Community open space psychological recovery effect evaluation method and system based on multi-modal perception
By integrating physiological, behavioral, and environmental data through multimodal perception technology, the psychological recovery effect of community open spaces is dynamically evaluated, solving the problems of modal fragmentation and insufficient rule generation in existing assessment methods, and achieving high-precision assessment of psychological recovery effect and guidance for transformation.
Patent Information
- Application Number
- CN202510892962.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-30
- Publication Date
- 2025-10-28
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing methods for evaluating the psychological recovery effects of open spaces in communities rely on a single image modality, which cannot effectively integrate physiological signals, behavioral trajectories, and eye-tracking data. This leads to discrepancies between the evaluation results and the actual psychological state, and neural network models struggle to generate clear design thresholds and transformation rules.
Multimodal sensing technology is used to simultaneously acquire physiological signals, movement trajectories, and panoramic image data. By dynamically determining the visible area, calculating information entropy and prediction error, and combining interpretable machine learning models to output recovery efficiency index and multidimensional mitigation indicators, actionable environmental modification rules are generated.
It enables precise quantitative assessment of the effects of psychological recovery, outputs clear guidelines for environmental modification, overcomes the modal fragmentation and black-box prediction defects of traditional assessment methods, and improves the accuracy and credibility of the assessment.
Smart Images

Figure CN120853875A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of environmental psychology assessment and spatial optimization decision-making technology for smart cities, and in particular to an assessment method and system for the psychological recovery effect of community open spaces based on multimodal perception. Background Technology
[0002] Against the backdrop of rapid urbanization, the importance of community open spaces for maintaining residents' mental health continues to rise. Existing assessment methods are insufficient to meet key needs, including capturing the dynamic interaction between users' physiological behaviors and the environmental space, and simultaneously establishing a correlation model between visual environmental elements and psychological recovery indicators. Therefore, it is urgent to construct a scientific and efficient evaluation system for psychological recovery effects.
[0003] Currently, mainstream technologies employ static assessment schemes based on image analysis. These schemes use computer vision algorithms to process street scene images and extract spatial distribution features of physical environmental elements, such as quantitative indicators like green coverage ratio and water visibility. A multi-layer neural network regression model is then established to map environmental feature vectors to standardized psychological recovery scores.
[0004] However, this approach has significant limitations. First, relying on a single image modality cannot integrate dynamic data such as physiological signals and behavioral trajectories, leading to discrepancies between the assessment results and the actual psychological state. Second, the neural network model can only provide a ranking of factor contributions and cannot generate modification rules with clearly defined design thresholds, such as the coordinated control requirements for greening rate and pedestrian walkway width. Summary of the Invention
[0005] This application provides a method and system for evaluating the psychological recovery effect of community open spaces based on multimodal perception, in order to solve the problem that the evaluation results in the prior art deviate from the actual psychological state.
[0006] Firstly, this application provides a method for evaluating the psychological recovery effect of community open spaces based on multimodal perception, including: Within the target environment, acquire the user's physiological signals, movement trajectory data, eye movement data, and panoramic image data of the target environment. Based on the user's position and line of sight, the visible area is dynamically determined, the panoramic image data is analyzed, the spatial distribution parameters of environmental elements within the visible area are extracted, and the information entropy value representing the complexity of the environment is calculated based on the spatial distribution parameters. Based on the magnitude of the deviation of the physiological signal from the user's resting state, the intensity of physiological stress deviation is calculated as the first type of prediction error. The degree of matching between the movement trajectory data and the environmental spatial structure is analyzed to quantify the path execution deviation as the second type of prediction error. Combining the eye movement characteristics in the eye movement data with the distribution map of environmental visual attraction, the degree of attention allocation deviation is calculated as the third type of prediction error. The first type of prediction error, the second type of prediction error, the third type of prediction error, and the information entropy value are input into the pre-built evaluation model to output the recovery efficiency index and the multidimensional mitigation index. The multidimensional mitigation index is used to characterize the degree of mitigation of each prediction error. By associating the spatial distribution parameters of the environmental elements with the multidimensional mitigation indicators, and using interpretable machine learning to extract the critical values of key environmental interventions and the information entropy recovery effect relationship function, and combining the dynamic recovery effect indicators, multidimensional mitigation indicators and the extracted rules, an evaluation result is generated.
[0007] Optionally, it also includes: Multiple sets of subjective evaluation data were collected using a semantic difference scale, and the subjective evaluation data were transformed into natural language descriptions using a large language model. The natural language description is input into the text encoder to generate text embedding vectors, and a dual-encoder variational autoencoder framework is constructed. A semantically shared latent space is built through the text encoder and decoder, and the KL divergence between the visual index distribution and the text embedding distribution is minimized when the text encoder or decoder is frozen. In the semantic shared latent space, a mapping is established between the text embedding vector and the psychological recovery effect evaluation result, and the text input path is enabled to directly encode the natural language description into the latent space; The recovery score is predicted by a regression model, and the semantic difference scale input path is enabled. The structured semantic difference scale score is converted into a natural language description, and the recovery score is predicted by the text path.
[0008] Optionally, it also includes: The recovery efficiency index is discretized into five recovery level labels, which include extremely low recovery level, low recovery level, medium recovery level, high recovery level and extremely high recovery level; Using the spatial distribution parameters of the environmental elements as input features and the five-level recovery level labels as prediction targets, a classification model based on a rule representation learner is trained. A weighted set of decision rules is extracted from the classification model, the set of decision rules containing the mapping relationship between environmental element threshold conditions and recovery levels; Based on the decision rule set, the critical values of key environmental interventions for multidimensional mitigation indicators are analyzed, and a nonlinear relationship function between information entropy and recovery efficiency index is fitted. The critical values of key environmental interventions and the nonlinear relationship function are then used as interpretable rule outputs.
[0009] Optionally, the first type of prediction error, the second type of prediction error, the third type of prediction error, and the information entropy value are input into a pre-built evaluation model to output a recovery efficiency index and a multidimensional mitigation index. The multidimensional mitigation index is used to characterize the degree of mitigation of each prediction error, including: The first type of prediction error, the second type of prediction error, the third type of prediction error, and the information entropy value are synchronously input into the inference interface of the pre-trained multi-channel fusion evaluation model. In the inference interface, the physiological stress bias intensity, path execution bias degree, and attention allocation bias degree are generated into error feature vectors through independent fully connected layers, and the error feature vectors are cross-attention weighted with the information entropy feature vectors. Based on the attention weight allocation results, the output scalar value of all feature channels is used as the recovery efficiency index, and three correction coefficients are generated for the three types of prediction errors as multidimensional mitigation indicators.
[0010] Optionally, the physiological stress bias intensity, path execution bias, and attention allocation bias are each generated into an error feature vector through an independent fully connected layer. The error feature vector is then cross-weighted with the information entropy feature vector, including: The physiological stress deviation intensity, path execution deviation, and attention allocation deviation are converted into error feature vectors of a unified dimension by independent feature converters, and all error feature vectors are integrated into an error feature set. The information entropy value is converted into an environmental entropy feature vector of uniform dimension by an environmental feature converter. The sum of the corresponding element-wise products of each error feature vector in the error feature set and the environmental entropy feature vector is calculated to generate an interaction value set. Divide each interaction value in the set of interaction values by the sum of the interaction value sets to generate a weighted set, and multiply each weight in the weighted set by the corresponding error feature to generate a weighted feature set; The weighted features in the weighted feature set are element-wise superimposed with the environmental entropy feature vector to generate the cross-attention weight allocation result representing the fused features.
[0011] Optionally, by associating the spatial distribution parameters of the environmental elements with the multidimensional mitigation indicators, and extracting key environmental intervention thresholds and information entropy recovery effect relationship functions through interpretable machine learning, and combining these with the dynamic recovery effect indicators, multidimensional mitigation indicators, and extraction rules, an evaluation result is generated, including: The spatial distribution parameters of the environmental elements are concatenated with the multidimensional mitigation index to form a set of associated features. The set of associated features is then processed by an interpretable machine learning model to extract a set of rule antecedents that meet the minimum support. The environmental element combination conditions that significantly improve the multidimensional mitigation index are extracted from the set of rule antecedents as key environmental intervention thresholds, and the functional relationship between information entropy value and recovery efficiency index is used as the information entropy recovery effect relationship function. Based on the critical values of key environmental interventions and the relationship function of information entropy recovery effect, combined with the data of the dynamic recovery effect index, the state data of the multidimensional mitigation index, and the parsed rules, an evaluation result including spatial optimization priority is generated.
[0012] Optionally, the spatial distribution parameters of the environmental elements are concatenated with the multidimensional mitigation index to form a correlated feature set. This correlated feature set is then processed by an interpretable machine learning model to extract a set of rule antecedents that satisfy the minimum support, including: The pixel area ratio of each environmental element is used as the spatial element configuration feature, and the three correction coefficients in the multidimensional mitigation index are used as the mitigation effect feature. The spatial element configuration feature and the mitigation effect feature are connected sequentially along the feature dimension to form an associated feature set. All combinations of environmental element types are traversed in the associated feature set to generate candidate condition items. The coverage ratio of each candidate condition item in the associated feature set sample is calculated, and only candidate condition items with a coverage ratio greater than a set threshold are retained. The retained candidate conditions are verified by rule confidence. The improvement of the multidimensional mitigation index when the candidate condition is true is calculated. When the improvement exceeds the baseline fluctuation range, it is marked as a valid condition. All the valid condition conditions are merged to form a set of rule antecedents that meet the minimum support.
[0013] Optionally, based on the user's position and line of sight, the visible area is dynamically determined, the panoramic image data is parsed, spatial distribution parameters of environmental elements within the visible area are extracted, and an information entropy value characterizing environmental complexity is calculated based on the spatial distribution parameters, including: Based on user position and line of sight data, the horizontal azimuth and vertical pitch boundaries of the line of sight are determined, and combined with depth sensor data, the three-dimensional visible area range is dynamically constructed. In the panoramic image data, image blocks corresponding to the visible area are cut out, the outline boundaries of all environmental elements in the image blocks are extracted, and the spatial distribution parameters occupied by each type of element are statistically analyzed according to the element attribute type. The spatial distribution parameters are used as a probability distribution, and the information entropy value, which characterizes the degree of disorder in environmental complexity, is obtained by weighted summation of probability logarithms.
