Method for evaluating user experience quality of subway station environment based on machine vision and passenger behavior simulation

Abstract

The subway station environment user experience quality evaluation method based on machine vision and passenger behavior simulation belongs to the field of intelligent transportation and artificial intelligence application. The application constructs a set of subway station user experience quality comprehensive evaluation system covering safety, convenience, comfort and pleasure; through deep convolutional neural network combined with target recognition posterior discriminant algorithm based on pulse timing dependence plasticity, the image data collected in the key area of subway station is extracted; with the help of semantic scene graph model based on brain-like random neural network, the position rationality of safety prompt signs and safety facilities is judged, and is converted into the third level index score in the comprehensive evaluation system; according to the comprehensive evaluation system, the user experience quality level of a single key area is calculated, and combined with the simulation in the loop subway station environment evaluation algorithm, the key area weight and the key facility layout rationality are determined according to the passenger flow, and then the overall environment user experience quality level of the station is obtained.

Description

Technical Field

[0001] This invention belongs to the field of intelligent transportation and artificial intelligence applications. It is an environmental user experience quality evaluation model based on deep learning and brain-like neural networks for feature extraction, supplemented by passenger behavior simulation. It can be deployed and applied to various subway stations and transportation hubs. Background Technology

[0002] With the acceleration of urbanization, the subway system, as a crucial component of urban transportation, plays a vital role in improving public transportation service quality, enhancing passenger loyalty, and promoting sustainable urban transportation development. Optimizing the subway environment primarily relies on a feedback control mechanism of "passenger evaluation - operation department response," making accurate and real-time acquisition of passengers' genuine feelings a necessary prerequisite for environmental optimization. Traditional passenger evaluation methods, mainly through subjective questionnaires, suffer from insufficient real-time accuracy and weak transferability of evaluation results. While methods based on approximate objective indicators offer quantitative analysis, they struggle to comprehensively reflect the actual environmental conditions and passengers' subjective feelings. In recent years, machine learning technology has made significant progress in image recognition and regression modeling, providing new solutions for the real-time evaluation and optimization of the subway environment. Summary of the Invention

[0003] This invention proposes a subway station environment evaluation method based on machine vision and passenger behavior simulation, which has the following characteristics: It constructs a comprehensive evaluation system for the user experience quality of subway stations, encompassing safety, convenience, comfort, and enjoyment; it extracts features from image data collected in key areas of the subway station using a deep convolutional neural network combined with a target recognition posterior discrimination algorithm based on impulse temporal dependence plasticity; it determines the rationality of the location of safety warning signs and safety facilities using a semantic scene graph model based on a brain-like random neural network, and converts these into three-level index scores in the comprehensive evaluation system; based on this, it calculates the user experience quality level of individual key areas according to the comprehensive evaluation system, and combines a simulation-in-the-loop subway station environment evaluation algorithm to determine the weight of each key area and the rationality of the layout of key facilities based on passenger flow, thereby obtaining the overall user experience quality level of the station environment, thus providing decision support for subway station environment optimization and service quality improvement.

[0004] The main contents of this invention are as follows:

[0005] (1) A set of evaluation index system for user experience quality of subway station environment was designed;

[0006] (2) A framework for extracting key environmental elements based on machine vision was designed;

[0007] (3) A quantitative algorithm for evaluating the quality of infrastructure user experience based on passenger behavior simulation was designed.

[0008] The core algorithm of the subway station environment user experience quality evaluation method based on machine vision and passenger behavior simulation is as follows:

[0009] Traditional methods for evaluating the environment of subway stations, particularly those based on subjective questionnaires, have significant limitations. Questionnaires rely on passenger feedback, resulting in poor real-time data collection and an inability to promptly reflect the dynamically changing subway environment. Furthermore, questionnaire results have poor transferability; evaluations of a specific subway station are difficult to apply to others. In addition, questionnaire results largely depend on passengers' subjective feelings, memory accuracy, and the rigor of their responses, thus limiting their objectivity and reliability, potentially leading to discrepancies between the obtained results and actual experiences.

[0010] Evaluation methods based on objective indicators also have certain limitations. These methods mainly rely on the quantitative analysis of observable indicators such as the number of facilities and operational efficiency. While they can reflect the system's operational status relatively intuitively, they often fail to comprehensively and truthfully reflect the level of service quality due to the lack of effective representation of passengers' actual feelings and experiences. For example, although operational efficiency can be calculated by statistically analyzing passenger flow and train punctuality rates, these indicators cannot directly reflect passengers' subjective feelings about crowding levels, environmental comfort, or station aesthetics, making it difficult for the results to reflect passengers' true experience.

[0011] In recent years, with the rapid development of convolutional neural networks with deep feature extraction capabilities, a number of machine vision-based evaluation index systems have been proposed. However, these methods still have limitations in terms of the comprehensiveness of evaluation indicators, spatial comprehensive evaluation capabilities, and the quantification capabilities of certain indicators (such as the appropriateness of the location of safety warning signs). To address these shortcomings, this invention proposes a method for evaluating the user experience quality of subway station environments based on machine vision and passenger behavior simulation (see...). Figure 1 ).

[0012] This invention uses image data collected inside subway stations as input, and achieves accurate analysis of the subway station environment through three steps: constructing an indicator system, image data analysis, and passenger behavior simulation (the specific process is as follows). Figure 2 (As shown). The specific steps are as follows:

[0013] (1) Construct a quality evaluation index system for user experience in subway station environment

[0014] This invention constructs a comprehensive environmental evaluation index system for subway stations through literature review and field research, covering four dimensions: safety, convenience, comfort, and enjoyment. The indicators are shown in Table 1.

[0015] Table 1. Evaluation Index System and Evaluation Method for User Experience Quality in Subway Station Environment

[0016]

[0017] In this embodiment, a weighted average method is used to calculate the total evaluation score in order to quantify the evaluation results. The total score is set at 5 points. Each of the four secondary indicators—safety, convenience, comfort, and enjoyment—has a maximum score of 5 points, with each indicator having a weight of 25%. This equal weighting is because the importance of each evaluation dimension varies across different scenarios. The weighting is primarily used to quantify the total score and does not serve as a constraint on the scenario evaluation, nor does it affect the evaluation results. The calculation formula is as follows:

[0018]

[0019] in, To evaluate the total score, The number of secondary indicators (safety / convenience / comfort / pleasure). For the first Secondary indicators The weight, taking values ​​of , Secondary indicator The rating.

[0020] The scores for secondary indicators are calculated based on the weighted scores of the tertiary indicators, with each tertiary indicator having a maximum score of 5 points. The calculation formula is as follows:

[0021]

[0022] in, Secondary indicator The number of third-level indicators, For the first Secondary indicators Next Three-level indicators The weight, taking values ​​of , Level 3 indicators The rating.

[0023] (2) Construct a visual feature extraction model based on a target recognition posterior discrimination algorithm using a deep convolutional neural network and pulse temporal dependent plasticity.

