time series node is the information on the diversion of a certain road due to temporary traffic control at 11 a.m.; the fourth time series node is the data on the impact of strong winds on outdoor walking comfort at 2 p.m.; and the fifth time series node is the forecast data on surrounding road congestion that may be caused by a supermarket promotion in the park at 5 p.m.

[0262] By integrating different random disturbance data into multiple time series nodes in sequence, the server can generate a forward disturbance feature that comprehensively considers various time stages and possible changes. This feature can more accurately reflect the uncertainty and dynamic changes faced by the campus navigation system at different times and under different circumstances, and provides an important foundation and basis for the subsequent generation of more accurate and practical navigation optimization decision results.

[0263] In another actual scenario, assuming that it is the weekend, the random disturbance data integrated into the first time series node by the server may be the gathering of tourists at a popular attraction in the park around 10 am; the second time series node may contain relevant data about the peak dining time at multiple restaurants in the park at around 12 noon; the third time series node may contain information about the impact of a small concert held in the park between 2 and 3 pm on the surrounding roads; the fourth time series node may contain data about power outages in some areas caused by sudden power equipment maintenance at around 4 pm; and the fifth time series node may contain data about the concentrated departure of people from a cinema in the park at around 7 pm.

[0264] Therefore, the generated forward disturbance features can fully capture the various complex and changeable situations in the smart park in different time periods, enabling the navigation decision system to respond to various potential challenges and changes more flexibly and effectively, and provide users with more accurate and timely navigation services.

[0265] In a possible implementation, the method further includes:

[0266] Acquire X navigation collection data, perform navigation collection element statistical analysis on the X navigation collection data, output navigation element data appearing in the X navigation collection data as primary navigation multimodal data, and output navigation element data in the navigation collection data other than the primary navigation multimodal data as secondary multimodal data.

[0267] Alternatively, in another possible implementation manner, the method further includes:

[0268] X navigation collection data are acquired, and navigation nodes are decomposed into the X navigation collection data to generate X navigation node data sets. A navigation node data in the navigation node data set refers to a minimum navigation data segment generated after the navigation collection data is decomposed into navigation nodes.

[0269] The X navigation node data sets are encoded to generate X navigation node encoding features, and the X navigation node encoding features are subjected to self-focus association processing to generate X navigation feature focus sets.

[0270] A primary navigation confidence and a secondary navigation confidence corresponding to each navigation node data in the X navigation node data sets are generated according to a self-focus weight coefficient corresponding to each navigation node data in the X navigation feature focus sets; primary navigation multimodal data are determined according to the primary navigation confidence corresponding to each navigation node data in the X navigation node data sets; and secondary multimodal data are determined according to the secondary navigation confidence corresponding to each navigation node data in the X navigation node data sets.

[0271] In this embodiment, the server first obtains X navigation collection data. Assume that X=5, and these 5 navigation collection data may come from different data sources, such as sensors in the park, navigation request records of users, data provided by the park management department, etc.

[0272] The server performs navigation collection element statistics analysis on the five navigation collection data. For example, the five data all contain information about the main roads in the park, such as road name, road length, road capacity, etc., and the navigation element data that appear in all the data are output as the main navigation multimodal data.

[0273] In addition to these common main road information, other element data, such as temporary restrictions on a certain road during a specific time period, special promotions at a certain parking lot, etc., are output as subordinate multimodal data.

[0274] In another possible implementation, after the server obtains the five navigation collection data, it decomposes the navigation nodes. Assuming that a navigation collection data is about the navigation route from the park entrance to the R&D building, the server will decompose it into multiple navigation nodes, such as from the park entrance to the first intersection, from the first intersection to the second intersection, from the second intersection to the entrance of the R&D building, etc. In this way, after the five navigation collection data are decomposed, five navigation node data sets are generated.

[0275] Then, the server encodes the five navigation node data sets. For example, for each navigation node, a specific code is assigned according to its location, connected roads, traffic conditions, etc. Through this encoding operation, five navigation node coding features are generated.

[0276] Next, the server performs self-focused association processing on these five navigation node coding features. This means that the server analyzes the association relationship between the elements within each coding feature. For example, if a navigation node coding feature represents an intersection where traffic congestion often occurs, the server will analyze other elements related to this intersection, such as surrounding buildings, nearby parking lot entrances, etc., to determine the potential association between them. Through this processing, five navigation feature focus sets are generated.