[0014] Optionally, based on the magnitude of the deviation of the physiological signal from the user's resting state, the intensity of physiological stress deviation is calculated as a first type of prediction error; the degree of matching between the movement trajectory data and the environmental spatial structure is analyzed to quantify the path execution deviation as a second type of prediction error; and the attention allocation deviation is calculated by combining the eye movement characteristics in the eye movement data with the environmental visual attraction distribution map as a third type of prediction error, including: The absolute difference between the amplitudes of the physiological signal and the skin conductance signal is calculated as the amplitude deviation between the physiological signal and the user's resting state reference value. The amplitude deviation is then smoothed through a sliding window to obtain the physiological stress deviation intensity as the first type of prediction error. Spatial matching is performed between the actual location sequence in the movement trajectory data and the node sequence of the predetermined path in the environmental spatial structure. The proportion of the number of offset nodes is used as the spatial matching difference. The path execution deviation is quantified by the spatial matching difference as the second type of prediction error. By combining the eye movement features in the eye movement data with the distribution relationship of high attraction regions in the environmental visual attraction distribution map, the positional deviation ratio between the fixation point and the high attraction region is calculated, and the attention allocation deviation is calculated as the third type of prediction error based on the positional deviation ratio.
[0015] Secondly, this application provides an evaluation system for the psychological recovery effect of community open spaces based on multimodal perception, including: The acquisition module acquires the user's physiological signals, movement trajectory data, eye movement data, and panoramic image data of the target environment space. The parsing module dynamically determines the visible area range based on the user's position and line of sight, parses the panoramic image data, extracts the spatial distribution parameters of environmental elements within the visible area, and calculates the information entropy value representing the complexity of the environment based on the spatial distribution parameters. The calculation module calculates the physiological stress deviation intensity based on the magnitude of the deviation of the physiological signal from the user's resting state, as the first type of prediction error; analyzes the degree of matching between the movement trajectory data and the environmental spatial structure, quantifies the path execution deviation, as the second type of prediction error; and combines the eye movement characteristics in the eye movement data with the environmental visual attraction distribution map to calculate the attention allocation deviation, as the third type of prediction error. The construction module inputs the first type of prediction error, the second type of prediction error, the third type of prediction error and the information entropy value into the pre-constructed evaluation model to output the recovery efficiency index and the multidimensional mitigation index. The multidimensional mitigation index is used to characterize the degree of mitigation of each prediction error. The generation module associates the spatial distribution parameters of the environmental elements with the multidimensional mitigation indicators, extracts key environmental intervention thresholds and information entropy recovery effect relationship functions through interpretable machine learning, and generates evaluation results by combining the dynamic recovery effect indicators, multidimensional mitigation indicators and the extracted rules.
[0016] This application constructs a multimodal perception foundation by simultaneously acquiring physiological signals, movement trajectories, eye-tracking data, and panoramic images, enabling holographic capture of the environmental space and user behavior. Based on dynamic gaze delineation of the visible area and calculation of environmental information entropy, it overcomes the limitations of static assessment and accurately quantifies the impact of spatial complexity on psychological recovery. By defining three types of prediction errors—physiological stress bias, path execution bias, and attention allocation bias—it establishes a dynamic assessment dimension of physical and mental response and environmental interaction. It integrates multi-source errors and entropy input assessment models, simultaneously outputting indices characterizing overall recovery effectiveness and sub-item error mitigation indicators, achieving integrated effect assessment and problem localization. Finally, by associating environmental parameters and mitigation indicators, it extracts quantifiable design critical values and entropy recovery functions, generating a priority decision report to guide spatial optimization, completely resolving the shortcomings of traditional solutions such as modal fragmentation, black-box prediction, and static adaptation.
[0017] Furthermore, subjective evaluation data is collected through a semantic difference scale. This data is then transformed into natural language descriptions using a large language model and input into a text encoder to generate embedding vectors. A dual-encoder variational autoencoder framework is constructed to create a semantically shared latent space, minimizing the distributional differences between visual indicators and text embeddings when freezing text components. Simultaneously, a mapping from text embeddings to psychological recovery effects is established. Through a dual-path prediction mechanism—the natural language direct input path and the scale transformation path—the recovery score is output, achieving efficient transformation of scale data and cross-modal semantic alignment. This overcomes the limitations of the separation between subjective perception and objective environment in traditional assessments, constructs a unified semantic representation framework, and provides stable predictive capabilities for cross-modal collaboration, significantly enhancing the assessment adaptability in data-deficient scenarios.
[0018] These or other aspects of this application will become more apparent in the following description of the embodiments. Attached Figure Description
[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0020] Figure 1 A flowchart is shown below illustrating an evaluation method for the psychological recovery effect of community open spaces based on multimodal perception, as provided in this application. Figure 2 The illustration shows a scenario diagram of an evaluation method for the psychological recovery effect of community open spaces based on multimodal perception, as provided in this application. Figure 3 The diagram shows a structural schematic of an evaluation system for the psychological recovery effect of community open spaces based on multimodal perception, as provided in this application. Detailed Implementation
[0021] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.
[0022] In some of the processes described in the specification, claims, and accompanying drawings of this application, multiple operations appearing in a specific order are included. However, it should be clearly understood that these operations may not be executed in the order they appear herein, or may be executed in parallel. The operation numbers, such as 101, 102, etc., are merely used to distinguish different operations and do not themselves represent any execution order. Furthermore, these processes may include more or fewer operations, and these operations may be executed sequentially or in parallel. It should be noted that the descriptions such as "first," "second," etc., in this document are used to distinguish different messages, devices, modules, etc., and do not represent a chronological order, nor do they limit "first" and "second" to different types.
[0023] Researchers have found that existing methods for assessing the psychological recovery effects of open spaces in communities have significant limitations. Relying on a single image modality fails to effectively integrate multidimensional dynamic data such as users' physiological signals, behavioral trajectories, and eye-tracking focus, leading to discrepancies between assessment results and actual psychological states. Furthermore, the neural network models used in existing methods can only rank the contributions of influencing factors, making it difficult to generate actionable environmental modification rules with clearly defined design thresholds, such as specific collaborative control requirements for green space ratios and walkway widths. Therefore, there is an urgent need for a method for assessing psychological recovery effects that can integrate multidimensional dynamic sensory data and output quantifiable operational rules.
[0024] To address the aforementioned issues, this invention proposes a method for evaluating the psychological recovery effect in community open spaces based on multimodal perception. Its core lies in dynamically integrating physiological, behavioral, eye-tracking, and environmental panoramic data to accurately quantify user psychological prediction errors and environmental complexity. Combined with an interpretable machine learning model, it outputs recovery efficacy indicators and extracts key environmental intervention rules. This method effectively solves the evaluation bias problem caused by a single data source and overcomes the shortcomings of existing models in rule generation, directly outputting specific transformation guidelines with clear critical values and collaborative control standards.
[0025] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0026] Figure 1 This application provides a flowchart of an evaluation method for the psychological recovery effect of community open spaces based on multimodal perception, as shown in the embodiments of this application. Figure 1 As shown, the method includes: 101. In the target environment, acquire the user's physiological signals, movement trajectory data, eye movement data, and panoramic image data of the target environment; In step 101 above, the target environment space refers to the open community space where the psychological recovery effect assessment is required. Physiological signals refer to signals related to the user's physiological state, such as heart rate, respiratory rate, or skin conductivity. Movement trajectory data refers to the location and speed data of the user's movement path within the target environment space. Eye movement data refers to the user's eye tracking information, including fixation point location and eye movement patterns. Panoramic image data refers to omnidirectional visual information of the target environment space, scene images captured from multiple angles.
[0027] In this embodiment, the user's physiological signals are first acquired using wearable biosensor technology installed on the user in the target environment, such as a heart rate monitoring bracelet or electroencephalography (EEG) device. This captures changes in the user's physiological state within an open community space, generating continuous heart rate or skin conductance signal data. Next, a portable mobile terminal or inertial measurement unit (IMU) with a built-in GPS module is deployed in the target environment. Continuous positioning and tracking technology records the user's walking path and speed, aggregating these location coordinates and timestamp data into a motion trajectory dataset for subsequent analysis of the user's movement patterns within the space. Then, wearable eye-tracking glasses are attached to the user's eyes, capturing eye movement trajectories, such as fixation point position and saccade data, using a high-speed camera and infrared light source. The user's visual attention distribution is recorded in the target environment, generating raw eye movement signals. Noise is then removed using a digital filtering algorithm to form structured eye movement data. Finally, fixed or mobile 360-degree panoramic camera equipment is deployed in key areas of the community's open space. Image acquisition technology is used to capture comprehensive visual scene information, and the images taken from multiple angles are stitched together and image processing algorithms are used to correct distortion, generating a complete panoramic image dataset of the target environment.
[0028] In a practical application, during an evaluation study exploring the psychological recovery effects of open spaces in a community, the evaluation team selected a typical waterfront vegetation area within the community as the target environmental space for multimodal data collection. A volunteer wearing an integrated sensor kit entered the area and sat quietly for 15 minutes. During this time, a device worn on their wrist continuously recorded physiological signals, using high-frequency sampling to obtain dynamic changes in skin conductance and heart rate. Movement was tracked via a foot positioning module, and the data clearly showed that the volunteer mainly stayed near a bench, with a relatively small overall activity range, exhibiting a predominantly static resting characteristic. An accompanying eye-tracking device fully recorded their eye movements, and analysis revealed that their gaze was primarily focused on the dynamic surface of the water and the surrounding tree landscape. Simultaneously, to comprehensively characterize the environmental features, panoramic cameras deployed in multiple directions within the space simultaneously captured a temporal sequence of panoramic images including water, vegetation, paving, and other active elements within the space, constructing a detailed three-dimensional spatiotemporal model of the environmental scene. By integrating and analyzing heterogeneous data such as physiological activity patterns, spatial movement patterns, visual attention preferences, and panoramic environmental elements, researchers effectively assessed the comprehensive restorative effect of the community's open space on the volunteer's mental state and observed positive signs of improvement in their physical and mental condition.
[0029] In the overall scheme of step 101 above, objective quantitative assessment of the psychological recovery process is achieved by simultaneously collecting users' physiological signals, movement trajectory data, eye movement data, and panoramic environmental image data in the community open space. Through coupled analysis of user physiological state monitoring, behavioral spatiotemporal feature analysis, visual attention distribution modeling, and environmental feature extraction, this method overcomes the subjectivity limitations of traditional questionnaire measurements and significantly improves the accuracy and reliability of psychological recovery effect assessment. Finally, a dynamic assessment model integrating environmental perception and individual response is constructed, providing scientific data support for the optimization of community open space design and the construction of a healthy living environment, forming an automated and objective psychological recovery effect assessment system.
[0030] 102. Based on the user's position and line of sight, dynamically determine the range of the visible area, analyze the panoramic image data, extract the spatial distribution parameters of environmental elements within the visible area, and calculate the information entropy value representing the complexity of the environment based on the spatial distribution parameters. Optionally, step 102 may specifically include the following steps: 1021. Based on the user's position and line of sight data, determine the horizontal azimuth and vertical pitch boundaries of the line of sight, and dynamically construct the three-dimensional visible area range by combining depth sensor data. 1022. In the panoramic image data, cut out the image blocks corresponding to the visible area range, extract the outline boundaries of all environmental elements in the image blocks, and statistically analyze the spatial distribution parameters occupied by each type of element according to the element attribute type. 1023. Using the spatial distribution parameters as a probability distribution, the information entropy value, which characterizes the degree of disorder in environmental complexity, is obtained by weighted summation of probability logarithms.