[0024] This invention proposes a multi-network fusion extraction framework for key elements of subway station environments based on machine vision (see...). Figure 3This invention enables target identification, instance segmentation, location determination, and target classification of key elements within subway stations. For target identification, an improved target detection model based on YOLOv11 and an impulse-time-dependent plasticity target recognition posterior discrimination algorithm is designed, achieving efficient detection of 61 categories of key elements within subway stations, including safety facilities, directional signs, and convenience facilities. For instance segmentation, an improved U-Net algorithm is used for image segmentation, accurately extracting key areas such as the ground, walls, and subway platform screen doors. For location determination, a semantic scene graph model based on a brain-like random neural network is proposed to determine the reasonableness of the location of safety signs and safety facilities. Furthermore, for open-world elements in subway stations, such as decorations, which are difficult to accurately identify and classify, a target classification model based on a convolutional neural network, MEP-Net, is proposed to improve the accuracy of evaluating the pleasantness of subway stations. The specific algorithms for target identification, instance segmentation, location determination, and target classification of key elements are described below:

[0025] The target recognition model is mainly used for automatic identification of key facilities and environmental elements in subway station images. The model uses YOLOv11 as the basic detection framework, consisting of a backbone network, a feature fusion network, and a prediction head. The input is an image captured in a subway station, and the output is the category, confidence score, and location parameters of the candidate target. Due to the complex environment of subway stations, including strong reflections, occlusions, and background interference, the detection rate of some small targets is low during target recognition. Therefore, this invention proposes a posterior target recognition algorithm based on impulse temporal-dependent plasticity to improve the detection rate (see...). Figure 4 ).

[0026] For candidate targets Let the confidence level of the target be . ,correspond The classification probability vectors for each category are:

[0027]

[0028] Then the candidate target belongs to the first The confidence level of a target class is defined as:

[0029]

[0030] in, Indicate candidate target Belongs to the Initial confidence level of the class The initial confidence scores of all candidate categories are used as input to the posterior discrimination module.

[0031] Since the low-confidence candidate targets are mostly objects such as safety signs in subway stations, we further extract their color saturation, average brightness, rectangularity, and positional relationship features relative to the reference target. Let the candidate targets be... The area is The area is Then its color saturation characteristics Defined as:

[0032]

[0033] in, Represents pixels Saturation components in the HSV color space. Candidate targets. Average brightness characteristics Defined as:

[0034]

[0035] in, Represents pixels The luminance component in the HSV color space.

[0036] Let the width and height of the minimum bounding rectangle of the candidate target be respectively... and Then its rectangularity Defined as:

[0037]

[0038] in, Let be a very small positive number, and take it as . This is used to avoid the denominator being zero.

[0039] Furthermore, the position determination algorithm described later is invoked to select the reference target category with the highest salience to the current candidate target, and the seven positional relationship salience values ​​of the candidate target relative to the reference target are calculated to form a relationship vector:

[0040]

[0041] in, , , , , , , These represent the salience values ​​for left, right, top, bottom, inside, outside, and nearby relationships, respectively. Based on these, candidate targets are constructed. The comprehensive input feature vector:

[0042]

[0043] The above-mentioned comprehensive feature vectors are encoded into input pulse sequences and then fed into a posterior discriminant network based on pulse temporal dependence plasticity learning.

[0044] The output layer of the posterior discriminant network is set to be related to the candidate category. There are 61 neurons of the same category, and these neurons compete with each other through lateral inhibition. Let the first... Each category of neuron at time [time] The membrane potential is Then its update method is:

[0045]

[0046] in, Indicates the first Each category of neuron at time [time] The membrane potential value was initially set to 0. This indicates the membrane potential value at the next moment; The membrane potential attenuation coefficient is set to... It is used to characterize the degree to which a neuron retains its state from the previous time step; The dimension of the input features; For the first The input neuron to the first The synaptic weights of neurons in each category are initially set to... ; For the first The input neurons at time 1 The input pulse signal; To output the total number of categories, set to ; For the first The class of neurons for the first... The transverse inhibition coefficients of neurons in each class are set as follows: ; For the first Each output neuron at time... The output pulse signal.

[0047] During the training phase, the first The membrane potential of each class of neuron is jointly determined by historical membrane potential retention terms, current input feature activation terms, and competitive inhibition terms from other class of neurons. Based on the true class labels of the training samples, the synaptic weights are updated using a reward-modulated pulse temporal-dependent plasticity rule, through continuous adjustment. This allows for dynamic adjustment of output pulses of different categories. The category corresponding to the output layer neuron that emits the first pulse is the final output category of the target. Let the time difference between the input pulse and the output pulse be... Then when When greater than 0:

[0048]

[0049] when When less than 0:

[0050]

[0051] in, and These are positive and negative magnitude updates, respectively. and These represent the positive and negative time constants, respectively. After training, the model saves the network structure and updated synaptic weights for posterior discrimination of low-confidence candidate targets in the subsequent inference stage. The final output category... for:

[0052]

[0053] Therefore, by leveraging the impulse temporal-dependent plasticity posterior discrimination module, and comprehensively utilizing the multi-class confidence, color and shape features of candidate targets, as well as their seven types of positional relationships relative to the reference target, we can further discriminate easily confused low-confidence targets in complex subway station environments, thereby improving the accuracy and robustness of target recognition.

[0054] The instance segmentation model can output pixel-level mask information for detected targets, enabling fine depiction of target contours. This segmentation model is used to extract walkable ground areas, wall areas, and platform screen door areas within subway stations, for subsequent congestion calculations, target location determination on walls and platform screen doors, and enhanced recognition.

[0055] In terms of network structure, this invention adopts a U-Net symmetrical encoder-decoder structure, with skip connections transmitting multi-scale features. The encoder extracts deep semantic information step-by-step through convolution and pooling, while the decoder restores spatial resolution through upsampling and feature fusion, ultimately outputting a ground region probability map of the same size as the original image. The formula for the output layer is:

[0056]

[0057] In the formula, The original output value of the model's output layer before the activation function is applied is called logit, which represents the unnormalized discrimination score of the target class at the current position. This value is generated by the convolution operation and then mapped by the Sigmoid function to obtain the corresponding probability value.

[0058] As mentioned above, the segmentation model is used for crowding calculation, and the enhanced recognition of targets on walls and subway platform screen doors is detailed in the target recognition model section. For location determination, please refer to the subsequent location determination algorithm. The crowding calculation process is as follows:

[0059] First, the pixel area of ​​the ground region is calculated using the segmentation model described above with the aid of thresholding. :

[0060]

[0061] In the formula, and These are the width and height of the image, respectively. Pixel position The predicted probability of belonging to the target region; Indicates an indicator function, when When the value is 1, it means that the pixel is identified as a pixel within the target area; otherwise, it is 0, meaning that the pixel is not identified as a pixel within the target area.

[0062] After the model outputs a pixel-level walkable area mask, the actual ground area can be obtained by combining it with the camera's field of view parameters. Assuming the environment image was taken with the camera horizontally positioned above the ground, and the camera's horizontal field of view (HFOV) is... Vertical field of view (VFOV) is The height of the camera above the ground is Then the camera's vertical coverage area for:

[0063]

[0064] Vertical coverage for:

[0065]

[0066] Based on the number of pixels in the ground area Total number of pixels in the image The actual ground area can be obtained by using the ratio. for:

[0067]

[0068] The aforementioned object detection model was then used to identify the number of passengers. In obtaining the ground area and passenger numbers Afterwards, the average area occupied by each passenger is:

[0069]

[0070] Finally, the congestion level can be determined from Table 2.

