Variable-controllable immersive evacuation illumination reaction data acquisition and modeling experiment method

Abstract

The invention provides a variable-controllable immersive evacuation illumination response data acquisition and modeling experiment method, which comprises the following steps: constructing an immersive evacuation scene, and adjusting illumination variables to form different illumination guide schemes; synchronously collecting eye movement physiology, behavior track and regional attention data of the subject; repeating experiments according to a preset design, and accumulating multi-scene multi-sample data; original data is processed, key features are extracted, and a multi-dimensional reaction feature vector is constructed and mapped to a high-dimensional space; inputting the feature vectors into a trained machine learning model, and classifying and predicting the evacuation reaction mode; a simulation environment is constructed, the state, action and reward function of reinforcement learning are defined, a deep reinforcement learning training agent is adopted, an evacuation behavior is simulated, and the illumination design is iteratively optimized. According to the method, a complete technical closed loop of multi-dimensional real reaction data acquisition, intelligent analysis and simulation verification is realized under a controllable condition, and a scientific and quantitative basis is provided for optimization of an evacuation lighting system.

Description

Technical Field

[0001] This invention relates to the technical field of emergency evacuation and public safety, and in particular to an experimental method for data acquisition and modeling of immersive evacuation lighting response with controllable variables. Background Technology

[0002] In emergencies in large public buildings, appropriate evacuation lighting and signage are crucial for guiding people to evacuate safely and quickly. Traditional studies on the effectiveness of evacuation guidance typically analyze human behavior and decision-making through evacuation drills, questionnaires, or video observations. However, these methods suffer from limitations in data dimensionality and accuracy. Relying solely on macro-level indicators such as personnel routes and evacuation times makes it difficult to discern micro-level cognitive responses during the evacuation process.

[0003] In recent years, the development of immersive virtual reality and augmented reality technologies has provided new avenues for evacuation behavior research, making it possible to construct realistic virtual emergency scenarios under safe and controllable experimental conditions. By constructing realistic virtual emergency scenarios and combining them with relevant data acquisition devices, conditions are created for in-depth research into the distribution of attention and decision-making mechanisms during emergency evacuations. These technologies can effectively reproduce the perceptual environment of real-world scenarios while avoiding the safety risks and uncontrollable conditions inherent in traditional field experiments, providing a more advantageous platform for related research.

[0004] With the rise of intelligent technologies such as machine learning and deep learning, their application in evacuation behavior analysis will become an industry trend. These technologies have shown great application potential in scenarios such as behavioral choice prediction and safety early warning in emergency situations. However, in the field of evacuation lighting and signage design, existing technologies still have significant shortcomings. A systematic solution to the problem has not yet been formed. It is unable to achieve precise control of key variables and effective collection of multi-dimensional data, and it also lacks the ability to deeply mine and intelligently analyze data, making it difficult to provide comprehensive and scientific technical support for the optimization of evacuation guidance. Summary of the Invention

[0005] The purpose of this invention is to provide an experimental method for acquiring and modeling immersive evacuation lighting response data with controllable variables. This method solves the technical problems of existing technologies, which are unable to acquire multi-dimensional real response data during evacuation under controllable conditions and lack the ability to intelligently mine and analyze the data in depth, thus failing to provide a scientific and quantitative basis for optimization of evacuation lighting design.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: This invention provides a variable-controllable experimental method for immersive evacuation lighting response data acquisition and modeling, comprising the following steps: S1. Construct an immersive evacuation scenario corresponding to a large public building, deploy programmable evacuation lighting devices in the scenario, and adjust lighting variables through a control system to form different lighting guidance schemes. S2. Using sensing and tracking devices, the subject's eye movement physiological data, behavioral trajectory data, and regional attention data are collected synchronously, and all data are recorded and integrated synchronously through a unified timestamp. S3. Place the subjects in an immersive scenario to simulate the real evacuation process, adjust the lighting variables according to the preset experimental design, and repeat the experiment to accumulate sample data from multiple scenarios and multiple subjects. S4. Process the collected raw data, extract key eye-tracking indicators, key features of behavioral trajectories, and regional attention features, and standardize or normalize the extracted features. S5. Combine the extracted multiple features to construct a multidimensional response feature vector, and map the response data of different subjects at different times to the same high-dimensional feature space; S6. Input the multidimensional response feature vector into a trained machine learning model to classify and predict the evacuation response patterns of the subjects. S7. Construct a simulation environment. Based on an immersive evacuation scenario and lighting control, define the state space, action space, and reward function for reinforcement learning. Use deep reinforcement learning algorithms to train the agent, simulate evacuation behavior, and iteratively optimize the lighting design.