[0031] In the above steps, user position refers to the user's current position coordinates in the target environment space. Line of sight direction refers to the azimuth vector data of the user's current gaze. Visible area refers to the three-dimensional spatial region visible from the user's perspective along the line of sight. Panoramic image data refers to the image dataset capturing omnidirectional visual information of the target environment space. Environmental elements refer to physical objects within the visible area, such as vegetation, buildings, or roads. Spatial distribution parameters refer to the area or volumetric proportion of each type of element in the environment within the visible area. Information entropy value is a numerical indicator quantifying the degree of disorder in the visual complexity of the environment. Horizontal azimuth boundary refers to the lateral span data defining the visible range from the left and right extreme angles in the horizontal direction. Vertical pitch boundary refers to the longitudinal span data defining the visible range from the up and down extreme angles in the vertical direction. Depth sensor data refers to depth information data acquired by sensors that measure distance, such as LiDAR. Three-dimensional visible area refers to the stereoscopic visible spatial region constructed based on azimuth, pitch, and depth data. Image patch refers to the corresponding two-dimensional image segment extracted from the panoramic image according to the visible area. Contour boundary refers to the boundary line data after detecting the edge contour of environmental elements. Element attribute type refers to the classification of environmental elements, such as natural or artificial categories. Probability distribution refers to the probability representation of spatial distribution parameters after normalization according to element type. Weighted sum of probability logarithms refers to the operation of calculating the sum of the logarithms of probability values using the entropy formula. The degree of disorder in environmental complexity refers to the level of randomness or diversity in the distribution of environmental elements.
[0032] In this embodiment of the application, firstly, the horizontal azimuth and vertical pitch boundaries of the line of sight are determined based on the user's position and line of sight direction data in step 1021. This process uses position tracking technology to process the user's position data and line of sight direction data. The user's position coordinate data and line of sight direction angle data are captured by a gyroscope sensor and a positioning system. Then, the horizontal azimuth boundary, vertical pitch boundary, and depth value of the line of sight are integrated using a depth information fusion algorithm in combination with depth sensor data. Finally, a three-dimensional visible area range is dynamically constructed and output as the basic data.
[0033] Secondly, in step 1022, image blocks corresponding to the visible area range are cut from the panoramic image data. This process uses image cropping algorithms in computer vision technology to process the dynamically constructed three-dimensional visible area range data. The image cropping algorithm locates the projection area of the three-dimensional visible area range on the two-dimensional panoramic image data. Then, the edge detection algorithm is used to process the image block data to extract the contour boundary data of all environmental elements. Then, according to the element attribute type, classification and recognition algorithms such as support vector machines or deep learning models are used to statistically analyze the pixel area ratio data occupied by each type of environmental element, and finally, a spatial distribution parameter dataset is generated and output.
[0034] Finally, in step 1023, the spatial distribution parameters are used as a probability distribution. This process employs data normalization techniques to process the spatial distribution parameter dataset, transforming the spatial distribution parameters into type probability values. Then, through a weighted sum of probability logarithms, the type probability value data is processed using the entropy formula, and finally, the type probability value data is processed using the information entropy calculation formula. Where: n represents the total number of environmental element types; Indicates the first The type probability value of each environmental element; ln represents the natural logarithm operator. The final output is an information entropy value representing the degree of disorder in environmental complexity. The calculation process is based on the cumulative sum of probability logarithms, ultimately outputting an information entropy value representing the degree of disorder in environmental complexity.
[0035] In a practical application, during a dynamic evaluation experiment on the psychological recovery effect of open spaces in a community, researchers quantified the visual complexity of the environment through the following steps: When a random subject sat quietly in a waterside rest area, the system collected data at a high frequency of 20 times per second, utilizing the head coordinates and gaze direction vectors with an accuracy within 0.1 degrees simultaneously output by the subject's smart glasses, combined with environmental point cloud data generated by LiDAR scanning. Based on a dynamic bounding box set with horizontal azimuth and vertical pitch angles (e.g., horizontal angle range of ±60 degrees, vertical angle range of -30 degrees to +15 degrees), the system automatically updated a cone-shaped visual area model composed of massive 3D coordinate points every second. The system then precisely cut out a specific image block corresponding to this field of view from the panoramic image library. This block clearly showed scenes such as birdsong and swaying tree shadows. Using a deep learning segmentation model, the system identified the contour boundaries of six major categories of elements within the block, including vegetation, water, and paving, and was able to statistically determine the specific spatial distribution ratio of each type of element in the field of view at that time. Finally, the proportional data of these elements were substituted into the entropy calculation formula to obtain a quantitative value characterizing the degree of disorder in the scene. By observing the trend of this value over time, such as the obvious downward trend from higher to lower values, the study successfully revealed the correlation between the evolution of environmental visual complexity and the degree of attention recovery of the subjects.
[0036] In the overall scheme of step 102 above, a three-dimensional visual area is dynamically constructed by capturing the user's gaze direction and position information. Corresponding blocks in the panoramic image are precisely segmented, and the contours of various environmental elements within the visual focus area are automatically identified and classified. The spatial distribution density and area proportion of each element type are quantified and statistically analyzed. Based on the probability characteristics of element distribution, an information entropy model is used to calculate the dynamic complexity of the visual environment, revealing the changing patterns of disorder in environmental elements from a microscopic perspective. This transforms the originally abstract spatial visual experience into a continuously quantifiable information entropy index, providing fundamental data support for the evolution of visual complexity in subsequent modeling of the correlation between psychological recovery effects and environmental characteristics.
[0037] 103. Based on the magnitude of the deviation of the physiological signal from the user's resting state, calculate the physiological stress deviation intensity as the first type of prediction error; analyze the degree of matching between the movement trajectory data and the environmental spatial structure, quantify the path execution deviation as the second type of prediction error; and combine the eye movement characteristics in the eye movement data with the distribution map of environmental visual attraction to calculate the attention allocation deviation as the third type of prediction error. Optionally, step 103 may specifically include the following steps: 1031. Calculate the absolute difference between the amplitudes of the physiological signal and the skin conductance signal, and use it as the amplitude deviation between the physiological signal and the user's resting state reference value. After smoothing the amplitude deviation through a sliding window, the physiological stress deviation intensity is obtained as the first type of prediction error. 1032. Spatial matching is performed between the actual position sequence in the movement trajectory data and the node sequence of the predetermined path in the environmental spatial structure. The proportion of the number of offset nodes is used as the spatial matching difference. The path execution deviation is quantified by the spatial matching difference as the second type of prediction error. 1033. Combining the eye movement features in the eye movement data with the distribution relationship of high attraction areas in the environmental visual attraction distribution map, calculate the positional deviation ratio between the fixation point and the high attraction area, and use the positional deviation ratio to calculate the attention allocation deviation as the third type of prediction error.
[0038] In the above steps, the deviation of the physiological signal from the user's resting state refers to the measured difference between the physiological signal value and the pre-established resting state baseline value. The physiological stress deviation intensity refers to the quantified intensity value obtained after processing the physiological signal deviation, serving as the prediction error. Type I prediction error refers to prediction errors related to physiological stress. Movement trajectory data refers to the position sequence and time data of the user's movement in the target environment space. Environmental spatial structure refers to the node sequence representation data of the predetermined path in the target environment space. Matching degree refers to the spatial similarity measurement value between the actual position sequence and the node sequence of the predetermined path. Path execution deviation refers to the quantified degree of deviation of the actual path from the predetermined path, serving as the prediction error. Type II prediction error refers to prediction errors related to path execution. Eye movement data refers to the user's eye-tracking feature data, including fixation point location data. Environmental visual attractiveness distribution map refers to the distribution representation data of high-attractiveness areas in the target environment space. Attention allocation deviation refers to the degree of deviation between the fixation point location and the high-attractiveness area location, serving as the prediction error. Type III prediction error refers to prediction errors related to attention allocation. Sliding window smoothing refers to the process of averaging data within a time window using a moving average algorithm. Spatial matching refers to the process of comparing the actual location sequence with the predetermined path node sequence. The percentage of offset nodes refers to the proportion of nodes in the actual location sequence that deviate from the predetermined path node sequence. Spatial matching difference refers to the path similarity difference value calculated based on the percentage of offset nodes. Location deviation ratio refers to the proportion of distance between the fixation point location and the high-attraction area location.
[0039] In this embodiment, firstly, step 1031 calculates the absolute difference between the amplitudes of the physiological signal and the skin conductance signal as the amplitude deviation between the physiological signal and the user's resting state reference value. This process uses a difference algorithm to process the physiological signal data and the user's resting state reference value, and obtains the amplitude deviation data through the absolute difference calculation operation. Specifically, it includes: calculating the difference between the physiological signal data and the user's resting state reference value, then taking the absolute value to generate the amplitude deviation data, and then using a moving average algorithm to calculate the average value of the amplitude deviation data within a time window through a sliding window smoothing operation. Specifically, it includes: setting a fixed time window length, accumulating the amplitude deviation data within the time window, and then dividing by the number of data points within the time window to calculate the average value, and finally obtaining the physiological stress deviation intensity data output as the first type of prediction error.
[0040] Secondly, in step 1032, the actual position sequence in the movement trajectory data is spatially matched with the node sequence of the predetermined path in the environmental spatial structure. This process uses a spatial matching algorithm to process the movement trajectory data and the environmental spatial structure data. The actual position sequence is aligned with the node sequence of the predetermined path through a position comparison operation, and the number of nodes deviating from the predetermined path is calculated. Specifically, this includes: using the Euclidean distance formula to calculate the distance between each actual position and the nearest node of the predetermined path. If the distance exceeds a preset threshold, it is marked as a deviating node. Then, the total number of deviating nodes is counted. Then, the spatial matching difference is calculated based on the proportion of the number of deviating nodes. Specifically, this includes: calculating the proportion of the total number of deviating nodes to the total number of nodes in the predetermined path, using this proportion as the spatial matching difference. Finally, the path execution deviation data is quantified by the spatial matching difference and output as the second type of prediction error.
[0041] Finally, in step 1033, the positional deviation ratio between the fixation point and the high-attraction region is calculated by combining the eye movement features in the eye movement data with the distribution relationship of high-attraction regions in the environmental visual attraction distribution map. This process uses a positional deviation calculation algorithm to process the fixation point position data in the eye movement data and the high-attraction region position data in the environmental visual attraction distribution map. The positional deviation ratio is calculated through distance measurement operations such as Euclidean distance. Specifically, this includes calculating the distance between each fixation point position and the center of the nearest high-attraction region using the Euclidean distance formula, which is: ,in, The distance between each fixation point location and the center of the nearest high-attraction area For the first j The coordinates of the gaze point; The center coordinates of the nearest high-attraction area are determined, and the ratio of this distance to the preset maximum distance is calculated as the position deviation ratio data. Then, the attention allocation deviation is calculated based on the position deviation ratio. Specifically, the position deviation ratio data of all fixation points are averaged to obtain the average deviation ratio value, which is used as the attention allocation deviation. Finally, the attention allocation deviation data is output as the third type of prediction error.