[0277] Afterwards, the server generates the main navigation confidence and the subordinate navigation confidence corresponding to each navigation node data according to the self-focus weight coefficient corresponding to each navigation node data in the five navigation feature focus sets. The self-focus weight coefficient may be calculated based on various factors such as historical data, real-time traffic conditions, and user feedback.

[0278] Assuming that a navigation node is the main entrance of the park, its self-focus weight coefficient is high, and the calculated main navigation confidence is also high, then the navigation node data is determined to be part of the main navigation multimodal data. Another navigation node is a side door that is not often used in the park, and its subordinate navigation confidence is high, so it is determined to be subordinate multimodal data.

[0279] For another example, for a main road leading to the core area of ​​a park that is frequently chosen by users, its primary navigation confidence will be higher and it will be determined as primary navigation multimodal data; while a small road that is only used during a specific time period or under specific circumstances will have a higher subordinate navigation confidence and will be classified as subordinate multimodal data.

[0280] Through the above series of fine processing and analysis, the server can accurately determine the main navigation multimodal data and the subordinate multimodal data, and provide accurate and valuable information for subsequent navigation decision optimization, thereby improving the performance and service quality of the smart park navigation system and better meeting user needs.

[0281] In a specific scenario, the five navigation data collected by the server cover different navigation routes and needs. The first data is the route for employees from the park entrance to the office building on weekday mornings; the second data is the route for visitors from the parking lot to the exhibition hall in the park on weekends; the third data is the route for cargo transport vehicles from the warehouse to the production workshop; the fourth data is the route for employees from the office building to the restaurant at noon; and the fifth data is the route for visitors from the coffee shop in the park to the exit in the afternoon.

[0282] When conducting statistical analysis of navigation collection elements, the server found that several main roads in the park, such as Xintong Avenue and Keji Road, appeared in these five data, so the relevant information of these roads was determined as the main navigation multimodal data. The use of elevators on a specific floor of an office building and the promotion time of a restaurant were determined as subordinate multimodal data.

[0283] For the decomposition of navigation nodes, taking the first data as an example, the server decomposes the route from the park entrance to the office building into the park entrance node, the first intersection node, the second intersection node, the node in front of the office building, etc. After similar decomposition of these five navigation collection data, five navigation node data sets are generated.

[0284] During the coding process, the park entrance node may be coded as "001", indicating that it is more important and has a larger traffic volume; the first intersection node may be coded as "002" and assigned a corresponding coding value based on factors such as the roads it connects and turning restrictions.

[0285] During the self-focus association processing, for the campus entrance node coded as "001", the server will analyze the roads directly connected to it, the surrounding facilities (such as signs, safety facilities, etc.), and the flow of people and vehicles in different time periods to determine the degree of association with these elements.

[0286] When calculating the confidence, if the self-focus weight coefficient of the "001" park entrance node is very high, its main navigation confidence may reach 0.9, and it is clearly identified as a key part of the main navigation multimodal data; while a relatively minor node, such as a staircase node on a certain floor in an office building, its subordinate navigation confidence may be 0.7 and be classified as subordinate multimodal data.

[0287] In this way, accurate and effective data support can be provided for different navigation needs, ensuring that the navigation services within the park are both efficient and in line with actual needs.

[0288] Figure 2 The hardware structure of the smart park service system 100 provided in the embodiment of the present application for implementing the above-mentioned multi-modal smart park navigation data analysis method is shown, as shown in FIG. Figure 2 As shown, the smart campus service system 100 may include a processor 110 , a machine-readable storage medium 120 , a bus 130 , and a communication unit 140 .

[0289] In one possible design, the smart park service system 100 can be a single server or a server group. The server group can be centralized or distributed (for example, the smart park service system 100 can be a distributed system). In some embodiments, the smart park service system 100 can be local or remote. For example, the smart park service system 100 can access information and / or data stored in a machine-readable storage medium 120 via a network. For another example, the smart park service system 100 can be directly connected to the machine-readable storage medium 120 to access the stored information and / or data. In some embodiments, the smart park service system 100 can be implemented on a smart park service system. By way of example only, the smart park service system may include a private cloud, a public cloud, a hybrid cloud, a community cloud, a distributed cloud, an internal cloud, a multi-layer cloud, etc. or any aggregation thereof.