[0007] Furthermore, in step S1, the immersive evacuation scene is constructed using virtual reality, augmented reality, or mixed reality technologies, and virtual emergency environment elements can be loaded to create an emergency atmosphere.

[0008] Furthermore, in step S1, the lighting variables include at least one of brightness, flashing frequency, color, and on / off mode.

[0009] Furthermore, in step S2, the eye movement physiological data includes at least one of fixation point position, fixation duration, number of saccades, saccade amplitude, and pupil diameter changes; The behavioral trajectory data includes at least one of walking path, gait characteristics, and head orientation.

[0010] Furthermore, in step S2, the area attention data is obtained by defining attention areas in the scene. The attention areas include at least one of the locations of safety exits, evacuation signs, and obstacle areas. The collected area attention data is the time and frequency of the subject's stay in each attention area.

[0011] Furthermore, in step S4, the key eye movement indicators include total number of fixations, average duration of a single fixation, total number of saccades, average saccade amplitude, maximum change in pupil diameter, and blink frequency; the key characteristics of the behavioral trajectory include total walking distance, average walking speed, number of obstacle avoidances, number of pauses, change in head orientation, and step length variation coefficient; the regional attention characteristics include the number of fixations in each attention region, fixation dwell time, and dwell time percentage.

[0012] Furthermore, in step S5, the dimensions of the multidimensional response feature vector include evacuation time, total duration of gaze at evacuation indicator signs, pupil diameter related indicators, walking path characteristics, head orientation change indicators, and number of pauses. Multiple sub-vectors can be constructed or weights can be introduced to highlight the importance of specific dimensions according to the comparison purpose.

[0013] Furthermore, step S6 specifically includes: Unsupervised clustering algorithms are used to cluster multidimensional response feature vectors to classify different response pattern categories; and / or, supervised classification algorithms are used to construct classification prediction models to identify or predict the response types of subjects. Deep learning methods are used to train and model response feature vectors or time series data to predict user evacuation behavior.

[0014] Furthermore, step S7 specifically includes: A reinforcement learning simulation environment is constructed, which maps the immersive evacuation scene and the current lighting control scheme into the state space and action space of reinforcement learning. The state space includes the position information, orientation, and state of the indicator lights in the environment of the agent, and the action space includes the selection of the next movement direction. A reward function associated with the evacuation target is set, and a deep reinforcement learning algorithm is used to train the agent based on the reward function, so that it gradually learns the optimal evacuation strategy under different lighting schemes. Using the final agent strategy, the evacuation trajectories of a large number of virtual individuals are simulated under a given lighting scheme, and the lighting scheme is evaluated and iteratively optimized.

[0015] Compared with the prior art, the present invention has at least the following beneficial effects: This invention constructs an immersive virtual evacuation scenario and precisely controls lighting variables, effectively inducing genuine responses from subjects under safe and repeatable experimental conditions. Simultaneously, it employs non-contact eye-tracking, motion capture, and other multi-sensor devices to synchronously collect subjects' eye-tracking physiological indicators, behavioral trajectories, and regional attention data. This achieves comprehensive recording of individual multi-dimensional, involuntary responses during evacuation, overcoming the limitations of traditional methods such as limited data and uncontrollable conditions, and significantly improving the richness and ecological effectiveness of the data.

[0016] This invention preprocesses multi-source data to construct a multi-dimensional feature vector that comprehensively represents individual response patterns, and maps it to a high-dimensional space for mathematical analysis. By introducing machine learning methods and deep learning models, it can automatically classify, deeply mine, and dynamically predict evacuation response patterns. This enables intelligent identification of behavioral patterns from massive amounts of data, evaluation of lighting scheme effects, and even prediction of individual behavioral trends, providing objective and quantitative decision-making basis for the precise optimization of evacuation guidance.

