Virtual reality immersion prediction method based on constraint perception depth regression

Abstract

The invention discloses a virtual reality immersion prediction method based on constraint perception depth regression, and belongs to the technical field of virtual reality display evaluation and human-computer interaction. In order to solve the problems that existing immersion assessment lacks a unified quantitative standard and is weak in cross-device generalization ability, the invention provides a constraint perception depth regression framework. According to the method, firstly, physical boundary constraint is applied to display parameters, and the display parameters are coupled with electroencephalogram Beta frequency band power to construct six-dimensional features; carrying out the statistics of boundary constraint input features through a training domain, and carrying out the modeling through a lightweight multi-layer perceptron regression network; and finally, combining feature importance analysis and a constrained single-parameter search mechanism to form a'prediction-interpretation-parameter adjustment 'integrated scheme. According to the method, the immersion prediction precision and the cross-device robustness can be effectively improved, and executable engineering suggestions are provided for parameter optimization of a virtual reality display system.

Description

Technical Field

[0001] This invention belongs to the field of display evaluation and human-computer interaction technology in Virtual Reality (VR) and Metaverse applications. Specifically, it relates to a virtual reality immersion prediction method and its application system based on constrained perception depth regression. In particular, it relates to a technical solution for quantitatively modeling and predicting immersion experience in virtual reality head-mounted displays and other terminals by comprehensively utilizing display optics and imaging parameters as well as user EEG signals. This can be categorized into the interdisciplinary fields of virtual reality display technology, human factors engineering evaluation, physiological signal perception, and intelligent optimization control. Background Technology

[0002] Currently, immersion, as an important technical indicator for measuring the quality of virtual reality (VR) system experiences, has become a key focus in this field. Existing technologies for evaluating and modeling immersive experiences can be broadly categorized into the following technical solutions: subjective assessment techniques based on questionnaires, objective modeling techniques based on physiological and behavioral signals, and comprehensive evaluation techniques that integrate subjective and objective information.

[0003] The first category comprises subjective assessment techniques based on questionnaire scales. Existing technologies have developed various scale systems for evaluating presence and usability, such as presence questionnaires, presence questionnaires, and system usability scales. Some literature has systematically analyzed the dimensional composition and reliability and validity of different presence scales, while other technical solutions have localized and psychometrically validated certain scales, demonstrating their applicability in dimensions such as spatial presence, involvement, and realism. These technical solutions share the common characteristics of being relatively simple to implement, having low deployment costs, and being easy to use in laboratory environments or small-scale user tests. However, these subjective assessment techniques generally rely on user self-reporting, are easily affected by psychological states, environmental factors, and experimental procedures, and are difficult to continuously quantify and model immersive experiences, nor can they directly support online parameter tuning of virtual reality devices during operation.

[0004] The second category comprises objective modeling techniques based on physiological and behavioral signals. In this category, physiological and behavioral signals are used to characterize the user's immersion state, with electroencephalography (EEG) signals being one of the most frequently used modalities. Existing solutions simultaneously collect EEG and other physiological signals in virtual reality or online learning scenarios, extracting features such as power spectral density and sample entropy. Traditional machine learning models, such as support vector machines, are then used to identify multi-class immersion levels or task states, validating the feasibility of predicting immersion levels based on physiological signals. Other solutions in virtual reality learning environments fuse multimodal behavioral features such as interaction logs, clickstreams, gaze trajectories, and facial expressions. An ensemble model is used to identify learning focus levels, demonstrating that multimodal fusion outperforms single-modal features and shows a certain correlation between perceived immersion and focus, as well as learning performance. While these objective modeling techniques have positive implications for utilizing physiological and behavioral signals, most still revolve around the chain of "scene or task—physiological response—subjective rating," with less attention paid to explicit joint modeling of virtual reality display parameters and neural signals. This makes it difficult to directly provide executable technical basis for the design and adjustment of display parameters in virtual reality devices.

[0005] The third category comprises comprehensive assessment technologies that integrate subjective and objective information. Some existing technologies attempt to jointly model immersive experiences using both subjective ratings and physiological signals. For example, in immersive virtual reality scenes, EEG signals are continuously recorded and combined with continuous emotional arousal scores, using deep learning structures such as Long Short-Term Memory (LSTM) networks to decode different arousal states; in virtual reality scenes with different levels of realism, EEG signals, presence scales, and task load scales are simultaneously collected, and the correlation between power changes in multiple frequency bands and subjective ratings is analyzed to reveal the comprehensive impact of visual realism on presence, cognitive load, and neural activity patterns. The above technical solutions provide a neurophysiological basis for immersive assessment based on the integration of subjective and objective information from EEG signals, but overall they still focus on correlation analysis and classification, with less attention paid to "immersion regression prediction for specific device configurations" and "output formats that can directly drive display parameter adjustments."

[0006] With the improvement of wearable sensors and graphics processing capabilities, deep learning has gradually become an important technical path for perception and prediction in virtual reality. Existing technologies, on the one hand, utilize structures such as convolutional neural networks, recurrent neural networks, and Transformers to automatically learn features and perform temporal modeling on representations such as power spectra, time-frequency graphs, and time series, overcoming the limitations of traditional support vector machines and random forests, which rely on handcrafted features and struggle to characterize high-dimensional time-varying signals. On the other hand, various multimodal deep fusion architectures have been proposed, mapping visual frames, head movements, and physiological signals to a unified feature space. Through modules such as attention mechanisms, modal weights are adaptively allocated, achieving good performance in classification or regression tasks such as dizziness level and comfort. To enhance generalization capabilities across scenarios and devices, some technical solutions introduce domain adaptation and distribution alignment mechanisms to reduce distribution differences between different content through adversarial loss. In addition, some technical solutions aim for "lightweight and interpretability," first using recurrent networks or multilayer perceptrons to build a dizziness detection model, and then combining feature importance analysis methods to screen key features. This allows for the reduction of model size and computational overhead while maintaining high classification accuracy, demonstrating the application potential of deep models in real-time monitoring.

[0007] Despite the significant progress made by the aforementioned deep learning technologies in automatic feature learning, multimodal fusion, and cross-scene modeling, the following common problems still exist: First, existing work mostly focuses on classification or level prediction tasks such as "dizziness / comfort," and the display configuration of virtual reality devices usually exists as latent variables, with fewer explicit inclusion of display parameters such as field of view, brightness, resolution, geometric distortion, and contrast in the modeling process; Second, existing models usually do not clearly characterize the physical boundaries of display parameters and lack explicit constraints on the distribution range of training data, making it difficult to guarantee prediction stability and engineering feasibility in input scenarios outside the training domain; Third, there is a relative lack of technical solutions that simultaneously address "immersion regression prediction based on EEG signals" and "outputting executable display parameter adjustment suggestions under engineering constraints" within a unified framework.