[0290] The machine-readable storage medium 120 may store data and / or instructions. In some embodiments, the machine-readable storage medium 120 may store data obtained from an external terminal. In some embodiments, the machine-readable storage medium 120 may store data and / or instructions used by the smart park service system 100 to execute or use to complete the exemplary methods described in this application.

[0291] During the specific implementation process, one or more processors 110 execute computer executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the multimodal smart park navigation data analysis method of the above method embodiment. The processor 110, the machine-readable storage medium 120 and the communication unit 140 are connected through the bus 130, and the processor 110 can be used to control the sending and receiving actions of the communication unit 140.

[0292] The specific implementation process of the processor 110 can be found in the various method embodiments executed by the above-mentioned smart park service system 100. The implementation principles and technical effects are similar and will not be repeated here in this embodiment.

[0293] In addition, an embodiment of the present application also provides a readable storage medium, in which computer executable instructions are set. When the processor executes the computer executable instructions, the above-mentioned multi-modal smart park navigation data analysis method is implemented.

[0294] It should be noted that in order to simplify the description disclosed in this application and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, multiple features are sometimes combined into one embodiment, drawings, or descriptions thereof. Similarly, it should be noted that in order to simplify the description disclosed in this application and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, multiple features are sometimes combined into one embodiment, drawings, or descriptions thereof.

Claims

1. A multi-modal smart park navigation data analysis method, characterized in that: The method comprises: The main navigation multimodal data is characterized to generate a first navigation feature vector, and X second navigation feature vectors are generated according to the first navigation feature vector and X slave multimodal data; X is a positive integer, the main navigation multimodal data is multimodal acquisition data of main navigation restriction requirements in the smart park, the slave multimodal data is multimodal acquisition data of slave navigation restriction requirements in the smart park, and the slave navigation restriction requirements represented by the X slave multimodal data are all different; a second navigation feature vector is generated by the first navigation feature vector and one slave multimodal data; Obtaining implicit disturbance features, loading the implicit disturbance features into a forward derivative transfer subnetwork in a target park navigation decision network, and using the forward derivative transfer subnetwork to output forward disturbance features; The forward perturbation feature is loaded into the reverse derivative transfer subnetwork in the target park navigation decision network, and X derivative transfer trajectories in parallel relationship are generated by using the reverse derivative transfer subnetwork; the reverse derivative transfer subnetwork is used to fuse the first navigation feature vector in a first derivative fusion channel, and is used to fuse a second navigation feature vector for each derivative transfer trajectory in a second derivative fusion channel; the loading of the second derivative fusion channel includes the generation of the first derivative fusion channel; Based on the forward disturbance feature, the first navigation feature vector fused in the first derivative fusion channel, and the X second navigation feature vectors fused in the second derivative fusion channel, a navigation optimization decision result corresponding to each of the X derivative transfer trajectories is generated; the main navigation restriction requirements in the X navigation optimization decision results are the same, and the subordinate navigation restriction requirements in the X navigation optimization decision results are different.