[0017] This invention further utilizes a reinforcement learning framework to abstract virtual scenes and lighting control into a simulation environment, training agents to learn optimal evacuation strategies. This allows for the simulation of the behavior of a large number of virtual individuals under a given lighting scheme, enabling the evaluation of overall evacuation efficiency and iterative optimization of lighting design in a low-cost and high-efficiency manner. This method forms a complete technical closed loop from data acquisition and intelligent analysis to simulation verification, and possesses good versatility. It can be adapted to emergency evacuation research and system optimization in various large public spaces such as airports, subways, and shopping malls, demonstrating significant practical value and promising prospects for wider application. Attached Figure Description

[0018] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 A flowchart of the experimental method provided in this embodiment; Figure 2 This is a schematic diagram of the immersive evacuation scenario and data acquisition system provided in this embodiment; Figure 3 This is a schematic diagram of the construction and cluster analysis of reaction feature vectors provided in this embodiment. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] This embodiment provides a variable-controllable experimental method for immersive evacuation lighting response data acquisition and modeling. Please refer to... Figures 1-3As shown, the technical solution specifically includes the following steps: Step 1: Immersive Scene Construction and Variable Control. Construct a highly immersive evacuation scenario simulation system, with scenarios corresponding to large spaces such as airport terminals, subway stations, and shopping mall corridors. Virtual Reality (VR) or Augmented Reality (AR / MR) technologies are preferred to recreate real-scale architectural environments and emergency evacuation scenarios in the laboratory. Deploy programmable evacuation lighting devices (LED light strips, emergency indicator lights, etc.) in the scenario, and adjust their brightness, flashing frequency, color, and on / off modes through a control system to create different lighting guidance schemes. Lighting variables can be pre-planned or adjusted in real time to explore the impact of various lighting conditions on subject behavior. Virtual environmental elements such as fire and smoke can also be added to the scenario to create a realistic emergency atmosphere.

[0022] The experimental system of this invention includes: a simulated indoor architectural environment, a controllable evacuation lighting device, subjects wearing eye-tracking devices, and cameras / locators for capturing behavioral trajectories, etc. Figure 2 As shown in the figure, arrows indicate the subject's movement path, and light source symbols indicate the direction of illumination.

[0023] Step 2: Experimental Equipment and Data Acquisition Setup. Multiple sensing and tracking devices are deployed without direct contact with the subjects to simultaneously collect physiological and behavioral responses. Eye-tracking devices are preferred for acquiring physiological data related to the subjects' gaze, such as fixation point location, fixation duration, saccade frequency and amplitude, and pupil diameter changes. The eye tracker can employ a high-frequency near-infrared camera (such as eye-tracking glasses or an eye-tracking module built into a VR headset) to ensure accurate eye movement information is captured even during movement. Simultaneously, a motion capture system records the subjects' behavioral trajectories and body posture information. This can be achieved through VR locators, inertial sensors, or cameras deployed around the scene to acquire data such as the subjects' walking paths, gait characteristics (e.g., stride length, cadence, walking speed), and head orientation (determining the approximate direction of their gaze based on head posture) during evacuation. Furthermore, several areas of interest (AOIs) are defined in the virtual scene, such as safety exit locations, evacuation signs, and obstacle areas. The time and frequency of the subjects' stay in these areas are automatically recorded, i.e., the "hotspot" dwell index. All data collection processes were completed in a non-contact manner to ensure that they did not interfere with the behavior of the subjects, and a unified timestamp was used to achieve synchronous integration of multi-source data.

[0024] Step 3: Experimental Procedure and Data Acquisition. Subjects (single or multiple) are placed in the aforementioned immersive scenario to simulate a real evacuation process. Before the experiment, subjects are familiar with the basic situation (such as the general structure of the building), but are not informed of specific evacuation signs and lighting details to ensure natural and realistic reactions. An emergency scenario is triggered (fire alarm sounds, lighting switches to emergency mode), and subjects must locate safe exits and evacuate according to instructions. During this process, the control system adjusts lighting variables in real time according to the experimental design (changing the flashing frequency or color of the lights at different stages) to present the predetermined evacuation lighting scheme. Eye-tracking and motion behavior data are recorded in real time throughout the evacuation process. The experiment can be repeated multiple times, each time with different combinations of lighting variables or scenario conditions, ensuring data coverage of various scenarios. After repeated testing by multiple subjects, a large amount of sample data is accumulated for subsequent analysis and modeling. A typical experimental procedure is shown in the attached diagram, including scenario preparation, subject calibration, evacuation triggering, data acquisition, and storage, ensuring a standardized and consistent process.