[0042] In practical applications, in an empirical study on the psychological recovery mechanism of environmental space, researchers used a multimodal data analysis system to quantify behavioral prediction errors: When a subject completed a pre-set rest activity in a community water area, the system first continuously collected their skin conductance signals. When the current conductance amplitude exceeded the upper limit of the resting baseline range three times consecutively, the system determined that the subject was in a state of physiological stress. A sliding window was used to smooth the deviation values for 180 consecutive seconds, and the waveform curve of the physiological stress deviation intensity was output as the first type of prediction error. Simultaneously recorded movement trajectory data interacted with the pre-set environmental path: the system performed spatial matching calculations between the subject's actual location sequence and key nodes in the planned path. When the actual trajectory deviated by more than 5 meters at the pavilion node and the dwell time exceeded the preset limit, the path execution deviation was calculated based on the geometric characteristics of the proportion of the number of deviated nodes to the total number of path nodes, serving as the second type of prediction error. At the visual level, the system combined the overlay analysis of the gaze coordinate sequence collected by an eye tracker with the environmental visual attractiveness distribution map: the system pre-divided the space into tens of thousands of grid units and generated a heat map of high attractiveness areas based on a hotspot detection algorithm. When a subject's fixation point is relatively likely to be located outside the core visual focus area during a specific time period, the dispersion of the Euclidean distance between their eye position and the geometric center of the high-attention area is calculated. This dispersion is then used to output an attention allocation deviation index over time as a third type of prediction error. These three types of quantified errors together constitute the key indicator set for the difference between the predicted behavior model and the actual state.
[0043] In the overall scheme of step 103 above, the intensity of physiological stress deviation is calculated as the first type of prediction error by continuously monitoring the deviation of the user's physiological signals from the resting baseline. Simultaneously, the spatial matching difference between the movement trajectory sequence and the predetermined path of the spatial structure is analyzed to quantify the path execution deviation as the second type of prediction error. Furthermore, the deviation ratio of the fixation point position is calculated by coupling eye movement characteristics with the distribution data of visual attraction in the preset environment to determine the attention allocation deviation as the third type of prediction error. This multi-prediction error system dynamically captures the deviation patterns of users at the level of behavioral execution path decision-making mechanisms, achieving comprehensive quantitative modeling from physiological state anomalies to behavioral path deviations and then to visual attention deviations. This allows for in-depth exploration of abnormal patterns in the endogenous response characteristics during the psychological recovery process in open community spaces, providing multi-dimensional heterogeneous data fusion support for the subsequent construction of a high-precision dynamic prediction model for psychological recovery effects, and significantly improving the overall robustness and mechanistic interpretability of the evaluation method.
[0044] 104. Input the first type of prediction error, the second type of prediction error, the third type of prediction error and the information entropy value into the pre-constructed evaluation model to output the recovery efficiency index and the multidimensional mitigation index. The multidimensional mitigation index is used to characterize the degree of mitigation of each prediction error. Optionally, step 104 may specifically include the following steps: 1041. The first type of prediction error, the second type of prediction error, the third type of prediction error and the information entropy value are synchronously input into the inference interface of the pre-trained multi-channel fusion evaluation model. 1042. In the inference interface, the physiological stress bias intensity, path execution bias degree, and attention allocation bias degree are respectively generated into error feature vectors through independent fully connected layers, and the error feature vectors are cross-attention weighted with the information entropy feature vectors. Step 1042 may specifically include the following processes: Converting the physiological stress bias intensity, path execution bias, and attention allocation bias into error feature vectors of a unified dimension using independent feature converters, and integrating all error feature vectors into an error feature set; converting the information entropy value into an environmental entropy feature vector of a unified dimension using an environmental feature converter, calculating the sum of the corresponding element-wise products of each error feature vector in the error feature set and the environmental entropy feature vector, generating an interaction value set; dividing each interaction value in the interaction value set by the sum of the interaction value sets, generating an allocation weight set, and multiplying each weight in the allocation weight set by its corresponding error feature, generating a weighted feature set; and element-wise superimposing all weighted features in the weighted feature set with the environmental entropy feature vector, generating a cross-attention weight allocation result representing the fusion feature.
[0045] 1043. Based on the attention weight allocation results, output the scalar value of all feature channels as the recovery efficiency index, and generate three correction coefficients for the three types of prediction errors as multidimensional mitigation indicators.
[0046] In the above steps, the first type of prediction error refers to the physiological stress deviation intensity data as the prediction error type. The second type of prediction error refers to the path execution deviation data as the prediction error type. The third type of prediction error refers to the attention allocation deviation data as the prediction error type. Information entropy value refers to the quantified value of the disorder of environmental complexity. Recovery efficiency index refers to the comprehensive scalar value of the model output. Multidimensional mitigation index refers to three sets of correction coefficients as data characterizing the degree of mitigation of each prediction error. Pre-built evaluation model refers to a multi-channel fusion evaluation model. Inference interface refers to the model's data input and output interface. Independent fully connected layer refers to a neural network layer used for feature transformation. Error feature vector refers to the transformed vector data. Information entropy feature vector refers to the transformed information entropy value vector. Cross-attention weight allocation refers to the weight allocation mechanism process. Independent feature converter refers to the feature converter that processes errors. Error feature set refers to the set of data integrating all error feature vectors. Environmental feature converter refers to the feature converter that processes information entropy values. Environmental entropy feature vector refers to the feature vector after information entropy value transformation. Corresponding element-wise product sum refers to the dot product calculation operation. Interaction value set refers to the set of dot product results. The weight allocation set refers to the normalized weight set. The weighted feature set refers to the weighted feature vector set. Element superposition refers to the vector element addition operation. The cross-attention weight allocation result representing the fused features refers to the fused feature output after cross-attention. The recovery efficiency index refers to the scalar value of the final output. The multidimensional mitigation index refers to the three sets of correction coefficients in the output.
[0047] In this embodiment of the application, firstly, the first type of prediction error, the second type of prediction error, the third type of prediction error and the information entropy value are synchronously input into the inference interface of the pre-trained multi-channel fusion evaluation model through step 1041. This process uses model input interface technology to process the input data, realizes the synchronous data reception operation through the pre-trained evaluation model inference interface, and finally completes the process of inputting the prediction error data and information entropy value into the evaluation model.
[0048] Secondly, in step 1042, the physiological stress bias intensity, path execution bias, and attention allocation bias are generated into error feature vectors through independent fully connected layers in the inference interface. These error feature vectors are then cross-weighted with the information digest feature vectors. This process first utilizes independent feature converters to process the physiological stress bias intensity data, path execution bias data, and attention allocation bias data. Specifically, this includes applying an independent fully connected layer to each error data point and generating error feature vectors of a uniform dimension through a linear transformation formula. For example, for the physiological stress bias intensity data… The error eigenvector is ,in and These are the pre-trained weights and biases. Then, all error feature vector data are integrated to generate an error feature set, specifically including: […]. Path execution deviation corresponding to Corresponding to attention allocation deviation Combined into an error characteristic matrix Next, the information entropy value data is processed using an environmental feature converter, specifically including: converting the information entropy value... Transformed into an environmental entropy feature vector through a fully connected layer. ,in, The information value represents scalar input data that characterizes the complexity of the environment; Represents the trainable weight matrix, whose dimension is... The size of the output feature vector is determined by matrix multiplication. Perform linear scaling; This represents a trainable bias vector used to translate and adjust the features after weighting. The environmental feature vector representing the output is a A dimensional vector, formed by a linear transformation Insults and bias This is obtained by upgrading the scalar entropy value to a high-dimensional feature representation for subsequent model processing. Then, the sum of the element-wise products of each error feature vector data and the environmental entropy feature vector data in the error feature set is calculated, and the dot product algorithm is used to generate the interaction value set data. Specifically, this includes calculating the interaction value set data for each error feature vector. Extracting feature vectors from the environment dot product as interaction value ,in For vector dimensions, This is the set of interaction values. Next, each interaction value in the set is divided by the sum of the interaction values in the set, and a normalization algorithm is used to generate a weighted set of data. Specifically, this includes applying the softmax function to normalize the interaction value set. , is a vector containing three elements, each (in The index (ranging from 1 to 3) corresponds to an interaction value, used to describe a specific interaction attribute. The assigned weights are then calculated using a formula. ,in, This represents the weight values in the weight set data, indicating the corresponding weight output after normalization. It is the result of an exponential function operation, obtained by... The natural exponential function (exp) is applied to convert the original interaction value into a positive value to improve normalization stability; the denominator Represents the normalization factor, which is the sum of all interaction values. (index Iterate from 1 to 3, summing the results after exponential operations to ensure the sum of the output weights is 1; calculate each weight in turn. Dividing the exponential value by the sum of the denominators maps the set of interaction values to a set of weighted data in the form of a probability distribution. This facilitates weight allocation operations in subsequent models. Then, each weight data in the allocated weight set is multiplied by its corresponding error feature vector data using a weighting algorithm to generate a weighted feature set. Specifically, this includes: for each error feature vector... Multiply by the corresponding weight Generate weighted feature vectors And combine them into a weighted feature matrix. Finally, the weighted feature data in the weighted feature set data is element-wise superimposed with the environmental feature vector data. A vector addition algorithm is then used to generate the cross-attention weight allocation result representing the fused features. Specifically, this includes: [the process of] weighting the feature matrix... The fused feature vector is obtained by summing the elements of all vectors in the vector matrix. .
[0049] Finally, in step 1043, a scalar value integrating all feature channels is output as the recovery efficiency index based on the attention weight allocation result, and three correction coefficients are generated for the three types of prediction errors as multidimensional mitigation indicators. This process utilizes model output interface technology to process the cross-attention weight allocation result data, and calculates and outputs the scalar value of the recovery efficiency index through a comprehensive feature fusion algorithm. Specifically, this includes: fusing the feature vectors... Input fully connected layer and activation function output scalar value ,in For the sigmoid function, and These are the pre-trained parameters. Simultaneously, three correction coefficients corresponding to three classes of prediction error data are generated as a multi-dimensional mitigation index output, specifically including: processing and fusing the feature vector through three independent output branches. Generate correction coefficients ,in These correspond to the first, second, and third types of prediction errors, respectively. The fused feature vector is a high-order feature representation of the environmental entropy feature vector and the interaction feature vector after cross-attention processing, and serves as the shared input for the three branches; The values represent correction coefficients, which are scalar outputs corresponding to the prediction errors of the first, second, and third types, respectively, and are generated by linear transformation. Represents an independent, trainable weight vector (dimension) ), where subscript express The specific weights for each branch are used to adjust the fused features. Perform linear projection; Representing an independently trainable bias scalar, is The corresponding branch's dedicated translation term is added to the weighted result to adjust the output bias; this process is implemented through three well-aligned fully connected branches, each branch using the same fusion feature. For input, but using independent parameters and Generate separately Ultimately, this constitutes a multidimensional mitigation index to differentiate and correct the three types of prediction errors.