[0071] Table 2. Pedestrian Crowding Experience Quality Rating

[0072]

[0073] The location determination algorithm automatically evaluates the rationality of the placement of various prompts and signs by leveraging the spatial positional relationships of elements. To determine rationality, building upon the aforementioned target recognition and instance segmentation algorithms for element categories and regions, we construct a semantic scene graph model based on a brain-like random neural network. This model describes the positional relationship of target elements relative to reference elements and includes seven brain-like random neurons. These represent seven typical positional semantic relationships: left, right, above, below, inside, outside, and nearby (see...). Figure 5 Since the target functions corresponding to different positional relationships differ in numerical range, to avoid interference with the neuronal membrane potential update and competitive determination process, this invention divides neurons into three groups. Neurons within a group can adjust their membrane potential (voltage value) by exchanging reward and inhibition pulse signals based on external feedback information. Then, the relative relationship between targets is determined based on the positional relationship of the neuron with the highest membrane potential value. Specifically, the target recognition or instance segmentation algorithm first outputs the contour region of the target element. and the outline area of ​​the reference element The outline region of the target element is denoted as:

[0074]

[0075] The reference element outline region is denoted as:

[0076]

[0077] Subsequently, in the outline area of ​​the target element respectively and reference element outline area Each sample is taken according to a uniform random sampling strategy. Each of the points is used as a target point. and reference points The aforementioned first The coordinates of the target point and the reference point are denoted as follows:

[0078]

[0079] This invention then defines objective functions for seven positional relationships to calculate reward and suppression pulse values. For the four directional relationships (left, right, top, and bottom), this invention is based on the target point. With reference point The relative directions are used to construct the orientation consistency coefficient. A larger orientation consistency coefficient indicates that the target point is closer to the corresponding ideal direction relative to the reference point. The consistency coefficients for the left, right, top, and bottom are defined as follows:

[0080]

[0081]

[0082]

[0083]

[0084] The non-zero Euclidean distance between the target point and the reference point for:

[0085]

[0086] in, It is a very small positive number, and in this invention it is set to... This is used to avoid the denominator being zero when two points coincide. Based on the above consistency coefficient, the objective functions for the left-side, right-side, upper-side, and lower-side relationships are defined as follows:

[0087]

[0088]

[0089]

[0090]

[0091] in, , , and The positive constant bias term is used to ensure that each objective function is always greater than 1, thereby affecting the reward function. The value is always less than 1 to ensure the stability of the neural network learning; in this invention, it is always set to 0.01. When the target point is closer to the ideal left direction of the reference point, The closer to 1, the better. The closer to the minimum value Similarly, the closer the target point is to the ideal right, top, and bottom directions of the reference point, , , The closer each value is to 1, the better the corresponding objective function. , , The closer to the minimum value When the target point deviates from the corresponding direction or even lies in the opposite direction, the consistency coefficient decreases, and the objective function increases accordingly. Therefore, the above four objective functions are used to characterize the significance of the relationship between the target element and the reference element to the left, right, above, and below, respectively.

[0092] Regarding internal and external relationships, this invention models the relationship based on the bidirectional sampling inclusion relationship between the target element contour region and the reference element contour region. This is because the contours of both the target and reference elements are typically irregular, and the complex environment of subway stations can introduce false detections and biases into target recognition and instance segmentation results. Therefore, relying solely on the area or center position of the bounding box is insufficient to accurately reflect the true occlusion relationship between the two. In contrast, by randomly sampling within both the target and reference element contour regions and statistically analyzing the proportion of each sampling point falling within the other's region, the enclosing relationship and hierarchical relationship between the two under two-dimensional image projection can be more effectively characterized. The mathematical expression is as follows:

[0093]

[0094]

[0095] in, A function to determine the inclusion of the target point relative to the reference region. The region containing the reference point relative to the target region is determined by a decision function. When the... Target sampling points Falling into the outline area of ​​the reference element Inside, The value is 1 if it is not 1, otherwise the value is 0; when the first Reference sampling points Fall into the outline area of ​​the target element Inside, The value is 1 if it is not 1, otherwise the value is 0.

[0096] To more precisely characterize cases where the target and reference overlap, this invention further introduces a penalty coefficient. The present invention sets Based on this, the first is defined. Consistency coefficient of internal relationship corresponding to each sampling point Consistency coefficient with external relations They are respectively:

[0097]

[0098]

[0099] in, Used to characterize the The significance of the internal relationships corresponding to each sampling point Used to characterize the The significance of external relationships corresponding to each sampling point. This is determined when the target sampling point is located inside the contour region of the reference element and the reference sampling point is not located inside the contour region of the target element. The maximum value is reached; when both fall within each other's region, a penalty term for mutual inclusion is introduced. This will decrease accordingly, thus reducing the salience of the internal relationship. Correspondingly, when the target sampling point is located outside the reference element contour region and the reference sampling point does not fall inside the target element contour region, A larger value indicates a more significant external relationship.

[0100] Based on the aforementioned consistency coefficients, we define the objective functions for internal relations respectively. Objective function of external relations for:

[0101]

[0102]

[0103] in, and This is a positive constant bias term used to ensure that the objective function is always greater than 0. This causes the subsequent reward function to always be less than To ensure the stability of neural network learning, this is all set to 0.01 in this invention. The more significant the internal relationships, the better. The closer objective function The closer to the minimum value The more significant the external relationship, The closer objective function The closer to the minimum value .

[0104] Regarding proximity relationships, this invention models them based on the Euclidean distance between the target point and the reference point. Specifically, it first models the Euclidean distance between the target point and the reference point as defined above. Normalization is performed to improve comparability under different image scale conditions. Let the current image width be... The height is Then the normalized distance Defined as:

[0105]

[0106] Further define the consistency coefficient of proximity relations for:

[0107]

[0108] in, This is the distance attenuation coefficient, used to adjust the sensitivity of proximity relationships to distance changes. In this invention, Set to 3. The closer the target element is to the reference element, the better. The smaller, The closer to 1; as the distance between the two increases, Increase Gradually decrease.

[0109] Based on this, the objective function for the proximity relation is defined as follows:

[0110]

[0111] in, This is a positive constant bias term used to ensure that the objective function is always greater than 0. This causes the subsequent reward function to always be less than To ensure the stability of neural network learning, this value is set to 0.01 in this invention. Therefore, the more significant the proximity relationship between the target element and the reference element, the better. The closer to 1, the better the objective function. The closer to the minimum value When the two are far apart, Decrease, objective function An increase indicates that the relationship between neighbors is gradually weakening.

[0112] To satisfy the reward-driven mechanism of brain-like stochastic neural networks, the above-mentioned objective functions are further transformed into reward functions. For any positional relationship... Its reward function is uniformly defined as:

[0113]

[0114] in, These correspond to left, right, top, inside, outside, and nearby relationships, respectively. Specifically, this refers to the... Round sampling, the objective function and reward function values ​​are respectively represented as Since all the above objective functions satisfy:

[0115]

[0116] Therefore, the corresponding reward function always satisfies:

[0117]

[0118] The smaller the objective function, the more closely the current sampling point pair conforms to the corresponding positional relationship, and the larger the reward function; conversely, the larger the objective function, the less significant the positional relationship, and the smaller the reward function. Each reward value serves as external feedback information input to the corresponding neuron. Neurons dynamically update the membrane potential value based on reward and inhibition pulses. Finally, the positional relationship corresponding to the neuron with the highest membrane potential value is used as the final positional determination result between the target element and the reference element. Since the objective functions corresponding to different positional relationships differ in numerical range, to avoid interference with the neuronal membrane potential update and competitive determination process, this invention divides the seven positional relationships into three groups of fully connected neuron clusters. Neurons within each group compete with each other. The left, right, upper, and lower neuron groups form a group. Internal and external neurons are combined into a group. Nearby neurons are grouped separately. The membrane potential update process of neurons is as follows:

[0119] 1. For each neuron, targeting the first... Round sampling, using an objective function to calculate the reward value. ;

[0120] 2. Calculate the current reward / penalty threshold for the current neuron group. For clusters of neurons and , For clusters of neurons , ,in The learning rate is set to 0.7. This serves as a reference value for the centroid distance between this type of target element and this type of reference element;

[0121] 3. Comparison and the average judgment threshold before the update Update neuron weights:

[0122]

[0123] 4. Update the total emission rate of signals (including reward and punishment signals) for each neuron:

[0124]

[0125] 5. Normalize the weights of each neuron:

[0126] 6. Calculate the membrane potential of each neuron:

[0127]

[0128] For wall decorations and three-dimensional decorations within the "Station Decoration" index of the "pleasure" dimension in the subway station environment evaluation system, traditional target detection methods struggle to accurately identify these objects due to their open-world nature, characterized by significant morphological differences, irregular structures, and blurred category boundaries. Therefore, this invention proposes a target classification and scoring model based on convolutional neural networks, MEP-Net (e.g., ...). Figure 6 This is used to automatically classify wall decorations and three-dimensional decorations, and to quantify the pleasure index.