[0025] Step 4: Data Preprocessing and Feature Extraction. The raw data obtained in Step 3 is preprocessed and analyzed to convert it into quantitative features. For eye-tracking data, key indicators are extracted using professional eye-tracking analysis software or a self-developed algorithm, including the total number of fixations, average duration of a single fixation, total number of saccades, average saccade amplitude, maximum pupil diameter change, and blink frequency for each subject during an evacuation. Simultaneously, the number of fixations and fixation duration in each region are calculated based on a predefined AOI, and the total duration and number of fixations on evacuation signs are statistically analyzed to measure attention distribution. For behavioral trajectory data, features such as total walking distance, average walking speed, number of obstacle avoidances, number of pauses, and head orientation change are calculated. Head orientation data is used to determine whether subjects frequently look back during evacuation (potentially indicating uncertainty or anxiety), and gait speed is used to determine hesitation. Gait features, such as the coefficient of variation of stride length (reflecting the degree of panic), can be extracted. All extracted features are normalized or standardized in a uniform format to eliminate dimensional differences and make them suitable for subsequent algorithm analysis.

[0026] Step 5: Response Vector Construction and High-Dimensional Mapping. The multiple features extracted in Step 4 are combined to construct the response feature vector corresponding to each experiment. Assuming n key features are selected, each experiment (one evacuation process for each subject) can be represented as an n-dimensional vector: R = [ x 1 , x 2 , ..., x n ], Among them, each component x nThese are the corresponding feature values. For example, a 6-dimensional reaction vector instance might contain: x 1 = Total evacuation time x 2 = Total time spent looking at evacuation signs. x 3 = Mean pupillary dilation, x 4 = The tortuosity of the walking path (ratio to the shortest path). x 5 = Total angle of change in head orientation x 6 = Number of pauses during evacuation, etc.

[0027] These features collectively characterize the physiological and behavioral response patterns of subjects under specific lighting conditions. This vectorized representation maps response data from different subjects at different times to the same high-dimensional feature space, facilitating mathematical analysis and pattern recognition. If needed, multiple sub-vectors can be constructed or weights introduced to highlight the importance of specific dimensions for different comparison purposes.

[0028] Step 6: Classification and prediction of evacuation response patterns.

[0029] Step 6.1: Response Classification and Cluster Analysis. In the feature vector space, a machine learning model is introduced to classify and cluster the subjects' response patterns. First, an unsupervised clustering algorithm (such as...) is applied. k Algorithms such as mean clustering, DBSCAN, and hierarchical clustering search for natural groups among a large number of response vectors. These algorithms measure similarity based on distances between vectors (e.g., Euclidean distance) and group vectors with similar response patterns into one class. The number of clusters is adjusted accordingly. k The optimal clustering scheme is determined by combining indicators such as the silhouette coefficient. Specifically, this invention uses the elbow method and the average silhouette coefficient method to determine the optimal clustering scheme. k First, calculate the sum of squares of the distance from each sample to the cluster center for different cluster numbers and plot the curves, selecting the locations where the curves show inflection points as candidates; second, calculate the sum of squares of the distances from each sample to the cluster center for different cluster numbers. k The average contour coefficient, select the value corresponding to the maximum value. k The final number of clusters is used to determine several response pattern categories. Each category corresponds to a group of evacuation responses with similar characteristics, potentially exhibiting different patterns such as "decisive" (rapid decision-making, short fixation time on the sign, straight path) and "hesitant" (multiple glances, many pauses, circuitous path). The cluster center vector for each cluster is calculated. C j = (1 / |C j |)∑i As representative features of this response category, the typical value differences across each dimension are analyzed to describe the behavioral characteristics of this category. Cluster analysis, in particular, can reveal the heterogeneity of responses to evacuation lighting among different population groups, contributing to the understanding of hidden pattern structures.

[0030] like Figure 3 As shown, this figure illustrates the distribution of response vectors in the feature space using two-dimensional or three-dimensional coordinate graphs, with different colors marking the different response categories obtained from clustering. The figure also indicates the location of the cluster center for each category and the typical feature differences of each category in two feature dimensions (e.g., "decisive" has lower values ​​for fixation time and evacuation time), helping to understand the classification results. By identifying the subject's physiological panic state through feature vectors, control commands can be output to adjust the brightness and frequency of the physical lighting device.