[0050] In practical applications, during an experiment assessing the effectiveness of restoring open spaces in a community, researchers used a dynamic monitoring system to achieve precise quantitative analysis. When subjects entered a landscape area for 15 minutes, their wrist sensors detected a peak skin conductivity fluctuation of 4.8 μS, exceeding the individual's resting baseline by 38%. The system smoothed this physiological deviation using a sliding window, generating a stress intensity curve lasting 120 seconds. Simultaneously collected movement trajectories showed that the average deviation between the subjects' actual stopping points and preset path nodes was 2.3 meters, with three key viewing nodes having deviations exceeding 5 meters. Based on this, the system calculated a path execution deviation of 42%. At the visual level, an eye tracker captured 237 fixation events. The five high-attraction areas marked in the environmental visual attractiveness model only attracted 31% of fixations, while 59% of fixations were scattered in peripheral areas. By calculating the Euclidean distance dispersion between fixation points and core attraction areas, the system quantified an attention allocation deviation of 68%. Meanwhile, based on segmented visual image blocks, the system identifies six environmental elements: water surface (35%), tree canopy (29%), and hard paving (18%), calculating an environmental complexity index of 2.7 using the information entropy formula. In the multimodal data fusion stage, the system simultaneously inputs three types of behavioral error parameters and the environmental entropy value of 2.7 into a 128-dimensional feature converter. At the cross-attention layer, the system automatically generates a dynamic weight matrix—physiological parameters receive a weight of 0.32, trajectory parameters 0.25, visual parameters 0.21, and environmental parameters 0.22. After weighted fusion, a comprehensive feature vector is formed. The final output is a dual-track result: the recovery efficiency index, mapped by the sigmoid function, is 0.67, while simultaneously generating three-dimensional correction coefficients—physiological relief coefficient 0.71, trajectory correction coefficient 0.63, and attention anchoring coefficient 0.58. This mechanism achieves dynamic diagnosis of spatial recovery efficiency with minute-level accuracy.
[0051] In the overall scheme of step 104 above, by synchronously tracking the user's physiological signal fluctuations, spatial movement trajectory characteristics, and eye gaze behavior, combined with the spatial distribution analysis of environmental elements within the dynamic visible area, an information entropy index representing environmental complexity is automatically generated; simultaneously quantifying three types of prediction errors—physiological stress deviation intensity, path execution difference, and visual attention deviation—and inputting them into a multi-channel fusion model for cross-attention weighted analysis, the final output is a comprehensive recovery efficacy index and mitigation coefficients for subdivided dimensions, realizing a dynamic and quantitative assessment of the psychological recovery effect of community open spaces, significantly improving the objectivity of the assessment results and the guidance value for space design optimization.
[0052] 105. By associating the spatial distribution parameters of the environmental elements with the multidimensional mitigation indicators, and extracting the critical values of key environmental interventions and the information entropy recovery effect relationship function through interpretable machine learning, and combining the dynamic recovery effect indicators, multidimensional mitigation indicators and the extracted rules, an evaluation result is generated.
[0053] Optionally, step 105 may specifically include the following steps: 1051. The spatial distribution parameters of the environmental elements are concatenated with the multidimensional mitigation index to form a set of associated features. The set of associated features is then processed by an interpretable machine learning model to extract the set of rule antecedents that meet the minimum support. Step 1051 may specifically include the following process: using the pixel area ratio of each environmental element as a spatial element configuration feature, using the three correction coefficients in the multidimensional mitigation index as mitigation effect features, and sequentially connecting the spatial element configuration feature and the mitigation effect feature along the feature dimension to form an associated feature set; traversing all environmental element type combinations in the associated feature set to generate candidate condition items, calculating the coverage ratio of each candidate condition item in the associated feature set sample, and retaining only candidate condition items whose coverage ratio is greater than a set threshold; performing rule confidence verification on the retained candidate condition items, calculating the improvement of the multidimensional mitigation index when the candidate condition item is valid, marking it as a valid condition item when the improvement exceeds the baseline fluctuation range, merging all the valid condition items to form a set of rule antecedents that meet the minimum support.
[0054] 1052. Extract from the set of rule antecedents the combination of environmental elements that significantly improves the multidimensional mitigation index as the critical value of key environmental intervention, and the functional relationship between information entropy value and recovery efficiency index as the information entropy recovery effect relationship function. 1053. Based on the critical value of the key environmental intervention and the relationship function of information entropy recovery effect, combined with the data of the dynamic recovery effect index, the state data of the multidimensional mitigation index and the parsed rules, an evaluation result including spatial optimization priority is generated.
[0055] In the above steps, the spatial distribution parameters of environmental elements refer to the dataset of pixel area proportions of various elements in the target environmental space. Multidimensional mitigation indices refer to the three sets of correction coefficients output by the evaluation model. Critical environmental intervention thresholds refer to the environmental element configuration condition thresholds that significantly improve mitigation effects. The information entropy recovery effect relationship function refers to the mathematical mapping relationship between information entropy values and recovery efficiency indices. Dynamic recovery effect indices refer to the changing recovery efficiency indices. Evaluation results refer to the analytical conclusions that include spatial optimization priorities. Feature association and concatenation refers to the operation of connecting different feature vectors by dimension to form composite features. The associated feature set refers to the concatenated composite feature dataset. An interpretable machine learning model refers to a machine learning model that can extract decision rules. The rule antecedent set refers to the minimum set of rules that satisfy the conditions. Spatial element configuration features refer to the feature representation of the pixel area proportion of environmental elements. Mitigation effect features refer to the feature representation of the three correction coefficients in the multidimensional mitigation indices. Candidate condition terms refer to the combination of environmental element types generated through traversal. Coverage ratio refers to the proportion of candidate condition terms appearing in the sample. Rule confidence verification refers to the process of verifying the correlation between condition terms and effect improvement. The magnitude of improvement refers to the change in the multidimensional mitigation index when a condition is met. The baseline fluctuation range refers to the random fluctuation range of the multidimensional mitigation index. Valid condition items refer to candidate condition items whose magnitude of improvement exceeds the baseline fluctuation range. Environmental element combination conditions refer to the combination of multiple environmental element configuration constraints. Spatial optimization priority refers to the priority order of environmental modification recommendations.
[0056] In this embodiment of the application, firstly, step 1051 involves concatenating the spatial distribution parameters of the environmental elements with the multidimensional mitigation index to form a associated feature set. This process uses the pixel area ratio data of each environmental element as a spatial element configuration feature and the three correction coefficients in the multidimensional mitigation index as mitigation effect features, sequentially concatenating them along the feature dimensions. A vector concatenation algorithm is then used to generate the associated feature set data. Next, all combinations of environmental element types in the associated feature set are traversed to generate candidate condition item data. Specifically, this includes: for each candidate condition item (such as an element combination)... ), calculate its coverage ratio in the associated feature set samples, the formula is coverage ratio The sample unit is a single observation record; only records with a coverage ratio greater than a set threshold are retained. (like The candidate condition data is then used to verify the rule confidence of the retained candidate condition items, specifically including: calculating the average value of the multidimensional mitigation index when the candidate condition item is true. The average value when it is not valid The increase is When the improvement exceeds the baseline fluctuation range, the data is marked as valid conditional data. Specifically, the baseline fluctuation range is defined as the overall standard deviation of the multidimensional mitigation index. times (e.g.) ), that is, if Then the mark is valid, where The overall standard deviation representing the multidimensional mitigation index is the statistical fluctuation range of the index across all sample points in the historical dataset; finally, all valid condition terms are merged to form the rule antecedent set data output.
[0057] Secondly, in step 1052, key environmental intervention thresholds and information entropy recovery effect relationship functions are extracted from the rule antecedent set. This process analyzes the environmental element combination condition data in the rule antecedent set, and extracts the environmental element combination condition with the highest improvement rate of multidimensional mitigation indicators as the key environmental intervention threshold output by sorting by the improvement rate. Specifically, this includes: sorting all effective condition items by improvement rate. Sort in descending order, select The maximum combination condition is used as the critical value, and the output is a combination of feature type and scale threshold (e.g., area scale of type A > 100%). Simultaneously fitting the information entropy value With recovery efficiency index The functional relationship generates the information entropy recovery effect relationship function output, specifically including: fitting the function using a linear regression algorithm. , where the coefficient and Calculated using the least squares method: , and The coefficients represent linear regression, and the slope (the strength of the effect of entropy on efficiency) and intercept (the baseline recovery level), respectively. This represents the number of samples, i.e., the total number of data points used for fitting. and Representing the The information extract and recovery efficiency index of each sample constitute the fitted dataset. This process is implemented using the least squares method: first, the predicted values of all samples are calculated. Next, calculate its value compared to the actual value. The squared difference is then iteratively optimized using algorithms such as gradient descent. and To minimize the sum of squared differences, the final output is a linear function. As a relational function of information entropy recovery effect.
[0058] Finally, in step 1053, an evaluation result is generated based on the key environmental intervention threshold and the information extraction recovery effect relationship function, combined with dynamic recovery effect indicators and multidimensional mitigation indicators. This process utilizes the key environmental intervention threshold to constrain the configuration conditions of environmental elements, specifically including: checking whether the current proportion of environmental element types meets the threshold condition (e.g., determining the proportion of type A). ), generate compliance flags; calculate the theoretical value of the recovery effectiveness index through the information digest recovery effect relationship function, specifically including: the current information digest value Input the relational function to calculate the theoretical value. Combined with dynamic recovery effect indicators The data and state data of multidimensional mitigation indicators are used to analyze the optimization rules in the set of rule antecedents, specifically including: comparing theoretical values. Compared with actual value deviation The well selects the increase magnitude from the set of rule antecedents. The largest rule is selected as the priority optimization term; an evaluation result output containing spatial optimization priorities is generated, specifically including: prioritizing the optimization rules according to... Sort the data, output a list of feature types and adjust the direction (e.g., increase the proportion of type B) as the evaluation result.