[0008] From an engineering perspective, the parameter configuration of a virtual reality system needs to simultaneously meet multiple constraints, including the physical limits of display devices, power consumption budget, and user comfort. Existing technologies in this field construct imaging models for head-mounted displays through optical simulation, formalizing indicators such as luminous efficacy, color gamut coverage, and brightness uniformity as objective functions. Under physical constraints such as pixel size and human eye resolution, parameter combinations are searched to provide feasible design boundaries for the display system. Other technologies dynamically limit the peripheral field of view or adjust optical flow intensity, redefining the usable field of view within physiologically acceptable limits to control user discomfort. Regarding constraint optimization, some research models brightness settings during immersive video playback as a time-series optimization problem with power consumption budget and contrast loss constraints. Given total power consumption and image quality thresholds, the brightness trajectory is solved, and the effect of balancing image quality and energy consumption is verified on actual devices. In terms of runtime parameter tuning, some technologies predict the degree of motion sickness based on head motion data and adjust rendering parameters accordingly, constructing a closed-loop system of "sensor signal—discomfort prediction—rendering parameter adjustment." Clear step sizes and boundaries are set for adjustment amplitude and update frequency to balance system responsiveness and stability.

[0009] In summary, while existing technologies have achieved significant progress in areas such as subjective scale construction, objective modeling based on EEG signals, multimodal deep learning, physical constraint modeling of display parameters, and multi-objective optimization of image quality and power consumption, they still have at least the following shortcomings:

[0010] 1. Most technical solutions model immersive experience as a relationship of "scene or task - physiological response - subjective rating", failing to explicitly co-model virtual reality display parameters and physiological signals such as EEG in a unified model, making it difficult to directly serve the engineering design and adjustment of display parameters;

[0011] 2. Existing deep learning models generally lack clear constraints on the physical feasible range of displayed parameters and the distribution range of training data. The models are not robust enough when dealing with inputs outside the training domain, and the prediction results lack reliability in engineering.

[0012] 3. Existing research usually separates the "prediction of immersion or discomfort level" from the "optimization of display parameters under constraints". There is still a lack of an integrated technical framework that tightly couples physiological signal-driven immersion regression prediction, training domain constraints and executable parameter tuning strategies in a unified feature space, and has lightweight features.

[0013] Therefore, under the above-mentioned technical background, there is an urgent need in this field for a new virtual reality immersion prediction technology that can perform immersion modeling by combining virtual reality display parameters and EEG signals under physical constraints and training domain constraints, and on this basis, provide technical support with engineering feasibility for parameter configuration and optimization of virtual reality display systems. Summary of the Invention

[0014] Purpose of the invention

[0015] In existing technologies, while various technical approaches have been developed for the evaluation and modeling of virtual reality immersive experiences, including subjective scale assessment, objective modeling of physiological and behavioral signals, and subjective-objective fusion, and techniques such as deep learning, multimodal fusion, and constraint optimization have been introduced, the following technical gaps still exist: First, there is a lack of a mechanism for explicit joint modeling of virtual reality display parameters and physiological signals such as EEG within a unified model, making it difficult to closely correlate "device configuration" with "immersion prediction"; second, there is a lack of explicit constraints on the physical feasibility range of display parameters and the distribution range of training data, resulting in a lack of engineering reliability in the prediction results of deep regression models under inputs outside the training domain; third, there is a lack of lightweight technical solutions that can transform immersion prediction results into executable display parameter adjustment suggestions that meet engineering constraints within the same framework.

[0016] Against this technical background, the purpose of this invention is to address the aforementioned problems by providing a virtual reality immersion prediction method based on constrained perception depth regression. This method constructs a compact coupled feature representation between virtual reality display parameters and EEG Beta band power, introduces physical boundary constraints on display parameters and training domain feature constraints, and combines feature importance analysis and a constrained parameter search mechanism to achieve the following specific technical objectives:

[0017] 1. Within a unified feature space, virtual reality display parameters and physiological signals such as EEG Beta band power are explicitly fused and modeled to establish an immersion quantitative prediction path based on device configuration, thereby achieving immersion regression prediction for specific device configurations.

[0018] 2. By setting physical boundary constraints on display parameters and applying projection and pruning constraints based on training domain statistics to input features, the deep regression model can learn and infer within physically feasible and data-supported feature regions, thereby improving the model's predictive stability and engineering feasibility under input conditions outside the training domain.

[0019] 3. Based on the above-mentioned constraint-aware immersion prediction model, feature importance measurement and constrained parameter search strategy are introduced to map the prediction results into display parameter adjustment suggestions that meet the engineering step size and amplitude limits, forming an integrated technical solution of "immersion prediction - feature interpretation - parameter tuning", providing feasible technical support for parameter optimization of virtual reality display systems under constraints.

[0020] Technical solution

[0021] This invention proposes a virtual reality immersion prediction method based on constraint-aware depth regression. The method is based on experimental samples collected from multiple subjects under multiple scene conditions. Each sample corresponds to one VR interaction trial and includes information such as display parameters, Beta band EEG power, and subjective immersion scores. Display parameters include diagonal field of view (FOV), screen brightness (b), color saturation (s), geometric distortion (d), and display resolution (number of pixels). and contrast adjustment Neurophysiological indicators were expressed using absolute power in the Beta band, denoted by symbols. The subjective indicator is the subjective immersion rating given by the participants in the IPQ questionnaire, expressed using symbols. This indicates that the participant identification ID is recorded simultaneously for subsequent cross-participant assessment and grouping strategy design.

[0022] To ensure that model learning occurs within the physically realizable parameter domain, hard constraint ranges are introduced for each display parameter based on the specifications of the VR headset's optical and electronic components and the characteristics of human vision. And clamp the original observations. For the first... Display parameters have

[0023] (1)

[0024] in, The effective value after clamping. , These are the physical lower and upper limits of the parameter, respectively. For example, for FOV constraints... Brightness constrained by cd / m², saturation constraint at Distortion and contrast adjustments are limited to preset safety ranges; for resolution pixel count... Simultaneously setting a lower and upper limit is used to avoid numerical instability caused by extremely low or extremely high resolution. After processing by equation (1), a clean dataset normalized under physical constraints can be obtained, providing a foundation for the subsequent training of the constraint-aware regression network.