[0031] Based on the clustering results, supervised learning methods can be combined to automatically classify and identify responses. If the experimental results are pre-labeled (responses are divided into "effective" and "ineffective" categories based on whether the subjects successfully escaped quickly or correctly followed instructions), a discrimination model can be built using classification algorithms such as Support Vector Machines (SVM) and neural networks. The model is trained using the response vector R as input features. f(R) Output the corresponding response category. Taking SVM as an example, during training, the decision hyperplane is optimized. f(R) = w·R + b This allows for maximum separation of categories in a high-dimensional space. The invention employs an SVM with a radial basis function (RBF) kernel, determining the penalty parameters through cross-validation grid search. C and kernel width γ . γ Too large an area can lead to overfitting of the model. γ Too small an exponent would result in underfitting; therefore, the optimal value was selected from the exponential sequence between 2^-5 and 2^5, with the training set classification accuracy used as the evaluation criterion. The trained model can be used to classify the evacuation response types of new subjects in real time: when someone is in the experimental scenario, based on the real-time vector formed by their initial data, the model can predict which response mode they belong to (e.g., whether they will escape successfully or may become confused), thus predicting the user's problem progression. This classification prediction helps to identify potential evacuation difficulties in a timely manner and verify whether there are common problems hindering evacuation under certain lighting schemes.

[0032] Step 6.2: Predictive Modeling Based on Deep Learning. To further improve prediction accuracy under complex patterns, deep learning models can be introduced to train and model response vectors or raw time-series data. Multi-layer fully connected neural networks or convolutional neural networks (CNNs) are constructed to map response features across dimensions to higher-level abstract representations, automatically learning nonlinear combination features to predict target outputs (such as evacuation success rate, individual risk level). For time-series data (such as time-varying sequences of eye movements and motion trajectories), models such as recurrent neural networks (RNNs) or long short-term memory networks (LSTMs) can be used to capture dynamic patterns of response evolution over time, thereby predicting user behavior and whether someone will deviate from the correct route within the next 10 seconds. Deep learning models, trained on large amounts of historical data, can discover complex features that are difficult to extract using manual rules. The neural network uses a two-layer fully connected structure with 64 and 32 hidden layer nodes, respectively, and the activation function is ReLU. For the LSTM model, 32 hidden units are used, and its input, forget, and output gates consist of fully connected layers and a sigmoid activation function. The formula for updating the memory unit is: C t = f t * C (t 1) + i t * t , in, To remember the state of the memory unit at the current moment, Output for the forget gate. C (t 1) The state of the memory unit in the previous moment. For input gate output, This represents the state of a candidate memory cell.

[0033] This allows for the capture of long-term dependencies while avoiding gradient vanishing, thus improving prediction accuracy and generalization ability. Notably, to prevent overfitting and ensure model reliability, this invention employs methods such as cross-validation to evaluate model performance and can integrate multiple models (e.g., ensemble learning) to enhance robustness.

[0034] Step 7: Reinforcement Learning-Based Behavior Simulation and Optimization. This invention also utilizes reinforcement learning methods to simulate and model evacuation behavior to verify and improve the lighting guidance strategy. A simulation environment is constructed, abstracting the virtual evacuation scenario and lighting control into a reinforcement learning environment state and action space. The state can include the current position and orientation of the subject (or agent), as well as the state of indicator lights in the environment; the action is defined as the agent's choice of movement direction in the next step. Specifically, the state space consists of two-dimensional position coordinates and lighting state; the action space includes eight movement directions (north, south, east, west, and four diagonal directions) and nine discrete actions, including remaining stationary. The reward function is designed as follows: +1 reward for reaching the exit; +0.1 reward if the distance to the exit is shortened after movement; -1 penalty if the distance to the exit increases or an obstacle is collided after movement; and 0 reward in other cases. The reward function is set to encourage the agent to quickly find the safe exit and follow the lighting guidance (e.g., high reward for reaching the exit, penalty for going in the wrong direction or staying for a long time). A deep reinforcement learning algorithm (such as DQN, A3C, PPO, etc.) is used to train an agent, enabling it to progressively learn the optimal evacuation strategy under different lighting schemes. During training, the agent simulates crowd decisions in various situations, iteratively updating its strategy until convergence. The final agent strategy can be used for behavioral simulation: under a given lighting scheme, the evacuation trajectories of a large number of virtual individuals are simulated to evaluate overall evacuation efficiency and potential problems. If the agent frequently gets lost or delayed under certain lighting configurations, it indicates that the scheme may be unsatisfactory. By adjusting the lighting variables and simulating again, the lighting design can be iteratively optimized. This reinforcement learning simulation module acts as a "virtual subject," capable of simulating crowd behavior and predicting the effects of different strategies, providing a reference for improving evacuation lighting systems. It is worth emphasizing that the reinforcement learning model can also be fine-tuned by incorporating knowledge extracted from the reactions of real subjects, making the simulation closer to real human behavior.