[0129] First, fixed-size image patches are extracted from high-attention areas for passengers, such as waiting areas, platform walls, and transfer passages, and then input into the MEP-Net scoring model for feature extraction. The network extracts color, texture, and spatial structure features from the decorative areas through a multi-layer convolutional structure, while enhancing key visual information and suppressing background noise. Subsequently, the extracted high-dimensional features are mapped through a fully connected classification layer at the end of the model, outputting the target category probability. Based on the maximum probability principle, the category is determined, classifying the decorative elements into two categories: wall decoration and three-dimensional decoration.

[0130] MEP-Net takes the target classification results and their spatial information as input, and uses a multi-layer convolutional network to fuse and weight features of different categories. The score reflects the degree of influence of the corresponding environmental elements on the passenger experience, and can be further summarized to obtain the overall environmental evaluation results of the subway station.

[0131] The input to this scoring model is decorative images of subway stations and online sources, labeled with subjective scores (1-5 points) from domain experts. The model consists of multi-layered feature extraction and classification modules, including: five 3×3 convolutional layers; five max-pooling layers; one rescaling layer for data normalization; two dropout layers; one fully connected layer; one dense layer; and one output layer. Detailed parameter configurations for each layer of the MEP-Net model are shown in Table 3. The model achieved an accuracy of over 96% in field testing.

[0132] Table 3. MEP-Net Network Structure

[0133]

[0134] (3) Construct a quantitative algorithm for evaluating the quality of infrastructure user experience based on passenger behavior simulation.

[0135] To calculate the overall user experience quality evaluation score of subway stations and address the challenge of directly quantifying certain infrastructure indicators under different spatial locations and passenger flow conditions, this invention proposes a user experience quality evaluation and quantification algorithm based on passenger behavior simulation. This algorithm simulates and analyzes passenger movement paths, spatial access frequency, and main movement directions within subway stations, quantifying the impact of relevant infrastructure on passengers as time-related indicators. It then determines key sampling points for user experience quality and their corresponding weights, and calculates the overall user experience quality level of the subway station environment based on this analysis.

[0136] Specifically, firstly, the space area of ​​the subway station Divided according to a spatial scale of 2 meters × 2 meters A grid cell, denoted as:

[0137]

[0138] in, Indicates the first One grid cell, This represents the total number of grid cells obtained from the spatial division of the subway station. Subsequently, based on the SUMO pedestrian simulation model, the travel paths of passengers from starting areas such as platforms, concourses, and transfer passages to entrances / exits, stairs, escalators, and other key facilities are set, and the number of times passengers pass through each grid cell and their main movement direction within that grid cell are recorded during the simulation.

[0139] For any grid cell The number of visits is defined as:

[0140]

[0141] in, This represents the total number of simulated passengers; To access the decision function, when the first... Passengers passed through the grid unit The value is 1 if the condition is met, and 0 otherwise. (Number of visits) The larger the value, the more frequently passengers pass by or pay attention to that grid cell, and the greater its impact on the overall user experience quality evaluation.

[0142] Furthermore, statistics on passengers in grid units The dominant orientation of the mesh cell is obtained by determining the main direction of motion within it. The dominant orientation is used to determine the acquisition direction of the environmental images at that location, ensuring that the acquired images reflect passengers' visual perception of the station environment during actual passage as accurately as possible. Therefore, the number of grid visits can be used to determine the optimal orientation. Dominant Orientation In addition to the spatial functional area where the grid is located, a set of key sampling points for user experience quality was obtained by filtering. , denoted as:

[0143]

[0144] in, Indicates the first Key points for sampling user experience quality. This indicates the number of key points selected. In one embodiment of the invention, grid cells with high access frequency, located on main passenger routes, and close to stairs, escalators, waiting areas, transfer passages, station halls, platforms, entrances / exits, and critical infrastructure can be preferentially selected as sampling key points (see [link to invention]). Figure 9 ).

[0145] For each sampling key point The environmental image is acquired based on the dominant orientation of its corresponding grid cell, and the user experience quality score of the key point is calculated using the aforementioned machine vision evaluation model, denoted as:

[0146]

[0147] in, Indicates the first The user experience quality score for each key point ranges from 0 to 5. This represents the scoring function obtained based on a comprehensive evaluation system and a machine vision model.

[0148] To reflect the varying contributions of different key points to the overall station environment evaluation results, this invention further calculates the weight of each key point based on the number of visits obtained from passenger behavior simulation. Let key points be... The number of visits to the corresponding grid cell is Then its weight is defined as:

[0149]

[0150] in, Indicates the first The weight of each key point in the overall evaluation. The more frequently a key point is visited, the greater its corresponding weight, indicating that the location has a more significant impact on the overall quality of the passenger experience.

[0151] Finally, the subway station Overall environment user experience quality evaluation score The weighted sum of the scores for each key point is obtained as follows:

[0152]

[0153] Therefore, this invention can combine passenger behavior simulation results with machine vision evaluation results, so that the overall evaluation results not only reflect the environmental quality level of each key area, but also reflect the impact of passenger flow distribution, travel path and spatial attention on user experience quality, thereby improving the authenticity and practicality of subway station environmental evaluation results.

[0154] Because the user experience quality of some infrastructure indicators is difficult to quantify directly using only image recognition results—such as exit signs, transfer guidance signs, medical facilities, and automated external defibrillators (AEDs)—the evaluation results depend not only on the existence of the facilities but also on their visibility, accessibility, guidance effectiveness, and service coverage during actual passenger passage. Therefore, this invention introduces a quantitative evaluation method based on passenger behavior simulation. This method combines machine vision recognition results with passenger behavior simulation results, transforming the aforementioned infrastructure indicators that are difficult to quantify directly into time-related evaluation indicators, and then using these indicators for quantitative scoring.

[0155] Specifically, firstly, the aforementioned machine vision recognition algorithm is used to detect targets such as exit signs, transfer guidance signs, medical facilities, and automated external defibrillators (AEDs) in subway station images, obtaining a set of spatial locations for these facilities. Then, combining the actual spatial structure and two-dimensional planar layout data of the subway station, areas such as the concourse, platforms, stairs, escalators, transfer passages, entrances / exits, and large open spaces are converted into a SUMO network map. In this SUMO network map, walkable areas within the station are represented as pedestrian movable spaces, spatial connections such as stairs, escalators, transfer passages, and entrances / exits are represented as pedestrian paths, and the locations of infrastructure such as exit signs, transfer guidance signs, medical facilities, and AEDs are mapped to corresponding simulation spatial coordinates. Thus, a simulation environment that reflects the real spatial structure, facility layout, and passenger traffic relationships of the subway station is constructed.

[0156] Based on this, passenger travel demands are generated under different origins, destinations, and passenger flow intensities according to historical passenger flow card swiping data at the station. The passenger travel demands include information such as the passenger's origin, target exit or transfer platform, departure time, and travel route, thereby ensuring that the simulation process can reflect the spatial distribution characteristics and traffic patterns of passengers in the real station.