[0025] Based on this, to enhance the coupling between the display parameters and the Beta band EEG response, six [other parameters] were constructed. The relevant derived features form a compact input vector.

[0026] (2)

[0027] in, This serves as the input feature vector for the subsequent regression network. These correspond to different coupling forms between field of view, saturation, brightness, distortion, resolution, and contrast and Beta power, respectively.

[0028] Field of view - Beta feature Defined as

[0029] (3)

[0030] in, This is the result after converting the diagonal field of view (FOV) (in degrees) to radians. This represents the absolute power in the Beta band for the corresponding test. It approximates the change in the visible solid angle in front of the eyes, used to describe the field of view coverage; and The multiplication reflects the combined effect of "field of view × Beta activation".

[0031] Saturation-Beta Feature Defined as

[0032] (4)

[0033] in, The normalized color saturation (value range is...) ), The square root of the Beta power is used to suppress the influence of extremely high-power samples on the feature distribution. This feature expresses the coupling relationship of "color saturation level × Beta response," giving moderate weight to visual stimuli at moderate saturation in the feature space. Brightness-Beta Feature Defined as

[0034] (5)

[0035] in, To display the brightness of the scene, The brightness constant calibrated for VR headsets; The square of the Beta power is used to highlight the nonlinear enhancement effect of Beta activation in bright environments. This feature characterizes the combination of "normalized brightness × quadratic Beta term," which helps the model distinguish the intensity differences of immersion-related neural responses under different brightness states. Distortion-Beta Feature Defined as

[0036] (6)

[0037] in, This is the geometric distortion variable; the smaller its value, the closer the image is to the ideal optical image. This feature is used to convert "distortion reduction" into a positive contribution; exponents of 0.4 and 0.9 are used to balance the magnitudes of Beta power and distortion amplitude in the feature. By setting power exponents for Beta and distortion amplitude respectively, this feature makes the combination of "low distortion + high Beta activation" more easily distinguishable by the model in the feature space. Resolution-Beta Feature Defined as

[0038] (7)

[0039] in, To display the total number of pixels, A positive scaling constant is used to control the input scale of the logarithmic function. In this invention, it is taken as... , It is a natural logarithmic function. This feature, by applying a logarithmic transformation to the resolution, describes the phenomenon of "diminishing returns to experience with increasing resolution," and is derived from... Weighting is applied to reflect the contribution of the "high resolution + high beta activation" combination to immersion, while suppressing the numerical range caused by extreme high resolution configurations.

[0040] Contrast - Beta Feature Defined as

[0041] (8)

[0042] in, Contrast adjustment amount (usually a decimal around 0, representing the standard) Offset on the curve). The gamma value obtained from its mapping, This is Beta power. This feature converts contrast adjustment to an approximation. The corrected intensity change, and then with Multiplication is used to describe the modulating effect of contrast modulation on immersion-related neural responses.

[0043] Based on the above construction, a six-dimensional Beta-coupled feature vector has been constructed based on the displayed parameters and the Beta band power.

[0044] (9)

[0045] And obtain the corresponding subjective immersion score. This invention builds upon this foundation by constructing a Constraint-Aware Deep Regression Method for Virtual Reality ImmersionPerception Prediction (CADL-IPP), achieving a non-linear mapping from Beta-coupled features to immersion ratings, and improving the model's robustness and controllability across device scenarios through a training domain projection mechanism. First, the samples in the training set are denoted as...

[0046] (10)

[0047] in, Let be the number of samples. To avoid extreme inputs that deviate far from the training distribution during the inference phase, a training domain constraint set is constructed in the feature space of the training set.

[0048] (11)

[0049] For any feature vector to be evaluated Define the training domain projection operator that clamps each dimension one by one.

[0050] (12)

[0051] And let the projected vector be... Furthermore, the six-dimensional features are standardized according to the statistics of the training set to eliminate dimensional differences. At this point, the features entering the regression network are all restricted to the feasible region obtained from the statistics of the training data. Inside.

[0052] By combining the physical boundary clipping of the display parameters, CADL-IPP applies constraints on both the "input parameter space" and the "Beta-coupled feature space" layers: on the one hand, it ensures that the parameter combination meets the physical specifications of the device, and on the other hand, it suppresses abnormal features that are far from the training experience distribution, thereby reducing the risk of unstable predictions when deployed across devices and configurations.

[0053] In terms of network architecture, CADL-IPP employs a lightweight multilayer perceptron regression model to achieve... Nonlinear mapping to immersion rating

[0054] (13)

[0055] in, For parameters feedforward neural networks For the first An estimated value of the immersion score for each sample. Specific implementation details.

[0056] In the middle, as input layer features The transformation is performed sequentially through three fully connected hidden layers:

[0057]

[0058] in, The first Layer weight matrix and bias vector, and For output layer parameters; This represents the LayerNorm normalization operator. The GELU activation function is used. To improve generalization performance, a Dropout operation is added after the hidden layer to randomly zero out the outputs of some neurons. The width of the hidden layer is set to... Its overall parameter scale is relatively small, making it easy to deploy on near-eye display devices with limited computing power.

[0059] The network training uses the minimum mean squared error (MSE) as the loss function. Let... Then in the parameters The following losses are

[0060] (15)

[0061] During optimization, a subset of samples from the training set is first allocated as a validation set to monitor model convergence. Parameter updates employ the AdamW optimization algorithm, combined with cosine annealing learning rate scheduling and gradient pruning strategies to suppress oscillations and overfitting risks in the later stages of training. Early stopping is triggered when the validation set MSE no longer decreases within several consecutive iterations, yielding the final model parameters. During the inference phase, given any set of display parameters satisfying physical hard constraints and their corresponding Beta-coupled features, the results are obtained via projection from the training domain. Afterwards, the security prediction score output by CADL-IPP can be recorded as:

[0062] (16)

[0063] This process unifies physical feasibility constraints and training domain statistical boundaries into the same prediction chain, enabling the resulting regression model to maintain good fitting accuracy while exhibiting stronger controllability and robustness to cross-device deployment and extreme parameter inputs. This provides a stable prediction foundation for subsequent feature contribution analysis and engineering-side parameter tuning scheme generation.