[0035] In summary, the technical solution of this invention closely integrates immersive experimental data acquisition with intelligent algorithm modeling, covering the complete process from data acquisition and feature quantification to model analysis and simulation verification.

[0036] This scheme can quantitatively characterize the physiological and behavioral response patterns of individuals under different lighting conditions, and can also use the latest artificial intelligence technology to classify, predict and optimize the patterns, providing scientific and advanced guidance for the design of evacuation lighting and signage systems. Specific Implementation The method of the present invention will be further described below with reference to specific embodiments. These embodiments are only used to explain the technical solution of the present invention and are not intended to limit the scope of the present invention.

[0038] Example: Immersive evacuation simulation experiment in an airport terminal.

[0039] The architectural layout of a large airport terminal was selected as the blueprint for the virtual scene, including elements such as long corridors, boarding gates, and signage. The virtual environment was constructed using engines such as Unity3D and connected to an HTC Vive Pro eye-tracking headset to achieve an immersive VR experience and high-frequency eye-tracking functionality. The evacuation lighting system was simulated by dynamic indicator lights installed on the virtual corridor floor (corresponding to real-world evacuation signs), and the lighting sequence and frequency could be controlled programmatically. This embodiment sets two lighting schemes: Scheme A is continuous constant illumination (all indicator lights are constantly lit), and Scheme B is dynamic progressive illumination (the nearest indicator light flashes at a high frequency, while distant lights illuminate sequentially to guide the way).

[0040] The experiment recruited 20 volunteers (half male and half female, aged 20-40) who entered a VR scene in batches to perform an emergency evacuation mission. Each volunteer wore a Vive Tracker on their foot to record their walking trajectory and speed. Simultaneously, the headset's built-in eye-tracking module collected their eye movement data at a frequency of 120Hz. Volunteers were placed at a virtual starting point in an airport terminal waiting area and, upon hearing a fire alarm, were required to quickly find the nearest emergency exit. The experiment employed a balanced Latin square design, with each participant performing one evacuation mission under both scenario A and scenario B. The following data was automatically recorded during the process: Eye-tracking data includes the screen coordinate sequence of the fixation point, the duration of each fixation, the timing and amplitude of saccades, and the sequence of pupil diameter size over time. After the experiment, the software calculates the total number of fixations, average fixation duration, maximum pupil diameter, and its increase for each experiment.

[0041] Behavioral data includes the volunteer's location trajectory over time (provided by the locator, with coordinates recorded every 0.02 seconds), walking speed curves, number and duration of stops (speed below a threshold), and head orientation angle over time. The software further calculates features such as total walking distance, average speed, and time taken to first discover the exit.

[0042] Area attention data: Based on the defined AOI area (the area where the exit sign is located), calculate the cumulative time and number of times volunteers' eyes fall on the area; similarly calculate the proportion of their time spent in the main route and secondary route areas, etc.

[0043] By summarizing the above data, a response feature vector for each experiment is constructed: This embodiment selects 12-dimensional features. The meaning of each dimension is as follows: x i1 Time (s) required to complete the evacuation; x i2 The spatial distance (m) between the exit sign and the exit sign when it is detected;x i3 The percentage of total time spent looking at the export sign in the total experimental time (%). xi 4 represents the average saccade amplitude (°); x i5 Maximum pupil diameter (mm); x i6 The maximum increase (%) in pupil diameter relative to the initial value. x i7 Total walking distance (m); x i8 Average walking speed (m / s); x i9 x represents the number of pauses (times). i10 The total angle of change in head orientation (°); x i11 The percentage of distance traveled in the wrong direction (%). x i12 The percentage of time spent in the main path hotspot. These dimensions can comprehensively reflect the evacuation behavior and involuntary response characteristics of volunteers under specific lighting conditions.