[0157] For wayfinding facilities such as exit signs and transfer guidance signs, this invention inputs the spatial location of the facilities, station network map, passenger origin-destination distribution, and passenger flow density into a passenger behavior simulation model. This simulation model models the actual passage of passengers within the station and outputs behavioral indicators related to the layout of the wayfinding facilities. These indicators include passenger hesitation, detours, incorrect direction selection, transfer travel time, and total travel time to the target exit or transfer facility near intersections, large open spaces, transfer nodes, and entrances / exits. Therefore, the simulation output results can be used to determine whether the location of the wayfinding facilities effectively supports passengers in making directional judgments and route selections.

[0158] Let the first The increase in travel time for each passenger during the simulation due to the layout of guidance facilities is: The weighted average travel time increment corresponding to the guidance facility is expressed as:

[0159]

[0160] in, This indicates the degree to which the layout of guidance facilities affects the overall efficiency of passenger traffic. The smaller the value, the more reasonable the layout of exit signs and transfer guidance signs, and the lower the possibility that passengers will hesitate, detour, or make the wrong choice of direction in the station. This indicates the number of simulated passengers participating in the evaluation.

[0161] Furthermore, the weighted average travel time increment corresponding to the guidance facilities is mapped to the guidance facility evaluation score:

[0162]

[0163] in, Indicates the evaluation score for guidance facilities; This represents the upper limit of the evaluation score, which is 5 in one embodiment of the present invention; This indicates the weighted average travel time increment caused by the layout of guidance facilities; This represents the acceptable threshold for incremental travel time. When When smaller, near ;when When the value increases, it indicates that the layout of the guidance facilities has a significant adverse impact on passenger traffic efficiency, and the evaluation score will decrease accordingly.

[0164] For emergency service facilities such as medical aid facilities and automated external defibrillators (AEDs), this invention inputs the spatial location of medical facilities, station network maps, passenger distribution, regional passenger flow density, and the location of the sudden medical event into a passenger behavior simulation model. This passenger behavior simulation model simulates the process of obtaining the nearest medical aid facility or AED from the event location and delivering it to the event location after a sudden medical event occurs, and outputs the time it takes to obtain the medical aid facility, the travel distance, and the actual response time affected by congestion. Therefore, the service coverage and emergency response efficiency of medical aid facilities and AEDs can be evaluated based on the simulation output results.

[0165] Let the first In a simulated emergency medical event scenario, the response time required to obtain medical aid facilities or AEDs and deliver them to the event location is: The weighted average response time for medical rescue facilities is expressed as:

[0166]

[0167] in, This represents the weighted average response time of medical rescue facilities or AEDs. The smaller the value, the more reasonable the spatial layout of medical rescue facilities and AEDs, and the stronger the emergency service coverage. Indicates the first The weight of each emergency medical event scenario in the evaluation is set in this embodiment as follows: .

[0168] Furthermore, the weighted average response time corresponding to medical rescue facilities is mapped to a medical facility evaluation score:

[0169]

[0170] in, This indicates the evaluation score for medical rescue facilities; This represents the upper limit of the evaluation score, which is 5 in one embodiment of the present invention; This represents the weighted average response time corresponding to medical emergency facilities; This represents the security response time threshold. In one embodiment of the invention, It can be set to 90 seconds. This is useful when the time for obtaining and delivering medical aid or an AED is short. near When the response time exceeds the safe response time threshold, the evaluation score drops rapidly, reflecting the emergency service risks caused by insufficient medical rescue facility layout.

[0171] Through the above methods, this invention can not only determine the existence of infrastructure such as exit signs, transfer guidance signs, medical rescue facilities and AEDs, but also further analyze their guidance efficiency, visibility, service coverage and emergency response efficiency in real operation scenarios by combining passenger behavior simulation results, so as to make the evaluation results of the user experience quality of subway station environment more scientific, detailed and practical.

[0172] Invention Effects

[0173] This invention proposes a subway station environment assessment method based on machine vision and passenger behavior simulation. This method successfully solves the problems of strong subjectivity, poor real-time performance, and difficulty in reflecting actual passenger experiences inherent in traditional assessment methods. It opens a new direction for intelligent assessment in the urban rail transit field, achieving precise quantitative analysis of the subway station environment and providing a reference for subway station environmental optimization and management. Specifically, the proposed model achieves 97.5% mAP0.5 and 92.5% mAP0.5:0.95 across all detection categories, accurately identifying key objects affecting the subway station environment, such as safety signs, facilities, and directional signs. In a case study at Jintai Road Station, the error between the model and actual user evaluations was only 8.13%, verifying the model's accuracy and reliability. Attached Figure Description

[0174] Figure 1 Comparison of different evaluation methods

[0175] Figure 2 Flowchart of a method for evaluating the user experience quality of subway station environments based on machine vision and passenger behavior simulation

[0176] Figure 3 Framework diagram for the multi-network fusion extraction of key elements of subway station environment

[0177] Figure 4 A posterior recognition algorithm for target recognition based on pulse temporal dependent plasticity (see diagram)

[0178] Figure 5 Location determination algorithm based on neuromorphic spiking neural network (Figure)

[0179] Figure 6 MEP-NET Structure Diagram for Decorative Item Evaluation Model

[0180] Figure 7 MEP-NET Confusion Matrix for Decorative Evaluation Model

[0181] Figure 8 Key locations of subway stations

[0182] Figure 9 Simulation evaluation graph based on SUMO Detailed Implementation

[0183] In this embodiment, the subway station environmental assessment model can be deployed at any subway station to automatically analyze and comprehensively evaluate the environmental images of key areas of the station. Its main workflow is shown in the figure.

[0184] First, environmental images are captured using mobile vision sensors at key locations within the subway station, including staircases, escalators, waiting areas, transfer passages, concourses, platforms, and entrances / exits. These mobile vision sensors are equipped with high-definition wide-angle lenses and can be handheld for shooting. The lens height is approximately 1.5 meters to simulate the eye level of an adult passenger; the shooting angle is controlled within ±10° in both the horizontal and vertical directions. To ensure image acquisition quality, extreme lighting and severe occlusion are avoided as much as possible during the acquisition process. For unavoidable passenger facial information, a privacy occlusion system based on the PaddleDetection framework is used for processing. This embodiment collects 20 representative key point images to ensure the comprehensiveness and accuracy of the evaluation results.

[0185] Subsequently, the preprocessed images are input into a subway station environmental assessment model deployed in a computing device. This computing device can be a server, workstation, or high-performance personal computer with image processing and model inference capabilities. The model automatically identifies, extracts features from, and comprehensively analyzes environmental elements in the input images, and outputs evaluation results corresponding to each key point image.

[0186] Furthermore, by summarizing and calculating the evaluation results of multiple key point images, the overall environmental evaluation result of the subway station can be obtained. This overall evaluation result reflects the station's comprehensive level in dimensions such as safety, convenience, comfort, and pleasantness, and can serve as a basis for subsequent management and optimization decisions. For key areas with lower evaluation results, managers can combine the model output results to take further targeted improvement measures, such as optimizing lighting conditions, improving environmental cleanliness, improving wayfinding facilities, improving spatial organization, and enhancing the configuration level of service facilities.

[0187] Furthermore, the environmental assessment process can be executed periodically, enabling continuous monitoring of the subway station's environmental condition. Managers can promptly identify weaknesses in the station environment based on changes in different time periods, areas, and assessment results, and make dynamic adjustments and optimizations in accordance with station operational needs. Through this approach, long-term, continuous, and automated evaluation of the subway station environment can be achieved, providing technical support for improving station environmental quality.