[0064] To systematically evaluate the performance of constraint-aware deep regression networks in immersion prediction tasks and quantify the contribution of six-dimensional beta-coupled features to subjective ratings, this invention constructs a multi-level evaluation system based on unified data partitioning and model structure, focusing on four aspects: regression accuracy, level recognition ability, extreme sample discrimination ability, and permutation feature importance. First, the test set samples are used... For the evaluation object, among which For the first The six-dimensional Beta-coupled feature vector of each sample. For the corresponding subjective immersion rating, This represents the number of samples in the test set. We then use the optimal regression model obtained in the previous section. Let its prediction on the test set be denoted as .

[0065] (17)

[0066] For evaluating regression performance, mean squared error (MSE), mean absolute error (MAE), and coefficient of determination were used. Three indicators:

[0067]

[0068] in, The mean of the true scores on the test set. MSE and MAE are used to characterize the average magnitude of the prediction error. These three metrics are used to measure the model's ability to explain the variance of the scores, and together they reflect the fitting accuracy and robustness of CADL-IPP on unseen data.

[0069] In engineering applications, immersion is often incorporated into configuration decisions in a hierarchical manner. Therefore, continuous scoring is used. Further discretization is used to construct classification labels. Assume the scores from the training set are predetermined. Level division points

[0070] (twenty one)

[0071] Then the first Grade labels for each sample Defined as

[0072] (twenty two)

[0073] Based on feature vectors With tags The gradient boosting tree ranking recognition model was trained using GroupKFold cross-validation with subjects grouped together, and the predicted label was denoted as... Classification performance is primarily measured by accuracy (Accuracy, Acc) and weighted F1 score (...). ) measurement, where

[0074] (twenty three)

[0075] (twenty four)

[0076] In the formula, This is an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise. Let k be the number of samples in the k-th class. and The first Precision and recall are calculated for each level. Acc reflects the overall level recognition accuracy. It pays more attention to the balance between different levels, and the combination of the two can evaluate the model's ability to distinguish levels under different granularities of 2 / 3 / 5 categories.

[0077] Considering that the separability between "extremely high immersion" and "extremely low immersion" samples is particularly important in practical deployment scenarios, this invention additionally constructs an extreme sample subset on the test set. Let a given quantile parameter... Record the score below and above The quantiles are respectively and Define the low-rated sample index set and the high-rated sample index set as follows:

[0078] (25)

[0079] In the sample set The above approach treats high-scoring samples as positive and low-scoring samples as negative, and can be based on continuous predicted values. The area under the ROC curve (AUC) and the average precision (AP) under the PR curve are calculated to characterize the model's discrimination ability in the binary distinction between "extremely high immersion" and "extremely low immersion". Furthermore, to evaluate the model's performance on the ranking task, pairwise ranking precision between high and low extreme samples is defined.

[0080] (26)

[0081] in, and These represent the number of extreme samples with high and low scores, respectively. PairAcc indicates the proportion of all "high-low" sample pairs where the model can consistently provide a ranking where "high score is greater than low score," used to evaluate the reliability of CADL-IPP in ranking extreme scenarios.

[0082] In terms of feature contribution analysis, this invention employs a permutation-based feature importance metric to measure the relative contribution of six-dimensional Beta-coupled features to the regression task. Let the baseline mean square error of the optimal regression model on the test set, while keeping the features constant, be... For the first Let there be _ ... Then the first The importance increment of each feature is defined as

[0083] (27)

[0084] Based on this, the relative weights are obtained by normalizing the importance increments of all features.

[0085] (28)

[0086] in, The larger the value, the more likely it is to disrupt the order. The more significant the impact of a feature on the model error, the more likely it is to have a corresponding effect. The larger the value, the more critical the feature is in immersion prediction. This is achieved through... Visualizations such as radar charts allow for a direct comparison of the relative contributions of derived features such as field of view, brightness, color saturation, darkness variation, resolution, and contrast to the overall decision-making process.

[0087] After training the constraint-aware regression network, the model output needs to be transformed into actionable parameter tuning suggestions for specific VR devices, thus forming a complete closed loop of "immersion prediction - display parameter optimization". This invention, based on physical hard constraints and training domain constraints, constructs a single-parameter search strategy for individual device configurations. It evaluates the gain of each of the six display parameters—view angle, brightness, saturation, distortion, resolution, and contrast—on the immersion score and provides an upgrade scheme that meets the principle of "high immersion - low cost". First, the display parameters of a specific VR device are denoted as a vector.

[0088] (29)

[0089] in, These represent the diagonal field of view (FOV) and brightness of the device, respectively. saturation , distorted variables Total number of pixels and contrast adjustment When generating the parameter tuning scheme, the Beta band power uses a reference value. This represents the typical neural response level of the target user group and participates in feature construction as a constant. Based on the defined Beta-coupled derived features, [the following is used]... Mapped to a six-dimensional feature vector

[0090] (30)

[0091] in, This represents a nonlinear mapping from physical parameters and reference Beta power to a six-dimensional Beta-coupled feature. Subsequently, the given training domain projection operator is utilized. Will Projected onto the feature space training domain The results are then input into the optimal regression network. The security prediction score for this device configuration is obtained as follows:

[0092] (31)

[0093] in, For combined mapping functions, To predict immersion scores under physical hard constraints and training domain constraints, These are the network parameters after early training stop. The current factory settings of the device are recorded as follows: Its corresponding baseline prediction score is

[0094] (32)

[0095] For each adjustable parameter This requires searching for candidate target values ​​within the engineering feasibility range and evaluating their gain on immersion scoring. To this end, within the physically hard constraint range... Based on this, an engineering-side "upgradeable range" limit is introduced. Let the allowed absolute adjustment range and relative adjustment ratio in engineering be respectively... and For parameters optimized in the "improvement" direction (such as FOV, brightness, saturation, resolution, and contrast), the search range is defined as follows:

[0096] (33)

[0097] in, The lower bound is fixed at the current value. The upper bound is determined by the physical limit. Absolute upgrade range Relative upgrade range All three factors are constrained, and their minimum value is taken; for parameters optimized in the "reduction" direction (such as distorted variables) The search interval is defined as follows:

[0098] (34)

[0099] in, Fixed to the current value , Physical lower limit Absolute decrease Relative decline The maximum value among the three. This is determined by... It will neither exceed the physically realizable boundary nor the engineering-acceptable single adjustment step size. After obtaining the adjustable range of a single parameter, for the ... A set of discrete candidate values ​​is constructed from the parameters.