[0044] Unsupervised clustering analysis was performed on 40 response vectors obtained from 20 volunteers under two lighting schemes, using k-means clustering. k=3 The algorithm groups response patterns to identify typical evacuation response categories under different lighting conditions. The k value is determined by a combination of elbow method and average profile coefficient analysis.

[0045] Category 1 (Rapid and Decisive): This cluster contains 15 samples, characterized by short evacuation times, minimal fixation time at exit signs, moderate pupil dilation, and near-straight walking paths. The cluster centers show they found the exit on average within 6 seconds, with almost no detours. This group largely comes from Solution B (Dynamic Indication), indicating that the dynamic indicator lights effectively guided their rapid evacuation.

[0046] The second category (hesitant searchers): This cluster has 18 samples, characterized by longer evacuation times (high), frequent pauses and glances around the scene (high number of pauses, large changes in head angle), and a relatively low percentage of time spent fixating on landmarks. Their pupils show significant increase (high), suggesting psychological tension. The cluster center shows an average evacuation time of 20 seconds, with frequent wrong turns and retracing (significantly higher than the first category). This category is present in both schemes A and B, indicating that some individuals will still search for too long due to anxiety even with indicator lights.

[0047] The third category (cautious and attentive): This category consists of 7 samples, whose characteristics fall between the first two categories. They exhibit moderate evacuation time but a high proportion of eye contact with exit signs, indicating they spend more time confirming instructions before acting. Their pupil dilation is not significant, their walking speed is slow and steady, and they almost never go astray. This category is considered to employ a cautious but effective strategy, taking slightly longer than the decisive category but ensuring a safe evacuation.

[0048] Next, using the clustering results as labels, SVM was employed for supervised classification training of the response vectors. The model was trained with 60% of the samples and validated with 40%, achieving a final classification accuracy of 85%. Applying this classification model, it can predict the volunteer's type within the first few seconds of an experiment based on real-time collected features. For example, a new volunteer exhibited 10% pupil dilation, paused twice, and failed to fixate on any marker within the first 5 seconds; the model determined this to be highly likely a "search-hesitant" volunteer, suggesting improvements to their guidance. This validated the model's predictive usability.

[0049] Finally, this embodiment utilizes Unity's built-in ML-Agents tool for reinforcement learning simulation. An intelligent agent is trained in a virtual terminal environment to search for exits from random locations. The agent's actions involve discrete movement in eight directions, with rewards for shortening the distance to the exit and penalties for moving further away. Finding the exit is given a high reward. After tens of thousands of training iterations, the agent successfully learned to follow the direction of the indicator lights, achieving fast convergence and optimal path selection under dynamic indicator scheme B. However, under the constant-light scheme A, the agent frequently lingered at intersections, resulting in poor convergence. This aligns with the results of real-world experiments—scheme B offers higher guidance efficiency. Reinforcement learning simulations extensively validated the advantages of scheme B and revealed that increasing the indicator light flashing frequency by 20% could further shorten agent evacuation time. Therefore, an improvement suggestion is proposed: when installing indicator lights in actual airports, adopt a high-frequency dynamic guidance mode to improve group evacuation efficiency.

[0050] In summary, the embodiments fully demonstrate the feasibility and effectiveness of the method of the present invention. Through immersive experiments combined with multidimensional data acquisition, we successfully captured the subtle differences in the responses of different individuals under evacuation lighting; with the help of intelligent modeling and simulation, we quantitatively evaluated the advantages and disadvantages of the lighting scheme and provided optimization directions. The technical solution of the present invention can be applied to the research and development and improvement of actual evacuation guidance systems, and has important practical value and promotion significance.