2. The multimodal smart park navigation data analysis method according to claim 1 is characterized in that: The generating, based on the forward disturbance feature, the first navigation feature vector fused in the first derivative fusion channel, and the X second navigation feature vectors fused in the second derivative fusion channel, a navigation optimization decision result corresponding to each of the X derivative transfer trajectories comprises: The first navigation feature vector is fused in a first derivative fusion channel by using the reverse derivative transfer subnetwork, the forward perturbation feature and the first navigation feature vector fused in the first derivative fusion channel are compressed to generate a first decision factor feature obtained by the first derivative fusion channel, and the first decision factor feature is loaded into the X derivative transfer trajectories respectively; In the second derivative fusion channel, a second navigation feature vector is respectively fused for each derivative transfer trajectory, and in each derivative transfer trajectory, the first decision factor feature and the second navigation feature vector fused in the second derivative fusion channel are respectively subjected to dimensionality upgrading processing to generate X second decision factor features obtained by the second derivative fusion channel, and navigation optimization decision results corresponding to each of the X derivative transfer trajectories are generated according to the X second decision factor features; The compressing process of the forward disturbance feature and the first navigation feature vector fused in the first derivative fusion channel to generate the first decision factor feature obtained by the first derivative fusion channel includes: The forward perturbation feature is loaded into the reverse derivative transfer subnetwork; the reverse derivative transfer subnetwork includes a perturbation estimation unit; the perturbation estimation unit includes a dimension reduction subunit, a derivative transfer subunit and a dimension increase subunit, the loading of the derivative transfer subunit is the generation of the dimension reduction subunit, and the loading of the dimension increase subunit is the generation of the derivative transfer subunit; the dimension reduction subunit includes a first self-focusing association branch and a first interactive focusing association branch; In the first self-focusing association branch, the forward disturbance feature is subjected to self-focusing association processing to generate a first self-focusing association vector, and in the first interactive focusing association branch, the first self-focusing association vector and the fused first navigation feature vector are subjected to interactive focusing association processing to generate a first decision factor feature obtained in the first derivative fusion channel; the number of decision fields of the first decision factor feature is less than the number of decision fields of the forward disturbance feature; The method further comprises: Copying the first decision factor features according to the derivation and transfer subunit, and distributing the copied X first decision factor features to the dimension-increasing subunit for outputting X second decision factor features in the second derivation and fusion channel; Among them, in the first interactive focus association branch, interactive focus association processing is performed on the first self-focus association vector and the fused first navigation feature vector to generate a first decision factor feature obtained in the first derivative fusion channel, including: Calculating the first self-focus correlation vector and a retrieval index array to generate a retrieval vector, calculating the first self-focus correlation vector and a primary key index array to generate a primary key vector, and calculating the first navigation feature vector and a numerical index array to generate a numerical vector; Generate self-focusing weight information according to the search vector and the primary key vector, compress the self-focusing weight information according to the number of fields of the primary key vector, perform regularization conversion on the compressed self-focusing weight information to generate regularization conversion features, and calculate the regularization conversion features and the numerical vector to generate a first decision factor feature; The X derivative transfer trajectories include a target derivative transfer trajectory; and the first decision factor feature and the second navigation feature vector fused in the second derivative fusion channel are respectively subjected to dimensionality increase processing in each derivative transfer trajectory to generate X second decision factor features obtained by the second derivative fusion channel, including: Loading the first decision factor feature into the dimension-raising subunit; the dimension-raising subunit includes a second self-focusing association branch and a second interactive focusing association branch corresponding to the target-derived transfer trajectory; In the second self-focus correlation branch, performing self-focus correlation processing on the first decision factor feature to generate a second self-focus correlation vector; In the second interactive focusing association branch, the second navigation feature vector fused with the second self-focusing association vector and the target-derived transfer trajectory is interactively focused and associated to generate an interactive focusing association vector, and a second decision factor feature obtained by the target-derived transfer trajectory is generated based on the interactive focusing association vector; the number of decision fields of the second decision factor feature is equal to the number of decision fields of the forward disturbance feature.

3. The multimodal smart park navigation data analysis method according to claim 1 is characterized in that: The reverse derivative transfer subnetwork includes Y disturbance estimation units, where Y is a positive integer; each disturbance estimation unit is used to fuse the first navigation feature vector in a first derivative fusion channel, and to fuse a second navigation feature vector for each derivative transfer trajectory in a second derivative fusion channel; The Y disturbance estimation units include a disturbance estimation unit, a is a positive integer not greater than Y; The generating, based on the forward disturbance feature, the first navigation feature vector fused in the first derivative fusion channel, and the X second navigation feature vectors fused in the second derivative fusion channel, a navigation optimization decision result corresponding to each of the X derivative transfer trajectories comprises: The feature to be processed is loaded into the disturbance estimation unit, and the disturbance estimation unit is used to compress the feature to be processed and the first navigation feature vector fused in the first derivative fusion channel to generate X first decision factor features obtained by the disturbance estimation unit, and the X first decision factor features obtained by the disturbance estimation unit are respectively loaded into the X derivative transfer trajectories obtained by the disturbance estimation unit; one first decision factor feature is loaded into one derivative transfer trajectory; if a is 1, the feature to be processed is the forward disturbance feature, and the X first decision factor features are all the same feature; if a is greater than 1, the feature to be processed includes X second decision factor features obtained by the disturbance estimation unit, and one first decision factor feature is determined by a second decision factor feature obtained by the disturbance estimation unit and the first navigation feature vector, and the disturbance estimation unit is the previous disturbance estimation unit of the disturbance estimation unit among the Y disturbance estimation units, and the generation of the disturbance estimation unit is the loading of the disturbance estimation unit; In the second derivative fusion channel, a second navigation feature vector is respectively fused for each derivative transfer trajectory of the disturbance estimation unit, and in each derivative transfer trajectory, a first decision factor feature obtained by the disturbance estimation unit and the fused second navigation feature vector are respectively subjected to dimensionality increase processing to generate X second decision factor features obtained by the disturbance estimation unit; If a is Y, then navigation optimization decision results corresponding to the X derived transfer trajectories of the disturbance estimation unit are generated according to the X second decision factor features obtained by the disturbance estimation unit.