[0059] In practical applications, in a dynamic evaluation experiment of the intervention effect in a community open space, researchers used an intelligent analysis system to optimize spatial elements. When continuous monitoring data accumulated for 42 minutes, the system automatically correlated the distribution status of environmental elements with behavioral correction effect parameters. Within a spatial unit where tree canopy coverage was stably maintained at 38% and flowering vegetation patches accounted for 12%, the physiological recovery coefficient in the multidimensional mitigation index reached a peak of 0.68, while the trajectory optimization coefficient exceeded the 0.7 threshold. The system, through an interpretable rule engine, deeply scanned 156 combinations of spatial behavioral features and found that under the configuration conditions where the water area ratio exceeded the critical value of 24% and the tree canopy continuity was greater than 75%, the three types of behavioral correction coefficients showed a synchronous leap: the attention focus aggregation effect was particularly significant, improving intervention effectiveness by nearly 30% compared to the baseline state. The system simultaneously analyzed the correlation between environmental complexity and recovery effect, establishing a quantitative response model of entropy change curves and a comprehensive recovery effectiveness index, clarifying that optimal recovery benefits can be obtained when visual complexity is controlled within a specific gradient range. Based on monitoring data from the northeastern landscape area, the current water body accounts for 18% of the area, and the entropy value exceeds the limit by 0.6 units. The system automatically triggers a spatial optimization early warning and generates a priority intervention plan: it is recommended to reduce spatial complexity through water body morphology modification, which is expected to improve the area's restoration efficiency index from the current level of 0.78 to 0.87. The final output includes an assessment report containing 5 key modification units and 12 key element adjustment parameters, providing precise guidance for upgrading the healing function of the community's open spaces.
[0060] In the overall scheme of step 105 above, a feature association set of spatial distribution parameters of environmental elements and multidimensional mitigation indicators is constructed, and an interpretable machine learning framework is used to analyze key laws: First, based on the feature concatenation of the element pixel area ratio and mitigation correction coefficient, candidate condition items are generated by traversing all combinations of environmental elements. Valid condition items with a coverage ratio higher than a set threshold and a mitigation improvement exceeding the baseline fluctuation are selected to form a set of rule antecedents. From this, critical values for environmental element configurations that significantly improve mitigation effects are extracted, and a nonlinear relationship function between information entropy and recovery efficiency index is simultaneously fitted. Finally, by combining dynamic recovery indicator data, multidimensional mitigation status, and analytical rules, an evaluation result containing spatial optimization priority ranking is automatically generated. This quantitatively reveals the psychological recovery contribution laws of core environmental elements such as the intervention threshold of vegetation coverage interface morphology combinations and the proportion of water system space, achieving a closed loop of precise design decision-making driven by environmental psychology mechanisms.
[0061] The following is a complete embodiment for steps 101 to 105: like Figure 2 As shown, in an experiment evaluating the psychological recovery effect in a community open space, researchers used a multimodal perception system to achieve full-process quantitative analysis. When a user sat quietly in a rest area surrounded by vegetation, a smart bracelet continuously collected their skin conductance signals, with a baseline value of 1.2 microsiemens. The system detected an instantaneous value climbing to 3.8 microsiemens, a deviation of 217% from the baseline value. Based on this, the system automatically calculated a physiological stress deviation intensity of 0.38 as the first type of prediction error. Simultaneously, the foot positioning module recorded and showed that the user's movement trajectory within 15 minutes contained 42 coordinate points. After comparison with the preset path nodes in the environment, three key deviations were found, with a maximum deviation of 4.7 meters. The system quantified the path execution deviation as a second type of prediction error of 0.42. The eye tracker captured 205 gaze events. After comparing them with 5 pre-set high-attraction core areas in the environment, it was found that the water spot area accounted for 35% of the core area, while the actual 59% of the gaze points were distributed in the non-core area, with an average distance difference of 2.8 meters. Based on this, an attention allocation deviation of 0.63 was generated as the third type of prediction error.
[0062] The dynamic gaze tracking system uses data from the user's head azimuth angle (±55 degrees) and pitch angle (±25 degrees) to segment a 1024 x 768 pixel visual field block. Six environmental element parameters are extracted through image recognition, including tree canopy coverage (38%) and hard paving (22%), and an environmental information entropy value of 2.6 is calculated based on the element morphological complexity. In the multimodal data fusion stage, the system simultaneously inputs three types of prediction error data (0.38, 0.42, and 0.63) and the environmental entropy value of 2.6 into the evaluation model. In the 128-dimensional feature space, a cross-attention mechanism assigns a weight of 0.32 to physiological indicators, 0.28 to trajectory parameters, 0.23 to visual parameters, and 0.17 to environmental parameters. After fully connected layer computation, a dual-track result is output: a recovery efficiency index of 0.71 (corresponding to a score of 84.3 on a percentage scale), and a three-dimensional relief index is generated, including a physiological relief coefficient of 0.74, a trajectory optimization coefficient of 0.68, and an attention anchoring coefficient of 0.61.
[0063] Deep correlation analysis revealed that when tree cover exceeded the 35% threshold and flowering vegetation density reached 12.5%, the system detected a support rate of 31% for this combination of conditions in 217 samples, at which point the mitigation coefficient showed an average jump of 0.21. The decision rule extraction engine identified two core criteria: a water area ratio of 24% is the optimal recovery critical point, at which point the entropy value stabilizes in the range of 1.8 to 2.2; when the hard interface ratio exceeds 40%, a recovery efficiency decay inflection point is triggered, with a decay slope of 0.15 decrease in efficiency for every 10% increase in interface ratio. The final evaluation report recommends increasing canopy density to improve coverage, which is expected to improve the recovery efficiency index from the current value, forming a quantitative spatial optimization scheme. Optionally, the method further includes: collecting multiple sets of subjective evaluation data through a semantic difference scale, and using a large language model to convert the subjective evaluation data into natural language descriptions; inputting the natural language descriptions into a text encoder to generate text embedding vectors, and constructing a dual-encoder variational autoencoder framework, constructing a semantic shared latent space through the text encoder and decoder, and minimizing the KL divergence between the visual indicator distribution and the text embedding distribution while freezing the text encoder or decoder; establishing a mapping from the text embedding vectors to the psychological recovery effect evaluation results in the semantic shared latent space, and enabling the text input path to directly encode the natural language descriptions into the latent space; predicting the recovery score results through a regression model, and simultaneously enabling the semantic difference scale input path to convert the structured semantic difference scale scores into natural language descriptions, and then predicting the recovery scores through the text path.
[0064] In the above steps, the semantic dissimilarity scale refers to a standardized questionnaire tool used to measure subjective perceptual differences. Subjective evaluation data refers to the dataset obtained by users rating their subjective feelings about the target environmental space using the semantic dissimilarity scale. The large language model refers to an artificial intelligence model, such as the Transformer architecture model, capable of understanding and generating natural language. Natural language description refers to the coherent textual narrative transformed from structured scores by the large language model. The text encoder refers to the component in a neural network structure that processes text input and generates vector representations. Text embedding vectors refer to the numerical vector representation data output by the text encoder. The dual-encoder variational autoencoder framework refers to a variational autoencoder machine learning architecture containing two encoders. The semantic shared latent space refers to the latent representation space region shared by multiple modalities. Freezing refers to an optimization technique that fixes certain component parameters during model training. The visual indicator distribution refers to the probability distribution representation based on visual features such as environmental image data. The text embedding distribution refers to the probability distribution representation data of text embedding vectors. KL divergence is a quantitative indicator that measures the degree of difference between two probability distributions. The psychological recovery effect evaluation result refers to the data evaluating the recovery efficacy index or multidimensional relief index in the model. Mapping refers to establishing a functional relationship model between input vectors and output variables. The text input path refers to the input channel through which natural language descriptions are directly input into the model. A regression model is a machine learning model used to predict continuous value outputs. The semantic difference scale input path refers to an alternative input channel through which raw structured scores are input into the model. The recovery score result refers to the continuous value score of the recovery effect predicted by the regression model.
[0065] In this embodiment, multiple sets of subjective evaluation data are first collected using a semantic difference scale, and then the subjective evaluation data is converted into natural language descriptions using a large language model. This process utilizes questionnaire collection technology to distribute the semantic difference scale to users in the target environment to obtain subjective evaluation data. Next, the subjective evaluation data is processed using a large language model algorithm, such as a pre-trained Transformer model. Natural language generation operations are then used to convert the structured ratings into coherent natural language descriptive text output. Specifically, this includes: inputting the subjective evaluation data into the decoder component of the Transformer model, generating a sequence of lexical units one by one using an autoregressive sampling algorithm, with the calculation formula being: the probability distribution of the next lexical unit. (input data), where Indicates the currently generated word element. The first step involves representing the historical word sequence and outputting the generated natural language description text, providing foundational data for subsequent text processing. Secondly, a natural language description input text encoder generates text embedding vectors. This process utilizes a text encoder architecture such as BERT or LSTM to process the natural language description text. The encoder's forward propagation algorithm extracts semantic features and generates fixed-dimensional text embedding vector data output. Specifically, this includes performing forward propagation calculations; the input text word sequence is processed through multiple layers of the encoder network. In the BERT model, the calculation formula is: [Formula omitted for brevity]. Layer output FeedForward ,in The initial embedding vector is used; finally, average pooling is applied to the output of the last hidden layer to generate the text embedding vector, with the formula: Embedding Vector ,in The number of lexical units, For the first The hidden states of each word are used as vector representations for the next step of building the model framework.
[0066] Next, a dual-encoder variational autoencoder framework is constructed. This framework builds a semantically shared latent space using a text encoder and decoder, and minimizes the KL divergence between the visual indicator distribution and the text embedding distribution while freezing either the text encoder or decoder. The process begins by initializing the dual-encoder variational autoencoder framework structure, including a text encoder component and an image decoder component. A parameter configuration algorithm is used to establish a link between the encoder output and the shared latent space. Then, some parameters of the text encoder or image decoder are frozen using parameter fixing techniques to prevent gradient updates. Finally, a KL divergence minimization algorithm, such as gradient descent, is used to calculate the distribution difference between the visual indicator distribution and the text embedding distribution and minimize this difference to align the distributions. Specifically, the formula for calculating the KL divergence value is: ,in Indicates the distribution of visual indicators. The text embedding distribution is represented; then the unfrozen parameters are updated using the gradient descent algorithm, with the optimization objective being to minimize the divergence value. The iterative formula is: parameter update. ,in For trainable parameters, The learning rate is used to output the optimized framework.
[0067] Finally, a mapping from text embedding vectors to the psychological recovery effect evaluation results is established in the semantic shared latent space, and the text input path is enabled to directly encode natural language descriptions into the latent space. This process first establishes a mapping relationship between the optimized text embedding vectors and the psychological recovery effect evaluation results within the semantic shared latent space through mapping functions such as linear regression or fully connected layers. Specifically, this includes applying a fully connected layer function, with the calculation formula as follows: Evaluation Results ,in The optimized text embedding vector. and For the weight matrix and bias vector, For activation functions such as sigmoid, the output is a mapping model; simultaneously, enabling text input paths allows for channel configuration, directly inputting natural language descriptions into the text encoder component for encoding into the shared latent space. Specifically, this includes: activating the text input path to connect the text encoder output to the shared latent space, generating latent representation data using the formula: Latent Representation Encoder (natural language description), where Encoder is a text encoder component.