Claims

1. A variable-controllable experimental method for immersive evacuation lighting response data acquisition and modeling, characterized in that, Includes the following steps: S1. Construct an immersive evacuation scenario corresponding to a large public building, deploy programmable evacuation lighting devices in the scenario, and adjust lighting variables through a control system to form different lighting guidance schemes. S2. Using sensing and tracking devices, the subject's eye movement physiological data, behavioral trajectory data, and regional attention data are collected synchronously, and all data are recorded and integrated synchronously through a unified timestamp. S3. Place the subjects in an immersive scenario to simulate the real evacuation process, adjust the lighting variables according to the preset experimental design, and repeat the experiment to accumulate sample data from multiple scenarios and multiple subjects. S4. Process the collected raw data, extract key eye-tracking indicators, key features of behavioral trajectories, and regional attention features, and standardize or normalize the extracted features. S5. Combine the extracted multiple features to construct a multidimensional response feature vector, and map the response data of different subjects at different times to the same high-dimensional feature space; S6. Input the multidimensional response feature vector into a trained machine learning model to classify and predict the evacuation response patterns of the subjects. S7. Construct a simulation environment. Based on an immersive evacuation scenario and lighting control, define the state space, action space, and reward function for reinforcement learning. Use deep reinforcement learning algorithms to train the agent, simulate evacuation behavior, and iteratively optimize the lighting design.

2. The experimental method for data acquisition and modeling of immersive evacuation lighting response with controllable variables according to claim 1, characterized in that, In step S1, the immersive evacuation scenario is constructed using virtual reality, augmented reality, or mixed reality technologies, and virtual emergency environment elements can be loaded to create an emergency atmosphere.

3. The experimental method for data acquisition and modeling of immersive evacuation lighting response with controllable variables according to claim 1, characterized in that, In step S1, the lighting variables include at least one of brightness, flashing frequency, color, and on / off mode.

4. The experimental method for data acquisition and modeling of immersive evacuation lighting response with controllable variables according to claim 1, characterized in that, In step S2, the eye movement physiological data includes at least one of fixation point position, fixation duration, number of saccades, saccade amplitude, and pupil diameter changes; The behavioral trajectory data includes at least one of walking path, gait characteristics, and head orientation.

5. The experimental method for data acquisition and modeling of immersive evacuation lighting response with controllable variables according to claim 1, characterized in that, In step S2, the area attention data is obtained by defining attention areas in the scene. The attention areas include at least one of the locations of safety exits, evacuation signs, and obstacle areas. The collected area attention data is the time and frequency of the subject's stay in each attention area.

6. The experimental method for data acquisition and modeling of immersive evacuation lighting response with controllable variables according to claim 1, characterized in that, In step S4, the key eye movement indicators include total number of fixations, average duration of a single fixation, total number of saccades, average saccade amplitude, maximum change in pupil diameter, and blink frequency; the key characteristics of the behavioral trajectory include total walking distance, average walking speed, number of obstacle avoidances, number of pauses, change in head orientation, and step length variation coefficient; the regional attention characteristics include the number of fixations in each attention region, fixation dwell time, and the percentage of dwell time.

7. The experimental method for data acquisition and modeling of immersive evacuation lighting response with controllable variables according to claim 1, characterized in that, In step S5, the dimensions of the multidimensional response feature vector include evacuation time, total duration of gaze at evacuation indicator signs, pupil diameter related indicators, walking path characteristics, head orientation change indicators, and number of pauses. Multiple sub-vectors can be constructed or weights can be introduced to highlight the importance of specific dimensions according to the comparison purpose.

8. The experimental method for data acquisition and modeling of immersive evacuation lighting response with controllable variables according to claim 1, characterized in that, Step S6 specifically includes: Unsupervised clustering algorithms are used to cluster multidimensional response feature vectors to classify different response pattern categories; and / or, supervised classification algorithms are used to construct classification prediction models to identify or predict the response types of subjects. Deep learning methods are used to train and model response feature vectors or time series data to predict user evacuation behavior.

9. The experimental method for data acquisition and modeling of immersive evacuation lighting response with controllable variables according to claim 1, characterized in that, Step S7 specifically includes: A reinforcement learning simulation environment is constructed, which maps the immersive evacuation scene and the current lighting control scheme into the state space and action space of reinforcement learning. The state space includes the position information, orientation, and state of the indicator lights in the environment of the agent, and the action space includes the selection of the next movement direction. A reward function associated with the evacuation target is set, and a deep reinforcement learning algorithm is used to train the agent based on the reward function, so that it gradually learns the optimal evacuation strategy under different lighting schemes. Using the final agent policy, under a given lighting scheme, the evacuation trajectories of a large number of virtual individuals are simulated, and the lighting scheme is evaluated and iteratively optimized.

Images