4. The multimodal smart park navigation data analysis method according to claim 2 is characterized in that: The generating, according to the X second decision factor features, navigation optimization decision results corresponding to the X derived transfer trajectories respectively includes: The X second decision factor features are loaded into a mapping subnetwork of the target park navigation decision network; in the mapping subnetwork, the implicit perturbation features are respectively mapped and output according to the X second decision factor features to generate X mapping knowledge point data; and X navigation optimization decision results are generated according to the X mapping knowledge point data.

5. The multimodal smart park navigation data analysis method according to claim 1 is characterized in that: The generating a first navigation feature vector by performing feature representation on the primary navigation multimodal data includes: Loading the main navigation multimodal data into a target park navigation decision network; the target park navigation decision network includes a feature extraction subnetwork; According to the feature extraction subnetwork, the main navigation multimodal data is disassembled to generate a navigation node data set corresponding to the main navigation multimodal data; a navigation node data in the navigation node data set refers to a minimum navigation data fragment generated after the smart park navigation data is disassembled; Feature extraction is performed on each navigation node data in the navigation node data set to generate a first navigation feature vector.

6. The multimodal smart park navigation data analysis method according to claim 1 is characterized in that: The obtaining of implicit disturbance features includes: Acquire a priori random disturbance data, load the a priori random disturbance data into a disturbance coding subnetwork in a target park navigation decision network, perform disturbance coding on the a priori random disturbance data according to the disturbance coding subnetwork, and generate mean characteristics and fluctuation characteristics of the a priori random disturbance data; Random selection is performed among the mean characteristics and fluctuation characteristics of the prior random disturbance data to generate implicit mean characteristics and implicit fluctuation characteristics, and implicit disturbance characteristics are generated based on the implicit mean characteristics and the implicit fluctuation characteristics.

7. The multimodal smart park navigation data analysis method according to claim 1 is characterized in that: The step of loading the implicit disturbance feature into the forward derivative transfer subnetwork in the target park navigation decision network and using the forward derivative transfer subnetwork to output the forward disturbance feature comprises: In the forward derived transfer subnetwork, a derived transfer feature distribution is obtained, and according to the forward derived transfer subnetwork, the implicit perturbation feature is respectively integrated into the random perturbation data in Y time series nodes to generate a forward perturbation feature; Y is a positive integer, the derived transfer feature distribution is the feature distribution after data fusion, the random perturbation data is randomly selected and generated according to the derived transfer feature distribution, the random perturbation data added in the Y time series nodes are different, and the time series nodes represent the forward time series depth of data fusion.

8. The multimodal smart park navigation data analysis method according to claim 1, characterized in that: The method further comprises: Acquire X navigation collection data, perform navigation collection element statistical analysis on the X navigation collection data, output navigation element data appearing in the X navigation collection data as primary navigation multimodal data, and output navigation element data in the navigation collection data other than the primary navigation multimodal data as secondary multimodal data.

9. The multimodal smart park navigation data analysis method according to claim 1, characterized in that: The method further comprises: Acquire X navigation collection data, decompose the X navigation collection data into navigation nodes, and generate X navigation node data sets; one navigation node data in the navigation node data set refers to a minimum navigation data segment generated after decomposing the navigation collection data into navigation nodes; Encoding the X navigation node data sets to generate X navigation node encoding features, and performing self-focus association processing on the X navigation node encoding features to generate X navigation feature focus sets; A primary navigation confidence and a secondary navigation confidence corresponding to each navigation node data in the X navigation node data sets are generated according to a self-focus weight coefficient corresponding to each navigation node data in the X navigation feature focus sets; primary navigation multimodal data are determined according to the primary navigation confidence corresponding to each navigation node data in the X navigation node data sets; and secondary multimodal data are determined according to the secondary navigation confidence corresponding to each navigation node data in the X navigation node data sets.

10. A smart park service system, characterized in that: The smart campus service system includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the multimodal smart campus navigation data analysis method described in any one of claims 1 to 9 above.