[0100] (35)

[0101] in, The number of sampling points. In the interval The first uniformly selected There are 10 candidate values. For each candidate value... Construction only in the first Parameter vector that changes in each dimension

[0102] (36)

[0103] And through the security prediction function Calculate the corresponding predicted score

[0104] (37)

[0105] Define the scoring gain relative to the baseline configuration as

[0106] (38)

[0107] To provide a recommended target value for each parameter, in the candidate set Find the candidate point with the largest score gain.

[0108] (39)

[0109] in, For the first Recommended values ​​for each parameter. This represents the corresponding maximum score gain. To avoid providing ineffective upgrades with limited benefits but high costs, a gain threshold is introduced. :when At that time, the judgment of the first The parameter is "not recommended to be adjusted", meaning it should be retained. Unchanged; when At that time, with As a candidate upgrade scheme for this parameter.

[0110] For all After completing the single-parameter constraint search, parameter tuning suggestions sorted by score gain can be obtained. These suggestions include parameter names, baseline values, recommended values, and corresponding predicted scores, facilitating the prioritization of tuning terms with the largest score gain under different hardware conditions and cost constraints. Furthermore, the potential synergistic effect of multi-parameter combination upgrades can be considered as an extension of the single-parameter recommendations, and can be applied sequentially. Evaluate the predicted score sequences on different upgrade paths to provide a quantitative basis for the trade-off between "upgrade magnitude - immersion benefits".

[0111] Technical effect

[0112] Corresponding to the aforementioned technical problems and solutions, this invention, through comprehensive design in areas such as physical constraints of display parameters, feature constraints of the training domain, construction of Beta-coupled features, and constraint-aware deep regression and parameter tuning strategies, can produce the following substantial technical effects. These technical effects are all directly derived from the technical features of this invention and have been verified in experiments involving multiple subjects and multiple scenarios.

[0113] First, by introducing physical boundary constraints on the display parameters and adopting the clamping strategy described in Equation (1), all combinations of display parameters involved in modeling and inference are limited to the feasible range allowed by the device's optics and electronics. On the one hand, this avoids the problem of traditional unconstrained regression models learning and extrapolating within physically unrealizable parameter regions, eliminating a potential problem of "predicted results being numerically reasonable but not practically feasible in engineering" from the source. On the other hand, in actual deployment, when the input parameters approach extreme operating points or deviate due to device differences, this invention can still project them back to the physically feasible region, ensuring the engineering feasibility and security of the prediction link throughout the entire process, thereby improving the stability of the virtual reality system under different devices and configurations.

[0114] Secondly, through the six-dimensional Beta-coupled derived features defined by equations (2) to (8), this invention maps multidimensional display parameters and absolute power of the EEG Beta band into a compact feature space, realizing an explicit coupling representation between "field of view—brightness—saturation—distortion—resolution—contrast" and neural response. Compared to modeling directly on the original parameters and power values, this construction effectively compresses the feature dimension while maintaining immersion-related information, reducing model complexity and enabling a lightweight feedforward network to achieve high fitting ability. At the same time, each derived feature has a clear physical and physiological interpretation, which is beneficial for subsequent feature importance analysis and for engineers to understand the model decision basis. In the comparative experiment, compared with the baseline method that did not introduce Beta-coupled derived features, this invention showed better predictive performance in terms of regression error and coefficient of determination on the test set, indicating that this compact feature representation can significantly improve the effectiveness of immersion modeling.

[0115] Furthermore, the training domain constraint set Ω defined by equations (11) and (12) D By employing its projection operator, this invention explicitly characterizes the feasible region supported by training data at the feature level. It performs dimensionality-based pruning and standardization on the Beta-coupled feature vectors input at all inference stages, restricting them to an empirical distribution range. This design effectively suppresses the unstable outward projection behavior of traditional deep regression models when faced with extreme inputs outside the training domain, making the model output insensitive to abnormal inputs when deployed across devices and scenarios. Comparative results in multi-subject cross-validation and extreme high / low immersion sample discrimination tasks demonstrate that this invention maintains high stability in metrics such as "extremely high vs. extremely low immersion" binary discrimination and pairwise ranking accuracy while preserving overall fitting accuracy, reflecting the robustness improvement brought about by training domain constraints.

[0116] Furthermore, by constructing a lightweight multilayer perceptron regression network using equations (13) to (16) and employing LayerNorm normalization, GELU activation, and Dropout regularization, this invention achieves nonlinear fitting of subjective immersion ratings under limited model size. On the one hand, the network structure is shallow and narrow with fewer parameters, making it suitable for deployment in near-eye display terminals or embedded platforms with limited computing and storage resources. On the other hand, with training strategies such as MSE loss, AdamW optimization, and learning rate scheduling, a convergent and stable regression model can be obtained under limited samples and across subject scenarios. Experimental results show that, compared with various unconstrained deep regression baseline models, this invention improves performance metrics such as prediction error, coefficient of determination, and weighted F1 at different rating levels to varying degrees, and maintains good generalization ability under cross-validation settings grouped by subject.

[0117] Furthermore, through the permutation-based feature importance analysis defined by equations (27) and (28), this invention can quantify the relative contribution of six-dimensional Beta-coupled features to immersion scoring at the model level, realizing a visual characterization of the role of "field of view—brightness—saturation—distortion—resolution—contrast" in immersion prediction. This technical feature makes the model no longer a "black box," but can clearly indicate which parameter adjustments are more conducive to improving immersion, providing an interpretable decision basis for subsequent parameter optimization and hardware upgrades, thereby significantly improving the engineering usability of the model while ensuring prediction performance.

[0118] Finally, through the single-parameter constraint search and scoring gain estimation mechanism described in equations (29) to (39), this invention achieves a unified mapping function g. The immersion prediction results are further transformed into display parameter adjustment suggestions that meet both physical hard constraints and engineering range constraints, achieving a closed-loop integration of prediction, interpretation, and parameter tuning. Specifically, by superimposing absolute and relative adjustment range limits on the physical range of each parameter, and discretely sampling and evaluating the prediction score gain within this constraint range, this invention can automatically select the parameter change with the maximum immersion benefit under the current constraints, and filter out adjustment schemes with insufficient benefits through a gain threshold mechanism. This technical feature allows virtual reality systems to directly obtain engineering-executable parameter tuning paths based on model output in actual deployment, avoiding large-scale blind searches and repeated trial and error, significantly reducing the experimental costs of parameter configuration and upgrades, and improving the efficiency of iterative optimization on the device side.