Claims

1. A method for evaluating the user experience quality of subway station environments based on machine vision and passenger behavior simulation, characterized in that... Includes the following steps: Using image data collected inside subway stations as input, the system achieves accurate analysis of the subway station environment through three steps: constructing an indicator system, image data analysis, and passenger behavior simulation. The specific steps are as follows: (1) Construct a quality evaluation index system for user experience in subway station environment A comprehensive evaluation index system for the subway station environment was constructed, covering four dimensions: safety, convenience, comfort, and enjoyment. The evaluation results are quantified, and the total evaluation score is calculated using a weighted average method. The total score is set at 5 points; among them, the four secondary indicators of safety, convenience, comfort, and pleasure are each worth 5 points, and the weight of each indicator is set at 25%; the calculation formula is as follows: ； in, To evaluate the total score, These are the secondary indicators, namely the number of safety, convenience, comfort, and pleasure aspects. For the first Secondary indicators The weight, taking values ​​of , Secondary indicator The rating; The scores for the secondary indicators are calculated based on the weighted scores of the tertiary indicators, with each tertiary indicator having a maximum score of 5 points; the calculation formula is as follows: ； in, Secondary indicator The number of third-level indicators, For the first Secondary indicators Next Three-level indicators The weight, taking values ​​of , Level 3 indicators The rating.

2. The method according to claim 1, characterized in that, The indicators are shown below: 。 3. The method according to claim 1, characterized in that, The visual feature extraction model based on a deep convolutional neural network and a pulse temporal-dependent plasticity-based posterior discriminant algorithm for target recognition is constructed as follows: The target recognition model is used to automatically identify key facilities and environmental elements in subway station images. The model adopts YOLOv11 as the basic detection framework and consists of a backbone network, a feature fusion network, and a prediction head. The input is the image collected in the subway station, and the output is the category, confidence, and location parameters of the candidate target. A target recognition posterior recognition algorithm based on impulse temporal dependent plasticity is proposed to improve the detection rate. For candidate targets Let the confidence level of the target's existence be . ,correspond The classification probability vectors for each category are: ； Then the candidate target belongs to the first The confidence level of a target class is defined as: ； in, Indicate candidate target Belongs to the Initial confidence level of the class The initial confidence scores of all candidate categories are used as inputs to the posterior discrimination module. Extract its color saturation, average brightness, rectangularity, and positional relationship features relative to the reference target; Let candidate targets... The area is The area is Then its color saturation characteristics Defined as: ； in, Represents pixels Saturation components in the HSV color space; candidate targets Average brightness characteristics Defined as: ； in, Represents pixels The luminance component in the HSV color space; Let the width and height of the minimum bounding rectangle of the candidate target be respectively... and Then its rectangularity Defined as: ； in, Let be a very small positive number, and take it as . This is used to avoid the denominator being zero; Furthermore, the position determination algorithm described later is invoked to select the reference target category with the highest salience to the current candidate target, and the seven positional relationship salience values ​​of the candidate target relative to the reference target are calculated to form a relationship vector: ； in, , , , , , , These represent the significance values ​​of relationships to the left, right, top, bottom, inside, outside, and nearby, respectively; based on these, candidate targets are constructed. The comprehensive input feature vector: ； The above-mentioned comprehensive feature vectors are encoded into input pulse sequences and then input into a posterior discriminant network based on pulse temporal dependence plasticity learning. The output layer of the posterior discriminant network is set to be related to the candidate category. Neurons of the same category; let the first... Each category of neuron at time [time] The membrane potential is Then its update method is: ； in, Indicates the first Each category of neuron at time [time] The membrane potential value was initially set to 0. This indicates the membrane potential value at the next moment; The membrane potential attenuation coefficient is set to... It is used to characterize the degree to which a neuron retains its state from the previous time step; The dimension of the input features; For the first The input neuron to the first The synaptic weights of neurons in each category are initially set to... ; For the first The input neurons at time 1 The input pulse signal; To output the total number of categories; For the first The class of neurons for the first... The transverse inhibition coefficients of neurons in each class are set as follows: ; For the first Each output neuron at time... The output pulse signal; During the training phase, the first The membrane potential of each class of neuron is jointly determined by historical membrane potential retention terms, current input feature activation terms, and competitive inhibition terms from other class of neurons. Based on the true class labels of the training samples, the synaptic weights are updated using a reward-modulated pulse-time-dependent plasticity rule, through continuous adjustment. This allows for dynamic adjustment of output pulses of different categories; the category corresponding to the output layer neuron that emits the first pulse is the final output category of the target; let the time difference between the input pulse and the output pulse be... Then when When greater than 0: ； when When less than 0: ； in, and These are positive and negative magnitude updates, respectively. and These represent the positive and negative time constants, respectively; after training, the model saves the network structure and updated synaptic weights for posterior discrimination of low-confidence candidate targets in the subsequent inference stage; the final output category... for: ； Therefore, by leveraging the impulse temporal-dependent plasticity posterior discrimination module, and comprehensively utilizing the multi-class confidence, color and shape features of candidate targets, as well as their seven types of positional relationships relative to the reference target, we can further discriminate easily confused low-confidence targets in complex subway station environments, thereby improving the accuracy and robustness of target recognition. The instance segmentation model can output pixel-level mask information for detected targets, achieving fine characterization of target contours. This segmentation model is used to extract walkable ground areas, wall areas, and subway platform screen door areas within subway stations, which are then used for subsequent congestion calculations, target location determination on walls and subway platform screen doors, and enhanced recognition. In terms of network structure, a U-Net symmetrical encoder-decoder structure is adopted, with skip connections in between to pass multi-scale features. The encoder extracts deep semantic information step-by-step through convolution and pooling, while the decoder restores spatial resolution through upsampling and feature fusion, ultimately outputting a ground region probability map of the same size as the original image. The formula for the output layer is: ； In the formula, The original output value of the model output layer before the activation function is applied, i.e., logit, is used to represent the unnormalized discrimination score of the current position belonging to the target class. This value is generated by the convolution operation and obtained by mapping through the Sigmoid function to obtain the corresponding probability value. As mentioned above, the segmentation model is used for crowding calculation, and the enhanced recognition of targets on walls and subway platform screen doors is detailed in the target recognition model section. For location determination, please refer to the subsequent location determination algorithm. The crowding calculation process is as follows: First, the pixel area of ​​the ground region is calculated using the segmentation model described above with the aid of thresholding. : ； In the formula, and These are the width and height of the image, respectively. Pixel position The predicted probability of belonging to the target region; Indicates an indicator function, when When the value is 1, it means that the pixel is identified as a pixel in the target area; otherwise, it is 0, which means that the pixel is not identified as a pixel in the target area. After the model outputs a pixel-level walkable region mask, the actual ground area can be obtained by combining it with the camera's field of view parameters; assuming the environment image is taken by a camera horizontally positioned above the ground, the camera's horizontal field of view (HFOV) is... Vertical field of view (VFOV) is The height of the camera above the ground is Then the camera's vertical coverage area for: ； Vertical coverage for: ； Based on the number of pixels in the ground area Total number of pixels in the image The actual ground area can be obtained by using the ratio. for: ； The aforementioned object detection model was then used to identify the number of passengers. In obtaining the ground area and passenger numbers Afterwards, the average area occupied by each passenger is: ； Depend on Indicators for assessing congestion levels; A semantic scene graph model based on a brain-like random neural network is constructed; the model is used to describe the positional relationship between target elements and reference elements, and includes 7 brain-like random neurons. These represent seven typical positional semantic relationships: left, right, above, below, inside, outside, and nearby. Since the target functions corresponding to different positional relationships differ in numerical range, to avoid interference with neuronal membrane potential updates and competitive judgment processes, neurons are divided into three groups. Neurons within a group can adjust their membrane potential (voltage) by exchanging reward and inhibition pulses based on external feedback information. The relative relationship between targets is then determined based on the positional relationship corresponding to the neuron with the highest membrane potential. Specifically, the target recognition or instance segmentation algorithm first outputs the contour region of the target element. and the outline area of ​​the reference element The outline region of the target element is denoted as: ； The reference element outline region is denoted as: ； Subsequently, in the outline area of ​​the target element respectively and reference element outline area Each sample is taken according to a uniform random sampling strategy. Each of the points is used as a target point. and reference points The first one mentioned above The coordinates of the target point and the reference point are denoted as follows: ； Subsequently, objective functions were defined for the seven positional relationships to calculate reward and inhibition impulse values. For the four directional relationships (left, right, top, and bottom), objective functions were defined based on the target point. With reference point The relative directions are used to construct the direction consistency coefficient; the larger the direction consistency coefficient, the closer the target point is to the corresponding ideal direction relative to the reference point; the consistency coefficients for the left, right, top, and bottom are defined as follows: ； ； ； ； The non-zero Euclidean distance between the target point and the reference point for: ； in, For the smallest positive number, the middle is set to This is used to avoid the denominator being zero when two points coincide; Based on the above consistency coefficient, the objective functions for the left, right, upper, and lower relationships are defined as follows: ； ； ； ； in, , , and The positive constant bias term is used to ensure that each objective function is always greater than 1, thereby affecting the reward function. The value is always less than 1, thus ensuring the stability of the neural network learning; the average value is set to 0.

