A virtual reality route-level moving speed stability prediction method and device, electronic equipment and storage medium
Patent Information
- Application Number
- CN202611200285.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-10
- Publication Date
- 2026-10-30
AI Technical Summary
[0006]本发明的目的是提出一种虚拟现实路线级移动速度稳定性预测方法、装置、电子设备及存储介质，以解决完整试次级指标无法反映局部空间差异、不同用户路线片段缺乏可比性以及有限标注条件下头眼时序模型跨用户泛化不足的问题
[0032]1、本发明基于空间拓扑先验和跨用户模板建立固定路线分段，使不同用户在相同路径位置的行为片段具有一致空间语义；
Smart Images
Figure CN123015914A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of virtual reality technology, and in particular to a method, apparatus, electronic device, and storage medium for predicting the stability of virtual reality route-level movement speed. Background Technology
[0002] During immersive virtual reality navigation, users need to continuously perform visual search, spatial positioning, route decision-making, and body posture adjustment. There is a continuous coupling between head orientation, eye movement direction, and body movement, and local velocity fluctuations can reflect the motion control state in different route areas.
[0003] Current virtual reality navigation evaluations typically use macro-level metrics such as total task time, path length, or average speed across all trials. These metrics compress the entire route into a single outcome, making it difficult to distinguish between local states with different spatial semantics, such as turning areas, straight-ahead areas, and path decision areas. Consequently, they are not conducive to predicting route-level speed stability.
[0004] On the other hand, head pose, eye movement direction, and body position data suffer from inconsistent sampling frequencies, timestamp jitter, local missing data, and individual differences. If training and testing data are not isolated according to user identity, the behavioral characteristics of the same user may participate in both model training and performance evaluation, making it difficult to reflect the actual predictive ability for unknown users. When labeled samples are limited, directly supervised training of temporal encoders also struggles to obtain stable cross-user dynamic representations.
[0005] Therefore, there is a need for a virtual reality route-level movement speed stability prediction method, device, electronic device, and storage medium that combines virtual scene spatial topology, physical movement trajectory, and head-eye temporal modeling to construct local movement speed stability indices on route segments with consistent spatial semantics and complete predictions for unknown users using limited labeled data. Summary of the Invention
[0006] The purpose of this invention is to propose a method, device, electronic device and storage medium for predicting the stability of virtual reality route-level movement speed, in order to solve the problems that complete trial sub-indices cannot reflect local spatial differences, the lack of comparability of different user route segments, and the insufficient cross-user generalization of head-eye time series models under limited annotation conditions.
[0007] A virtual reality route-level movement speed stability prediction method includes:
[0008] Acquire the user's waist center of gravity position sequence, head posture sequence, eye movement direction sequence, and corresponding timestamps during virtual reality navigation;
[0009] Transform the waist center of mass position sequence to a fixed coordinate system in the virtual scene, and align the waist center of mass position sequence, head pose sequence, and eye movement direction sequence to a unified time grid;
[0010] A normalized path process is established based on the path topology and turning decision node information of the virtual scene. Turning nodes are determined based on the normalized path process, the semantic prior of the turning decision nodes, and the cross-user turning node template. Turning segments are generated with the turning nodes as the center. The trajectory between adjacent turning segments is defined as a straight segment. Route segments that do not meet the validity conditions are filtered out.
[0011] A four-channel time-series input is constructed within each effective route segment to extract single-modal statistical features and cross-modal relationship features;
[0012] The four-channel time series input is input into the sequence coding branch of the prediction model, and the single-modal statistical features and cross-modal relationship features are input into the feature coding branch of the prediction model for training.
[0013] The high-level representations output by the sequence coding branch and the feature coding branch are concatenated and input into the regression head. The regression head outputs the predicted value of the velocity variation coefficient, which characterizes the stability of the segmented movement speed of the target route.
[0014] Furthermore, the process for determining the turning node is as follows:
[0015] For the resampled waist centroid trajectory, the heading angle is calculated based on the planar displacement direction of adjacent trajectory points. The absolute value of the period difference between adjacent heading angles is accumulated within the sliding window to obtain the cumulative steering angle. Then, the turning candidate score is constructed based on the cumulative steering angle and the semantic prior of the steering decision node. The local maximum value of the turning candidate score is selected as the turning node in the preset search neighborhood of each cross-user template position.
[0016] Furthermore, the construction process of the four-channel timing input is as follows:
[0017] The forward head direction vector and the eye movement direction vector are converted into yaw angle and pitch angle, respectively. The yaw angle sequence is periodically continuous and the difference is calculated between adjacent unified time grid points to obtain a four-channel timing input.