[0119] In summary, this invention achieves high-quality prediction of virtual reality immersion and executable optimization of display parameters through a collaborative design of physical boundary constraints, training domain constraints, Beta-coupled compact features, lightweight deep regression, permutation feature importance, and constrained single-parameter search, while ensuring engineering feasibility and model interpretability. Compared with existing technologies, it has significant advantages in prediction accuracy, robustness across subjects and devices, and engineering parameter tuning feasibility. Attached Figure Description

[0120] Figure 1 This is the overall process structure of the virtual reality immersion prediction method of the present invention. Detailed Implementation

[0121] The following is in conjunction with the appendix Figure 1 The technical solution of the present invention will be further described below with reference to specific embodiments.

[0122] like Figure 1As shown, the overall process of the method of the present invention includes: First, acquiring the field of view, brightness, resolution, and other display parameters from the virtual reality device, while simultaneously collecting the raw EEG data and subjective questionnaire scores for the corresponding trials. The aforementioned display parameters and EEG data are input into the feature engineering module, where they are normalized, combined, and extracted to form various feature indices corresponding to the field of view, brightness, resolution, etc.

[0123] Then, these feature indices and subjective questionnaire scores are used as samples to constrain the perceptual regression network for modeling and training. During the training phase, methods such as N-fold cross-validation can be employed to optimize network parameters through multiple training and validation processes, resulting in a stable prediction model. Finally, the feature indices of the scene to be evaluated are input into the trained network to output the corresponding immersion prediction results.

[0124] The following detailed explanation of each step of the above process is based on a specific embodiment.

[0125] 1. System Construction and Data Acquisition

[0126] In this embodiment, a virtual reality immersion prediction system is constructed for verifying and applying the method of the present invention. The system includes a head-mounted VR display device, an EEG acquisition device, and a computer workstation for running the algorithm of the present invention. The head-mounted VR display device is an HTC Vive Pro Eye, with configurable display parameters such as field of view, brightness, resolution, and contrast. The EEG acquisition device is a MUSE 2 EEG acquisition device, used to collect the user's brain signals during virtual reality interaction. The computer workstation runs data acquisition and modeling software, connects to the HTC Vive Pro Eye via a video interface, and connects to the MUSE 2 via Bluetooth or a wired connection, thereby synchronously recording display parameters, EEG signals, and subjective immersion scores in each VR interaction trial, and executing the immersion prediction and parameter optimization method described in this invention.

[0127] In this embodiment, 36 healthy participants were selected to participate in the experiment, covering different genders and levels of virtual reality experience. Each participant wore a VR headset and an EEG device, and experienced several virtual scenes in sequence. The scene types could include virtual roaming scenes, interactive task scenes, and immersive video scenes. For each scene, multiple independent trials were set for each participant, and each participant completed 10-15 interactive trials in each scene. Before each VR interactive trial, a set of display parameter configurations was sent to the VR device through the workstation, including the diagonal field of view (FOV), screen brightness (b), color saturation (s), geometric distortion (d), and display resolution (pixels R). px and contrast adjustment amount c ctrDuring the trial, the EEG acquisition device continuously collected EEG signals and extracted the absolute power β of the Beta band within a predetermined time window. After the trial, the subject completed an immersion questionnaire to obtain the subjective immersion score y for that trial, and the subject's ID was written into the data record. After removing invalid data such as obvious motion artifacts and recording interruptions, this embodiment ultimately obtained approximately 800-1000 valid samples, each containing a set of display parameter vectors (FOV). The corresponding Beta band power β, subjective immersion score y, and participant identification .

[0128] 2. Physical constraint handling

[0129] After data acquisition, this embodiment first applies physical constraints to the samples at the display parameter level. Based on the technical specifications of the HTC Vive Pro Eye head-mounted display device and the human eye's visual comfort range, this embodiment sets corresponding physical lower limits for field of view, brightness, saturation, geometric distortion, resolution pixels, and contrast adjustment. and upper limit For example: the field of view (FOV) is limited to between the minimum and maximum diagonal FOV supported by the device; the brightness (b) is limited to between the minimum and maximum brightness nominally provided by the device; the color saturation (s) is limited to the normalized effective range of [0, 1]; the geometric distortion (d) is limited to a predetermined distortion compensation range; the resolution (pixels) (Rpx) is limited to between the minimum and maximum rendering resolution supported by the device; and the contrast adjustment (c) is limited to... ctr The parameters are then restricted within a preset adjustment range. Subsequently, according to formula (1) in the instruction manual, the displayed parameters for each sample are clamped, and the original observed values ​​are... Mapped to physically feasible valid values This process yields a set of normally regulated display parameters under physical constraints, providing clean input for subsequent Beta coupling feature construction and regression model training.

[0130] 3. Construction of Six-Dimensional Beta-Coupling Features

[0131] After completing the physical constraint processing, this embodiment constructs a six-dimensional Beta-coupled derived feature vector for each sample based on the physical clipping display parameters and the corresponding Beta band absolute power β, according to the aforementioned equations (2) to (8) in the specification. .in, Calculations are performed after converting the diagonal field of view (FOV) to radians. Then multiply by β to characterize the coupling effect between field coverage and Beta activation; The normalized saturation s is combined with the square root of the Beta power to describe the relationship between color saturation and neural response. The brightness b is set according to the equipment calibration constant. After normalization, it is multiplied by the square of the Beta power to highlight the nonlinear enhancement effect in high-brightness environments; By applying power function transformations to the geometric distortion d and Beta power respectively, and mapping "distortion reduction" to a positive contribution, the combination of "low distortion + high Beta activation" is more easily distinguished in the feature space. For resolution and number of pixels Logarithmic transformation combined with β-weighting is used to characterize the diminishing immersive benefits of resolution improvement; Adjust the contrast Mapped to gamma value This is then multiplied by β to characterize the modulating effect of contrast adjustment on immersion-related neural responses. Thus, in this embodiment, each sample can be represented as a pair of... ,in It is a six-dimensional Beta-coupling feature. For the corresponding subjective immersion rating, the sample set can be represented as the dataset defined in equations (9) and (10) above in the instruction manual. .