01. When the target point is closer to the ideal left direction of the reference point, The closer to 1, the better. The closer to the minimum value Similarly, the closer the target point is to the ideal right, top, and bottom directions of the reference point, , , The closer each value is to 1, the better the corresponding objective function. , , The closer to the minimum value When the target point deviates from the corresponding direction or even is in the opposite direction, the corresponding consistency coefficient decreases, and the objective function increases accordingly. Therefore, the above four objective functions are used to characterize the significance of the relationship between the target element and the reference element to the left, right, above, and below, respectively. For internal and external relationships, modeling is based on the bidirectional sampling inclusion relationship between the target element contour region and the reference element contour region. This is because the contours of both the target and reference elements are often irregular, and the complex subway station environment can introduce false detections and biases into target recognition and instance segmentation results. Therefore, relying solely on the area or center position of the bounding box is insufficient to accurately reflect the true front-to-back occlusion relationship between the two. In contrast, by randomly sampling within both the target and reference element contour regions and statistically analyzing the proportion of each sampling point falling within the other's region, the enclosing relationship and front-to-back hierarchical relationship between the two under 2D image projection can be more effectively characterized. The mathematical expression is as follows: ； ； in, A function to determine the inclusion of the target point relative to the reference region. The region inclusion determination function is defined for the reference point relative to the target region; when the... Target sampling points Falling into the outline area of ​​the reference element Inside, The value is 1 if it is not 1, otherwise the value is 0; when the first Reference sampling points Fall into the outline area of ​​the target element Inside, The value is 1 if it is set to 1, otherwise the value is 0. To more precisely characterize cases where the target and reference overlap, a penalty coefficient is further introduced. , in settings Based on this, the first is defined. Consistency coefficient of internal relationship corresponding to each sampling point Consistency coefficient with external relations They are respectively: ； ； in, Used to characterize the The significance of the internal relationships corresponding to each sampling point Used to characterize the The significance of external relationships corresponding to each sampling point; when the target sampling point is located inside the contour region of the reference element and the reference sampling point is not located inside the contour region of the target element. The maximum value is reached; when both fall within each other's region, a penalty term for mutual inclusion is introduced. This will decrease accordingly, thereby reducing the salience of the internal relationship; correspondingly, when the target sampling point is located outside the contour region of the reference element and the reference sampling point does not fall inside the contour region of the target element, A larger value indicates a more significant external relationship; Based on the aforementioned consistency coefficients, we define the objective functions for internal relations respectively. Objective function of external relations for: ； ;in, and This is a positive constant bias term used to ensure that the objective function is always greater than 0. This causes the subsequent reward function to always be less than This ensures the stability of neural network learning, and the average value is set to 0.01; the more significant the internal relationship, the better. The closer objective function The closer to the minimum value The more significant the external relationship, The closer objective function The closer to the minimum value ; For proximity relationships, modeling is based on the Euclidean distance between the target point and the reference point; specifically, the Euclidean distance between the target point and the reference point as defined above is first... Normalization is performed to improve comparability under different image scale conditions. Let the current image width be... The height is Then the normalized distance Defined as: Further define the consistency coefficient of proximity relations. for: ； in, This is the distance attenuation coefficient, used to adjust the sensitivity of neighborhood relationships to changes in distance. Set to 3; the closer the target element is to the reference element, the better. The smaller, The closer to 1; as the distance between the two increases, Increase Gradually decrease; Based on this, the objective function for the proximity relation is defined as follows: ； in, This is a positive constant bias term used to ensure that the objective function is always greater than 0. This causes the subsequent reward function to always be less than This ensures the stability of neural network learning, and is set to 0.01; thus, the more significant the proximity relationship between the target element and the reference element, the better. The closer to 1, the better the objective function. The closer to the minimum value When the two are far apart, Decrease, objective function An increase indicates that the proximity relationship is gradually weakening; To satisfy the reward-driven mechanism of brain-like stochastic neural networks, the above-mentioned objective functions are further transformed into reward functions; for any positional relationship Its reward function is uniformly defined as: ； in, These correspond to left, right, top, inside, outside, and nearby relationships, respectively; specifically targeting the first... Round sampling, the objective function and reward function values ​​are represented as follows: Since all the above objective functions satisfy: ； Therefore, the corresponding reward function always satisfies: ； The smaller the objective function, the more closely the current sampling point pair conforms to the corresponding positional relationship, and the larger the reward function; conversely, the larger the objective function, the less significant the positional relationship, and the smaller the reward function. Each reward value serves as external feedback information input to the corresponding neuron. Neurons dynamically update the membrane potential value based on reward and inhibition impulses. Finally, the positional relationship corresponding to the neuron with the highest membrane potential value is used as the final positional determination result between the target element and the reference element. Since the objective functions corresponding to different positional relationships differ in numerical range, to avoid interference from this difference in the neuronal membrane potential update and competitive determination process, the seven positional relationships are divided into three groups of fully connected neuron clusters. Neurons within each group compete with each other. The left, right, upper, and lower neuron combinations form a group. Internal and external neurons are combined into a group. Nearby neurons are grouped separately. The neuron's membrane potential update process is as follows: 1) For each neuron, targeting the first... Round sampling, using an objective function to calculate the reward value. ; 2) Calculate the current reward / penalty threshold for the current neuron group. For clusters of neurons and , For clusters of neurons , ,in The learning rate is set to 0.