[0068] In practical applications, in a study evaluating the efficacy of psychological recovery in open community spaces, researchers innovatively integrated multi-source data stream acquisition and analysis technologies. They systematically acquired user subjective experience data through a specially designed semantic difference scale, and then used an advanced large language model to automatically transform these structured evaluations into concrete natural language descriptive text. These descriptive texts were processed by a dedicated text encoder to generate high-dimensional text embedding vectors, and a dual-channel variational autoencoder framework was simultaneously constructed to achieve cross-modal semantic fusion. During training, fixed text encoder-decoder parameters were used to construct a shared latent semantic space, and the matching degree between the distribution of visual physiological indicators and the probability distribution of the text embedding space was continuously optimized, ultimately compressing the KL divergence to 0.32 to achieve cross-modal semantic alignment. Within this joint framework, a direct mapping mechanism from text embedding vectors to psychological recovery assessment results was established, forming the predictive capability of the natural language input channel. During the validation phase, the system used two prediction paths in parallel: on the one hand, the subjects' free descriptive text was directly input, and the psychological recovery score was output after encoding conversion and regression calculation; on the other hand, the original scores of the semantic difference scale were automatically translated into standardized descriptive text, achieving score prediction through the same text channel. Key data shows that the correlation coefficient of the dual-path prediction results remained stable at 0.89. The natural language path captured an additional 22% of psychological details not covered by the scale, successfully revealing the potential correlation between visual dynamic features and emotional description statements, and providing quantitative guidance for community space design.
[0069] The above steps involve collecting multiple sets of subjective evaluation data by introducing a semantic difference scale, transforming them into natural language descriptions using a large language model, and constructing a dual-encoder variational autoencoder framework to achieve deep fusion of subjective and objective data. Specifically, a text encoder maps natural language descriptions into text embedding vectors, aligning them with the distribution of objective visual indicators in a semantically shared latent space using KL divergence minimization to establish a mapping relationship between text embeddings and psychological recovery effects. A dual-path input mechanism is simultaneously designed: the natural language path directly encodes subjective descriptions into the latent space, while the scale path converts structured scores into text and predicts recovery scores. Finally, a regression model outputs psychological recovery scores, generating a psychological recovery correlation map between environmental features and subjective perceptions in the latent space. This overcomes the limitations of separating subjective and objective data in traditional assessments, significantly improving the ecological validity and interpretive depth of the assessment results, and realizing a multi-source heterogeneous data collaborative computation framework for the psychological recovery effects of environmental space. Optionally, the method further includes: discretizing the recovery efficacy index into five recovery level labels, the five recovery level labels including extremely low recovery level, low recovery level, medium recovery level, high recovery level, and extremely high recovery level; using the spatial distribution parameters of the environmental elements as input features and the five recovery level labels as prediction targets, training a classification model based on a rule representation learner; extracting a weighted decision rule set from the classification model, the decision rule set containing the mapping relationship between environmental element threshold conditions and recovery levels; analyzing the key environmental intervention critical values of the multidimensional mitigation index according to the decision rule set, fitting a nonlinear relationship function between the information entropy value and the recovery efficacy index, and outputting the key environmental intervention critical values and the nonlinear relationship function as interpretable rules.
[0070] In the above steps, discretization refers to the process of dividing the continuous recovery efficacy index into discrete levels. The five-level recovery level labels include labels representing five levels of recovery effectiveness: extremely low recovery level, low recovery level, medium recovery level, high recovery level, and extremely high recovery level. Input features refer to the input variable data of the training model, such as the spatial distribution parameters of environmental elements. The prediction target refers to the output category predicted by the model, such as the recovery level label. A classification model based on a rule representation learner refers to a machine learning model that can extract interpretable rules, such as a decision tree architecture. Training refers to the process of optimizing model parameters. The decision rule set refers to the extracted set of rules. The weighted decision rule set refers to the set of rules with attached weight values. Environmental element threshold conditions refer to the critical constraint conditions of the element proportion values in the rules. The recovery level mapping relationship refers to the correlation and correspondence between environmental element threshold conditions and recovery level labels. The critical environmental intervention threshold value refers to the environmental element threshold condition data that significantly improves the multidimensional mitigation indicators. Fitting refers to the process of estimating function parameters. The nonlinear relationship function refers to the complex mathematical mapping function between the information entropy value and the recovery efficacy index. Interpretable rules refer to the readable decision rule data of the output.
[0071] In this embodiment, the recovery efficiency index is first discretized: the continuous recovery efficiency index is divided into five level intervals using the equal interval method. Specifically, the dataset range of the recovery efficiency index is determined. ,in To restore the efficiency index to its minimum value, To restore the maximum value of the efficiency index, the boundary points are then calculated. ,in For boundary point indices, take values Indicates the first The boundary points were used to divide the recovery efficiency index into five levels, such as the interval... This corresponds to "extremely low recovery level". Secondly, a decision tree classification model is trained using the spatial distribution parameters of environmental elements as input features and the five recovery levels as prediction targets: during recursive node splitting, the Gini impurity is calculated. ,in To restore the hierarchical index, take the value. These correspond to five levels respectively. The level in the current node The sample proportion The Gini impurity of the representative node is then used to select a feature threshold combination that maximizes the reduction of Gini impurity to split the node. The formula for calculating the reduction is as follows: ,in The Gini value of the parent node. Identifies the left or right branch after the split, with a value of "left" or "right". For branches The set of child node samples, The total number of samples in the parent node. This represents the Gini value of the child nodes. After generating the fitted model, the decision rules are extracted by traversing the backtracking path of the leaf nodes, such as "the proportion of element A > 0". "High recovery level", and assign weights to each rule. Accuracy, of which denoted as the number of leaf node samples, and denoted as the accuracy scalar value representing the model evaluation precision. This represents the rule weights. Finally, the critical values for key interventions are analyzed. ,in This refers to the threshold for the proportion of environmental elements in the rules. To determine the minimum critical value, a quadratic polynomial regression was used to fit the information entropy value. With recovery efficiency index Relationship, generating function ,in These are the regression coefficients; this nonlinear function and the critical value are ultimately output as interpretable rule data. The entire process integrates steps from data partitioning, decision tree training, rule weighting to nonlinear fitting, with all symbols strictly adhering to their definitions to maintain consistency.
[0072] In practical applications, researchers systematically constructed an interpretable decision-making model in a study analyzing intervention mechanisms in community open spaces. First, the continuously monitored recovery efficiency index was discretized into five recovery level labels based on specific threshold ranges. For example, an efficiency index below 0.25 was labeled as extremely low recovery, 0.25-0.45 as low recovery, and so on up to 0.85 as extremely high recovery. Then, using spatial distribution parameters of environmental elements as input features, including quantitative indicators such as tree canopy coverage and water area ratio, and using the five recovery levels as prediction targets, a rule-based classification model was trained. This model automatically extracted a set of key decision rules through an iterative rule generation mechanism. For example, when tree coverage exceeded 35% and flowering vegetation density was higher than 2.6 plants per square meter, the model predicted a high recovery level with 78% confidence; when the hard surface area ratio exceeded 40% and the visual entropy value was higher than 3.2, it was associated with a low recovery level. The model further analyzes the environmental intervention thresholds implicit in the rule set, clearly identifying the key inflection point where the water body edge morphology index exceeds 1.7, resulting in a step-like increase in restoration efficiency. Simultaneously, through nonlinear fitting techniques, it reveals a specific functional relationship between environmental information entropy and the restoration efficiency index: when the entropy value is maintained in the 1.8-2.3 range, restoration efficiency enters the optimal growth zone. The final interpretable rule package integrates 12 element threshold conditions and a 3D entropy-efficiency response surface, accurately guiding the optimization of community spatial element configuration. Experimental verification shows that the model achieves an accuracy of 85% in predicting level transformations and successfully identifies key modification parameters for increasing the water area in the core area.
[0073] The above steps discretize the continuous recovery efficacy index into five recovery level labels, construct a rule classification model of environmental element distribution parameters and recovery levels, and realize the visual extraction of psychological recovery assessment rules. First, based on the dynamic recovery efficacy index, five recovery levels from extremely low to extremely high are defined. The spatial distribution parameters of environmental elements are used as input features to train a rule representation learner. A path backtracking mechanism is used to extract a set of decision rules with confidence weights, forming a mapping relationship between environmental threshold conditions and recovery levels. By analyzing the correlation between element configuration and recovery levels in the rule set, key environmental intervention thresholds are extracted, and a nonlinear relationship function between information entropy and the recovery efficacy index is established simultaneously. Finally, an interpretable rule base integrating environmental intervention thresholds and psychological recovery mechanisms is output.
[0074] Figure 3 This application provides a schematic diagram of the structure of an evaluation system for the psychological recovery effect of community open spaces based on multimodal perception, as shown in the embodiments of this application. Figure 3 As shown, the system includes: The acquisition module 31 acquires the user's physiological signals, movement trajectory data, eye movement data, and panoramic image data of the target environment space in the target environment space. The parsing module 32 dynamically determines the visible area range based on the user's position and line of sight, parses the panoramic image data, extracts the spatial distribution parameters of environmental elements within the visible area, and calculates the information entropy value representing the complexity of the environment based on the spatial distribution parameters. The calculation module 33 calculates the physiological stress deviation intensity based on the magnitude of the deviation of the physiological signal from the user's resting state, as the first type of prediction error; analyzes the degree of matching between the movement trajectory data and the environmental spatial structure, quantifies the path execution deviation, as the second type of prediction error; and combines the eye movement characteristics in the eye movement data with the environmental visual attraction distribution map to calculate the attention allocation deviation, as the third type of prediction error. The construction module 34 inputs the first type of prediction error, the second type of prediction error, the third type of prediction error and the information entropy value into the pre-constructed evaluation model to output the recovery efficiency index and the multidimensional mitigation index. The multidimensional mitigation index is used to characterize the degree of mitigation of each prediction error. The generation module 35 associates the spatial distribution parameters of the environmental elements with the multidimensional mitigation indicators, extracts the critical values of key environmental interventions and the information entropy recovery effect relationship function through interpretable machine learning, and generates evaluation results by combining the dynamic recovery effect indicators, multidimensional mitigation indicators and the extracted rules.
[0075] Figure 3 The aforementioned evaluation system for the psychological recovery effect of community open spaces based on multimodal perception can be executed. Figure 1 The implementation principle and technical effects of the evaluation method for the psychological recovery effect of community open spaces based on multimodal perception described in the illustrated embodiment will not be repeated here. The specific methods of operation of each module and unit in the evaluation system for the psychological recovery effect of community open spaces based on multimodal perception in the above embodiments have been described in detail in the embodiments related to this method, and will not be elaborated upon here.