[0132] 4. Applying constraints to the training domain and preprocessing features

[0133] After feature construction is completed, this embodiment further applies training domain constraints at the feature level to improve the model's robustness to inputs outside the training domain. Specifically, based on all training samples... In this embodiment, the minimum value of each feature dimension in the training set is calculated according to the aforementioned formula (11) in the specification. With the maximum value Construct the training domain constraint set During the model training phase, for each input feature vector, the original data is first used directly based on the identity of the sample it belongs to in the training or validation partition. And its standardized form. During the model inference phase, for any Beta-coupled feature of a sample to be predicted... In this embodiment, the feature values ​​of each dimension are restricted to the corresponding range according to the aforementioned formula (12) in the specification. Within, the projected vector is obtained. Furthermore, the features are standardized according to the training set statistics, ensuring that the features entering the regression network fall within both the physically feasible region and the empirical distribution supported by the training data. By introducing constraints at both the explicit parameter level and the feature level, this embodiment effectively suppresses unstable outreach behavior that may occur under cross-device deployments or extreme configurations.

[0134] 5. Network Structure, Training, and Inference

[0135] Regarding network structure and training, this embodiment uses the lightweight multilayer perceptron regression network defined in equations (13) to (16) above in the specification as the immersion prediction model. Specifically, the six-dimensional feature vector projected and standardized from the training domain is used as the input layer. The network passes through three fully connected hidden layers, each followed by LayerNorm normalization and GELU activation. A Dropout operation is applied after the hidden layer output to suppress overfitting. In this embodiment, the widths of the three hidden layers can be set to 128, 64, and 32 respectively, resulting in a moderate overall network parameter scale, facilitating deployment on GPU workstations or subsequent embedded terminals. The output layer is a one-dimensional linear unit, outputting the immersion prediction score for the corresponding sample. During training, all collected samples are split into training, validation, and test sets according to the subject division principle. A five-fold cross-validation (GroupKFold) strategy is adopted to ensure that the data of the same subject only appears in either the training or test set, so as to guarantee cross-subject generalization performance. The loss function adopts the mean squared error (MSE) defined in the specification (15). The network parameters are iteratively updated by the AdamW optimization algorithm, and the cosine annealing learning rate scheduling and gradient clipping are combined to suppress the oscillation in the later stage of training. When the validation set loss no longer decreases significantly within a predetermined number of rounds, early stopping is triggered to obtain the final model parameters. In the inference stage, for any input sample including display parameters and Beta band power, this embodiment first processes it according to the above physical constraints and training domain projection to obtain a standardized feature vector. Then utilize the trained network Output safety prediction score This process is consistent with the aforementioned formula (16) in the instruction manual.

[0136] 6. Model Performance Validation

[0137] To verify the predictive performance and robustness of the model described in this embodiment, MSE, MAE, and R are calculated on the test set according to equations (17) to (20) above in the specification. 2The regression evaluation indexes are used, and the continuous scores are further discretized into several levels according to equations (21) to (24) in the specification. The gradient boosting tree level recognition model is trained, and the accuracy and weighted F1 value are evaluated. In terms of extreme sample discrimination, this embodiment constructs extreme subsets of high and low scores according to equations (25) and (26) in the specification, and calculates the area under the ROC curve (AUC), average precision of the PR curve (AP), and pairwise ranking precision (PairAcc) to evaluate the model's discrimination ability in the case of "extremely high immersion vs. extremely low immersion". Compared with several deep regression baseline models that do not introduce physical constraints, training domain constraints, or do not use Beta-coupling features, the model obtained in this embodiment shows stable results that are better than the baseline in terms of regression error on the test set, coefficient of determination, and level recognition performance. It also has good generalization consistency across subject divisions, indicating that the method of this invention can obtain high-quality immersion prediction performance under actual experimental conditions.

[0138] 7. Parameter Optimization Application Examples

[0139] After completing model training and validation, this embodiment also applies the immersion prediction model to optimize the display parameters of specific virtual reality devices. Taking a factory configuration of an HTC Vive Pro Eye device as an example, let its current display parameter vector be... Each component corresponds to FOV, brightness, saturation, distortion, resolution (number of pixels), and contrast adjustment, respectively. In this embodiment, the median Beta power of all subjects in the training dataset is selected as the Beta band power reference value β representing the average state of the target user group. ref According to the aforementioned equation (30) and the training domain projection operator in the instruction manual. Construct the Beta-coupled features of the baseline configuration and input them into the regression network to obtain the baseline predicted score for that configuration. , where g It is consistent with the definitions in formulas (31) and (32) of the instruction manual.

[0140] Subsequently, taking into account the physical limits of the equipment and the allowable adjustment range in engineering, this embodiment sets an upper limit for the absolute adjustment range for each display parameter dimension. and the upper limit of relative adjustment ratio And determine the single-parameter adjustable range of each parameter according to the aforementioned formulas (33) and (34) in the instruction manual. Within each adjustable interval, this embodiment uniformly samples a set of candidate values ​​according to equation (35). For each candidate value Construct a parameter vector that changes only in the k-th dimension. (Example (36)), and through combination mapping g Calculate the corresponding predicted score (Equation (37) in the instruction manual), and then calculate the score gain relative to the baseline configuration according to Equation (38). .

[0141] After calculating the score gain for all candidate points, this embodiment calculates the candidate set for each parameter according to the aforementioned formula (39) in the specification. Perform a search to find the recommendation value with the greatest rating gain. and its corresponding maximum gain Simultaneously, a preset gain threshold ε is set; when the maximum gain of a certain parameter... In this embodiment, it is considered that adjusting this parameter would yield limited benefits under the current constraints; therefore, it is not recommended to adjust this parameter and keep its original value. ;when When this happens, the recommended target value for this parameter is set to... After completing the single-parameter constraint search for all six display parameters, this embodiment can output a list of parameter tuning suggestions, including parameter names, current configuration values, recommended configuration values, and corresponding predicted score gains, arranged from highest to lowest score gain. Engineers can use this list to prioritize parameter adjustments that contribute most to improving immersion, within constraints such as hardware cost, power consumption budget, and user comfort. For example, in a practical application, this embodiment might provide recommended combinations such as "moderately increase FOV" and "slightly increase scene brightness and appropriately reduce distortion" to improve the predicted immersion score within the device's capabilities.

[0142] As can be seen from the above specific embodiments, the virtual reality immersion prediction method based on constrained perception depth regression described in this invention can be implemented on specific VR hardware and EEG acquisition platforms. It utilizes a limited number of multi-subject, multi-scene samples to complete model training and outputs engineering-executable display parameter optimization suggestions on real device configurations. Those skilled in the art can make equivalent substitutions or appropriate adjustments to the device model, number of subjects, scene type, number of network layers, and parameter settings without departing from the spirit and substance of this invention; all such substitutions should fall within the protection scope of this invention.