7. This serves as a reference value for the centroid distance between this type of target element and this type of reference element; 3) Comparison and the average judgment threshold before the update Update neuron weights: ； ； ； ； ； ； 4) Update the total emission rate of signals (including reward and punishment signals) for each neuron: ； 5) Normalize the weights of each neuron: ; ; 6) Calculate the membrane potential of each neuron: ; ； We propose a target classification and scoring model MEP-Net based on convolutional neural networks to automatically classify wall decorations and three-dimensional decorations and quantify the pleasure index. First, fixed-size image patches are extracted from high-attention areas for passengers, such as waiting areas, platform walls, and transfer passages, and then input into the MEP-Net scoring model for feature extraction. The network extracts color, texture, and spatial structure features from the decorative areas through a multi-layer convolutional structure, while enhancing key visual information of the decorative areas and suppressing background noise. Subsequently, the extracted high-dimensional features are mapped through a fully connected classification layer at the end of the model, outputting the target category probability, and classifying the decorative elements into two categories: wall decoration and three-dimensional decoration, based on the principle of maximum probability. MEP-Net takes the target classification results and their spatial information as input, and uses a multi-layer convolutional network to fuse and weight features of different categories. The score reflects the degree of influence of the corresponding environmental elements on the passenger experience, and can be further summarized to obtain the overall environmental evaluation results of the subway station. The scoring model takes as input decorative images of subway stations and networks as input and consists of a multi-level feature extraction and classification module, including: 5 3×3 convolutional layers; 5 max pooling layers; 1 rescaling layer for data normalization; 2 dropout layers; 1 fully connected layer; 1 dense layer; and 1 output layer.

4. The method according to claim 1, characterized in that, (3) The specific steps for constructing a quantitative algorithm for evaluating the quality of infrastructure user experience based on passenger behavior simulation are as follows: By simulating and analyzing passengers' travel paths, space access frequency, and main movement directions in subway stations, the impact of relevant infrastructure on passengers is quantified into time-related indicators. Key sampling points for user experience quality and their corresponding weights are determined, and the overall user experience quality level of the subway station environment is calculated based on this. Specifically, firstly, the space area of ​​the subway station Divided according to a spatial scale of 2 meters × 2 meters A grid cell, denoted as: ； in, Indicates the first One grid cell, This represents the total number of grids obtained from the spatial division of the subway station. Subsequently, based on the SUMO pedestrian simulation model, the travel path of passengers from the starting areas such as the platform, station hall, and transfer passage to the entrances / exits, stairs, escalators, and other key facilities is set, and the number of times passengers pass through each grid cell and the main movement direction within that grid cell are recorded during the simulation. For any grid cell The number of visits is defined as: ； in, This represents the total number of simulated passengers; To access the decision function, when the first... Passengers passed through the grid unit The value is 1 if the condition is met, and 0 otherwise; the number of visits. The larger the value, the more frequently passengers pass by or pay attention to the grid cell, and the greater its impact on the overall user experience quality evaluation. Furthermore, statistics on passengers in grid units The dominant orientation of the mesh cell is obtained by determining the main direction of motion within it. The dominant orientation is used to determine the acquisition direction of the environmental image at that location, so that the acquired image can reflect the passenger's visual perception of the station environment during actual passage as much as possible; thus, it can be determined based on the number of grid visits. Dominant Orientation In addition to the spatial functional area where the grid is located, a set of key sampling points for user experience quality was obtained by filtering. , denoted as: ； in, Indicates the first Key points for sampling user experience quality. Indicates the number of key points selected; For each sampling key point The environmental image is acquired based on the dominant orientation of its corresponding grid cell, and the user experience quality score of the key point is calculated using the aforementioned machine vision evaluation model, denoted as: ； in, Indicates the first The user experience quality score for each key point ranges from 0 to 5. This represents the scoring function obtained based on a comprehensive evaluation system and a machine vision model; To reflect the different contributions of various key points to the overall station environment evaluation results, the weights of each key point are further calculated based on the number of visits obtained from passenger behavior simulation; key points are defined as follows: The number of visits to the corresponding grid cell is Then its weight is defined as: ； in, Indicates the first The weight of each key point in the overall evaluation; the more times a key point is visited, the greater its corresponding weight, indicating that the location has a more significant impact on the overall quality of the passenger experience; Finally, the subway station Overall environment user experience quality evaluation score The weighted sum of the scores for each key point is obtained as follows: 。 5. The method according to claim 1, characterized in that, Also includes: First, the aforementioned machine vision recognition algorithm is used to detect exit signs, transfer guidance signs, medical facilities, and automated external defibrillators (AEDs) in subway station images to obtain the spatial location set of the corresponding facilities; then, the station hall, platform, stairs, escalators, transfer passages, entrances and exits, and large open spaces are converted into SUMO network maps. In the SUMO network map, the passable area within the station is represented as the pedestrian movable space, and the spatial connections such as stairs, escalators, transfer passages and entrances and exits are represented as pedestrian paths. The locations of exit signs, transfer guidance signs, medical rescue facilities and AED infrastructure are mapped to the corresponding simulation space coordinates. Based on historical passenger flow card swiping data at the station, passenger travel demands are generated under different origins, destinations, and passenger flow intensities; the passenger travel demands include the passenger's generated location, target exit or transfer platform, departure time, and travel route; For wayfinding facilities such as exit signs and transfer guidance signs, the spatial location of the facilities, the station network map, the distribution of passenger origins and destinations, and passenger flow density are input into the passenger behavior simulation model. The passenger behavior simulation model is used to simulate the actual passage process of passengers in the station and output behavioral indicators related to the layout of wayfinding facilities, including hesitant behavior, detour behavior, incorrect direction selection, transfer passage time, and total passage time to reach the target exit or transfer facility at intersections, large open spaces, transfer nodes, and near entrances and exits. Therefore, the simulation output results can be used to determine whether the location of the guidance facilities can effectively support passengers in making direction judgments and route selections. Let the first The increase in travel time for each passenger during the simulation due to the layout of guidance facilities is: The weighted average travel time increment corresponding to the guidance facility is expressed as: ； in, This indicates the degree to which the layout of guidance facilities affects the overall efficiency of passenger traffic. The smaller the value, the more reasonable the layout of exit signs and transfer guidance signs, and the lower the possibility that passengers will hesitate, detour, or make the wrong choice of direction in the station. This indicates the number of simulated passengers participating in the evaluation; Map the weighted average travel time increment corresponding to the wayfinding facility to the wayfinding facility evaluation score: ； in, Indicates the evaluation score for guidance facilities; This indicates the upper limit of the evaluation score. This indicates the weighted average travel time increment caused by the layout of guidance facilities; This indicates the acceptable threshold for incremental travel time; when When smaller, near ;when When the value increases, it indicates that the layout of the guidance facilities has a significant adverse impact on passenger traffic efficiency, and the evaluation score will decrease accordingly. For emergency service facilities such as medical aid facilities and automated external defibrillators (AEDs), the spatial location of the medical facilities, station network map, passenger distribution, regional passenger flow density, and the location of the sudden medical event are input into the passenger behavior simulation model. The passenger behavior simulation model is used to simulate the process of obtaining the nearest medical aid facility or AED from the location of the event and delivering it to the event location after a sudden medical event occurs, and outputs the time for obtaining the medical aid facility, the travel distance, and the actual response time affected by congestion. The service coverage and emergency response efficiency of medical aid facilities and AEDs are evaluated based on the simulation output results. Let the first In a simulated emergency medical event scenario, the response time required to obtain medical aid facilities or AEDs and deliver them to the event location is: The weighted average response time for medical rescue facilities is expressed as: ； in, This represents the weighted average response time of medical rescue facilities or AEDs. The smaller the value, the more reasonable the spatial layout of medical rescue facilities and AEDs, and the stronger the emergency service coverage. Indicates the first The weight of each emergency medical event scenario in the evaluation is set as follows: ; The weighted average response time corresponding to medical emergency facilities is mapped to a medical facility evaluation score: ； in, This indicates the evaluation score for medical rescue facilities; Indicates the upper limit of the evaluation score; This represents the weighted average response time corresponding to medical emergency facilities; This indicates the threshold for security response time.

Images