[0076] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.
Claims
1. A method for evaluating the psychological recovery effect of community open spaces based on multimodal perception, characterized in that, include: Within the target environment, acquire the user's physiological signals, movement trajectory data, eye movement data, and panoramic image data of the target environment. Based on the user's position and line of sight, the visible area is dynamically determined, the panoramic image data is analyzed, the spatial distribution parameters of environmental elements within the visible area are extracted, and the information entropy value representing the complexity of the environment is calculated based on the spatial distribution parameters. Based on the magnitude of the deviation of the physiological signal from the user's resting state, the intensity of physiological stress deviation is calculated as the first type of prediction error. The degree of matching between the movement trajectory data and the environmental spatial structure is analyzed to quantify the path execution deviation as the second type of prediction error. Combining the eye movement characteristics in the eye movement data with the distribution map of environmental visual attraction, the degree of attention allocation deviation is calculated as the third type of prediction error. The first type of prediction error, the second type of prediction error, the third type of prediction error, and the information entropy value are input into the pre-constructed evaluation model to output the recovery efficiency index and the multidimensional mitigation index. The multidimensional mitigation index is used to characterize the degree of mitigation of each prediction error. By associating the spatial distribution parameters of the environmental elements with the multidimensional mitigation indicators, and using interpretable machine learning to extract the critical values of key environmental interventions and the information entropy recovery effect relationship function, and combining the dynamic recovery effect indicators, multidimensional mitigation indicators and the extracted rules, an evaluation result is generated.
2. The method according to claim 1, characterized in that, Also includes: Multiple sets of subjective evaluation data were collected using a semantic difference scale, and the subjective evaluation data were transformed into natural language descriptions using a large language model. The natural language description is input into the text encoder to generate text embedding vectors, and a dual encoder variational autoencoder framework is constructed. A semantically shared latent space is built through the text encoder and decoder, and the KL divergence between the visual index distribution and the text embedding distribution is minimized when the text encoder or decoder is frozen. In the semantic shared latent space, a mapping is established between the text embedding vector and the psychological recovery effect evaluation result, and the text input path is enabled to directly encode the natural language description into the latent space; The recovery score is predicted by a regression model, and the input path of the semantic difference scale is enabled. After the structured semantic difference scale score is converted into a natural language description, the recovery score is predicted by the text path.
3. The method according to claim 1, characterized in that, Also includes: The recovery efficiency index is discretized into five levels of recovery level labels, which include extremely low recovery level, low recovery level, medium recovery level, high recovery level and extremely high recovery level; Using the spatial distribution parameters of the environmental elements as input features and the five-level recovery level labels as prediction targets, a classification model based on a rule representation learner is trained. A weighted set of decision rules is extracted from the classification model, the set of decision rules containing the mapping relationship between environmental element threshold conditions and recovery levels; Based on the decision rule set, the critical values of key environmental interventions for multidimensional mitigation indicators are analyzed, and a nonlinear relationship function between information entropy and recovery efficiency index is fitted. The critical values of key environmental interventions and the nonlinear relationship function are then used as interpretable rule outputs.
4. The method according to claim 1, characterized in that, The first type of prediction error, the second type of prediction error, the third type of prediction error, and the information entropy value are input into a pre-constructed evaluation model to output a recovery efficiency index and a multidimensional mitigation index. The multidimensional mitigation index is used to characterize the degree of mitigation of each prediction error, including: The first type of prediction error, the second type of prediction error, the third type of prediction error, and the information entropy value are synchronously input into the inference interface of the pre-trained multi-channel fusion evaluation model. In the inference interface, the physiological stress bias intensity, path execution bias degree, and attention allocation bias degree are generated into error feature vectors through independent fully connected layers, and the error feature vectors are cross-attention weighted with the information entropy feature vectors. Based on the attention weight allocation results, the output scalar value of all feature channels is used as the recovery efficiency index, and three correction coefficients are generated for the three types of prediction errors as multidimensional mitigation indicators.
5. The method according to claim 4, characterized in that, The physiological stress bias intensity, path execution bias, and attention allocation bias are each processed through an independent fully connected layer to generate error feature vectors. These error feature vectors are then cross-weighted with the information entropy feature vector, including: The physiological stress deviation intensity, path execution deviation, and attention allocation deviation are converted into error feature vectors of a unified dimension by independent feature converters, and all error feature vectors are integrated into an error feature set. The information entropy value is converted into an environmental entropy feature vector of a unified dimension by an environmental feature converter. The sum of the corresponding element-wise products of each error feature vector in the error feature set and the environmental entropy feature vector is calculated to generate an interaction value set. Divide each interaction value in the set of interaction values by the sum of the interaction value sets to generate a weighted set, and multiply each weight in the weighted set by the corresponding error feature to generate a weighted feature set; The weighted features in the weighted feature set are element-wise superimposed with the environmental entropy feature vector to generate the cross-attention weight allocation result representing the fused features.
6. The method according to claim 1, characterized in that, By associating the spatial distribution parameters of the environmental elements with the multidimensional mitigation indicators, and extracting key environmental intervention thresholds and information entropy recovery effect relationship functions through interpretable machine learning, and combining the dynamic recovery effect indicators, multidimensional mitigation indicators, and extraction rules, an evaluation result is generated, including: The spatial distribution parameters of the environmental elements are concatenated with the multidimensional mitigation index to form a set of associated features. The set of associated features is then processed by an interpretable machine learning model to extract a set of rule antecedents that meet the minimum support. The environmental element combination conditions that significantly improve the multidimensional mitigation index are extracted from the set of rule antecedents as key environmental intervention thresholds, and the functional relationship between information entropy value and recovery efficiency index is used as the information entropy recovery effect relationship function. Based on the critical values of key environmental interventions and the relationship function of information entropy recovery effect, combined with the data of the dynamic recovery effect index, the state data of the multidimensional mitigation index, and the parsed rules, an evaluation result including spatial optimization priority is generated.
7. The method according to claim 6, characterized in that, The spatial distribution parameters of the environmental elements are concatenated with the multidimensional mitigation indicators to form a correlated feature set. This correlated feature set is then processed using an interpretable machine learning model to extract a set of rule antecedents that satisfy the minimum support, including: The pixel area ratio of each environmental element is used as the spatial element configuration feature, and the three correction coefficients in the multidimensional mitigation index are used as the mitigation effect feature. The spatial element configuration feature and the mitigation effect feature are connected sequentially along the feature dimension to form an associated feature set. All combinations of environmental element types are traversed in the associated feature set to generate candidate condition items. The coverage ratio of each candidate condition item in the associated feature set sample is calculated, and only candidate condition items with a coverage ratio greater than a set threshold are retained. The retained candidate conditions are verified by rule confidence. The improvement of the multidimensional mitigation index when the candidate condition is true is calculated. When the improvement exceeds the baseline fluctuation range, it is marked as a valid condition. All the valid condition conditions are merged to form a set of rule antecedents that meet the minimum support.
8. The method according to claim 1, characterized in that Based on the user's position and line of sight, the visible area is dynamically determined, the panoramic image data is analyzed, spatial distribution parameters of environmental elements within the visible area are extracted, and information entropy values characterizing environmental complexity are calculated based on these spatial distribution parameters, including: Based on user position and line of sight data, the horizontal azimuth and vertical pitch boundaries of the line of sight are determined, and combined with depth sensor data, the three-dimensional visible area range is dynamically constructed. In the panoramic image data, image blocks corresponding to the visible area are cut out, the outline boundaries of all environmental elements in the image blocks are extracted, and the spatial distribution parameters occupied by each type of element are statistically analyzed according to the element attribute type. The spatial distribution parameters are used as a probability distribution, and the information entropy value, which characterizes the degree of disorder in environmental complexity, is obtained by weighted summation of probability logarithms.
9. The method according to claim 1, characterized in that, Based on the magnitude of the deviation of the physiological signals from the user's resting state, the intensity of physiological stress deviation is calculated as the first type of prediction error. The degree of matching between the movement trajectory data and the environmental spatial structure is analyzed to quantify the path execution deviation as the second type of prediction error. Combining the eye movement characteristics in the eye movement data with the environmental visual attraction distribution map, the attention allocation deviation is calculated as the third type of prediction error, including: The absolute difference between the amplitudes of the physiological signal and the skin conductance signal is calculated as the amplitude deviation between the physiological signal and the user's resting state reference value. The amplitude deviation is then smoothed through a sliding window to obtain the physiological stress deviation intensity as the first type of prediction error. Spatial matching is performed between the actual location sequence in the movement trajectory data and the node sequence of the predetermined path in the environmental spatial structure. The proportion of the number of offset nodes is used as the spatial matching difference. The path execution deviation is quantified by the spatial matching difference as the second type of prediction error. By combining the eye movement features in the eye movement data with the distribution relationship of high attraction regions in the environmental visual attraction distribution map, the positional deviation ratio between the fixation point and the high attraction region is calculated, and the attention allocation deviation is calculated as the third type of prediction error based on the positional deviation ratio.
10. An evaluation system for the psychological recovery effect of community open spaces based on multimodal perception, characterized in that, include: The acquisition module is used to acquire the user's physiological signals, movement trajectory data, eye movement data, and panoramic image data of the target environment space in the target environment space. The parsing module is used to dynamically determine the visible area range based on the user's position and line of sight, parse the panoramic image data, extract the spatial distribution parameters of environmental elements within the visible area, and calculate the information entropy value representing the complexity of the environment based on the spatial distribution parameters. The calculation module is used to calculate the intensity of physiological stress deviation based on the magnitude of the deviation of the physiological signal from the user's resting state, as a first type of prediction error; analyze the degree of matching between the movement trajectory data and the environmental spatial structure, quantify the path execution deviation, as a second type of prediction error; and combine the eye movement characteristics in the eye movement data with the distribution map of environmental visual attraction to calculate the attention allocation deviation, as a third type of prediction error. The construction module is used to input the first type of prediction error, the second type of prediction error, the third type of prediction error and the information entropy value into the pre-constructed evaluation model to output the recovery efficiency index and the multidimensional mitigation index. The multidimensional mitigation index is used to characterize the degree of mitigation of each prediction error. The generation module is used to associate the spatial distribution parameters of the environmental elements with the multidimensional mitigation indicators, extract key environmental intervention thresholds and information entropy recovery effect relationship functions through interpretable machine learning, and generate evaluation results by combining the dynamic recovery effect indicators, multidimensional mitigation indicators and the extracted rules.
Citation Information
Cited By
Unmanned aerial vehicle re-planning method for real-time inspection image analysis
CN121325940A
Campus building and landscape transition-oriented collaborative recovery effect evaluation method
CN121598808A
A Method for Assessing the Synergistic Restoration Effects of Campus Building and Landscape Transition
CN121598808B