Claims

1. A method for predicting and optimizing virtual reality immersion based on constrained perception depth regression, characterized in that, Includes the following steps: (S1) Collect multiple display parameters, synchronous EEG beta frequency absolute power and subjective immersion score of the subject during the interaction process of virtual reality scene. The display parameters include field of view, screen brightness, color saturation, geometric distortion, total number of pixels of the displayed image and contrast adjustment. Physical boundary constraint processing is performed on each of the display parameters to limit its value to a preset physical feasible range. (S2) The various display parameters after physical constraint processing are nonlinearly combined with the power of the EEG Beta band to construct a six-dimensional beta coupling feature vector that reflects the six coupling relationships of field of view, brightness, color saturation, geometric distortion, resolution and contrast. (S3) Based on the training dataset, the empirical distribution boundary of the six-dimensional beta coupled feature vector is statistically determined, and a training domain constraint set is constructed. During model training and inference, all feature vectors are projected into the training domain constraint set and standardized. The feature vectors after projection and standardization and the corresponding subjective immersion scores are used to supervise the training of a constraint-aware deep regression network to obtain an immersion prediction model. (S4) For the target virtual reality device, its display parameters and the reference beta power representing the target user group are processed in the same way as in steps (S1) to (S3) and input into the immersion prediction model to obtain the immersion prediction score of the current configuration; a single parameter search is performed within the physical boundaries and engineering adjustment constraints of each display parameter to evaluate the gain of parameter adjustment on the immersion score and output parameter optimization suggestions.

2. The method according to claim 1, characterized in that, In step (S1), the physical boundary constraint processing specifically includes: The field of view is limited to between the minimum acceptable angle of view and the maximum achievable angle of view supported by the device; The screen brightness is limited to the safe brightness range of the display device; The color saturation is limited to a normalized effective range between zero and one. The geometric distortion is limited to a predetermined distortion compensation range; The total number of pixels in the displayed image is limited to between the minimum and maximum number of pixels supported by the device. The contrast adjustment amount is limited to a preset contrast adjustment range.

3. The method according to claim 1, characterized in that, In step (S2), the construction of the six-dimensional beta coupling feature vector is achieved through the following set of formulas: Field of view coupling characteristics: ; Color saturation coupling feature: ; brightness Coupling characteristics: ; Geometric distortion coupling characteristics: ; resolution Coupling characteristics: ; Contrast coupling characteristics: ,in ; in, The value in radians after the field of view is converted. The power of the EEG Beta band. The normalized color saturation, For screen brightness, This is the brightness calibration constant. For the geometric distortion, The total number of pixels in the displayed image. This is a resolution scaling constant. This refers to the contrast adjustment amount.

4. The method according to claim 1, characterized in that, In step (S3), the construction of the training domain constraint set and the projection processing specifically involve: Calculate the minimum value of each feature dimension of the six-dimensional beta-coupled feature vector in the training dataset. and maximum value This constitutes a six-dimensional bounded space. As the set of constraints for the training domain; For any input feature vector to be processed Clamping projection by dimension: ,in ; The projected vector Standardize according to the statistics of the training dataset.

5. The method according to claim 1, characterized in that, In step (S3), the constraint-aware deep regression network is a multilayer perceptron structure, which includes: The input layer is used to receive the six-dimensional standardized feature vector; There are at least three fully connected hidden layers, with the number of neurons in each hidden layer decreasing layer by layer. Layer normalization and GELU activation functions are set between adjacent hidden layers. The output layer is used to output the immersion prediction score; A random deactivation mechanism is introduced in the hidden layer; the network is trained with mean squared error as the loss function, using the AdamW optimization algorithm combined with a cosine annealing learning rate scheduling strategy, and training stops when the validation set loss no longer decreases within a preset number of rounds.

6. The method according to claim 1, characterized in that, It also includes a permutation-based feature importance analysis step, used to quantify the contribution of each feature in the six-dimensional beta-coupled feature vector to the prediction: Calculate the baseline prediction error of the immersion prediction model on the test set; The individual feature dimensions in the six-dimensional beta-coupled feature vector are randomly permuted in turn, and the prediction error is recalculated. The error difference before and after the replacement of each feature dimension is used as its importance index, and normalization is performed to obtain the relative contribution weight of each feature.

7. The method according to claim 1, characterized in that, In step (S4), the single-parameter search within the physical boundary and engineering adjustment constraints specifically includes the following sub-steps: (S4.1) Determine the search interval: For the first segment to be optimized... A display parameter, at its physical boundary Based on this, the upper limit of the absolute adjustment step size allowed in engineering is superimposed. With the upper limit of relative adjustment ratio Together, they determine the feasible search range for this parameter in the current optimization step. ; (S4.2) Discrete sampling and gain evaluation: within the search interval Uniform sampling is used to generate a set of candidate values; for each candidate value, a new configuration is constructed in which only that parameter changes while the other parameters remain at the baseline value, and its predicted score is calculated by the immersion prediction model, and the score gain compared to the baseline configuration is calculated. (S4.3) Generate optimization suggestions: In the candidate value set of each parameter, select the candidate value with the largest score gain as the recommended value; if the maximum gain is lower than the preset gain threshold, it is determined that the parameter does not need to be adjusted.

8. The method according to claim 7, characterized in that, In step (S4.1), the search interval The rules for determining it are as follows: For parameters whose values ​​need to be increased to improve immersion, the upper bound of the search range is... The lower realm , The physical upper limit; For parameters whose values ​​need to be reduced to improve immersion, the lower bound of the search range is... Upper Realm ; This is the physical lower limit; where, This is the current baseline value for this parameter.

9. The method according to claim 7, characterized in that, Following step (S4), a multi-parameter upgrade path evaluation step is also included: The recommended upgrade values ​​of each parameter are superimposed on the baseline configuration in different orders to form multiple upgrade paths; For each upgrade path, parameter changes are applied sequentially, and the updated prediction score is calculated using the immersion prediction model at each step to obtain the immersion enhancement trajectory. Based on the different paths of immersion enhancement, the balance between overall upgrade costs and immersion benefits is evaluated, providing a basis for decision-making on comprehensive upgrade solutions.

10. A virtual reality immersion prediction and display parameter optimization system, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program, and when the processor executes the program, it implements the virtual reality immersion prediction and optimization method based on constraint-aware depth regression as described in any one of claims 1 to 9.

Images