[0018] Furthermore, the cross-user turning node template is obtained by aggregating the median or robust mean of the normalized path process positions of multiple training users at the same topology turning decision node; the validity conditions include: the segment boundary is within the complete trajectory range, the segment space length is within a preset range, the number of valid position samples and the number of valid head-eye samples reach preset values, and the proportion of valid samples reaches a preset proportion.
[0019] Furthermore, during the training phase, the prediction model calculates the instantaneous velocity sequence based on the waist centroid position sequence within each effective route segment, and uses the ratio of the standard deviation to the mean of the instantaneous velocity sequence as the segment-level velocity variation coefficient label.
[0020] Furthermore, the single-modal statistical features include the mean, standard deviation, minimum, maximum, range, root mean square, and absolute mean difference extracted from the yaw and pitch motion sequences of the head and eye movements; the cross-modal relationship features include statistics on the difference between head and eye movements along the same axis, the ratio of head movement to eye movement amplitude, and zero-delay or low-order delay correlation coefficients; the sequence encoding branch is a temporal convolutional encoder with dilated convolution and residual connections; and the feature encoding branch is a multilayer perceptron.
[0021] Furthermore, prior to supervised training, self-supervised contrastive pre-training is performed using only the four-channel temporal input from the training user. Two random temporal augmentations are applied to the same four-channel temporal input to form positive sample pairs. The two augmented views are then input into a shared-parameter temporal convolutional encoder and a projection head to obtain normalized representations. Normalized temperature-scaled cross-entropy loss is used to update the temporal convolutional encoder. After pre-training, the projection head is removed, and the parameters of the temporal convolutional encoder are used as the initialization parameters for the sequence coding branch. Notably, the self-supervised pre-training, supervised training, feature normalization parameter calculation, missing value imputation parameter calculation, and model selection do not use the target user's data.
[0022] A virtual reality route-level movement speed stability prediction device, comprising:
[0023] The data acquisition module is used to acquire the user's waist center of gravity position sequence, head posture sequence, eye movement direction sequence and corresponding timestamps during the virtual reality navigation process;
[0024] The data transformation module is used to transform the waist center of mass position sequence to a fixed coordinate system of the virtual scene, and to align the waist center of mass position sequence, head posture sequence and eye movement direction sequence to a unified time grid.
[0025] The route segmentation module is used to establish a normalized path process based on the path topology and turning decision node information of the virtual scene, determine turning nodes based on the normalized path process, semantic priors of turning decision nodes and cross-user turning node templates, generate turning segments with turning nodes as the center, define the trajectory between adjacent turning segments as straight segments, and filter out route segments that do not meet the validity conditions.
[0026] The feature extraction module is used to construct a four-channel time series input within each effective route segment and extract single-modal statistical features and cross-modal relationship features;
[0027] The model training module is used to train the prediction model by inputting four-channel time-series inputs into the sequence coding branch, and inputting single-modal statistical features and cross-modal relational features into the feature coding branch of the prediction model.
[0028] The prediction output module is used to concatenate the high-level representations output by the sequence coding branch and the feature coding branch and input them into the regression head. The regression head outputs the predicted value of the velocity variation coefficient, which characterizes the stability of the segmented movement speed of the target route.
[0029] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement various steps in a virtual reality route-level movement speed stability prediction method.
[0030] A storage medium storing a computer program that, when executed by a processor, implements the various steps of a virtual reality route-level movement speed stability prediction method.
[0031] The beneficial effects of this invention are as follows:
[0032] 1. This invention establishes fixed route segments based on spatial topology priors and cross-user templates, so that the behavioral segments of different users at the same path location have consistent spatial semantics;
[0033] 2. This invention preserves continuous head-eye time series and explicit statistical summaries, reducing the loss of temporal structure information caused by using only statistical quantities;
[0034] 3. This invention improves the model initialization and cross-user prediction capabilities under limited labeling conditions by performing self-supervised pre-training of the temporal convolutional encoder using unlabeled sequences trained on users.
[0035] 4. The segmented velocity variation coefficient output by this invention can be used as an input parameter for virtual reality training evaluation, rehabilitation monitoring, and adaptive navigation interaction. Attached Figure Description
[0036] Figure 1 This is a flowchart of the overall process for predicting the stability of virtual reality route-level movement speed according to the present invention.
[0037] Figure 2 This is a schematic diagram of the fixed segmentation of the route based on spatial topology priors in this invention.
[0038] Figure 3 This is a schematic diagram of the self-supervised pre-training and late-stage fusion prediction model of the present invention.
[0039] Figure 4 This is a structural diagram of the virtual reality route-level movement speed stability prediction device of the present invention.
[0040] Figure 5 This is a schematic diagram of the structure of the electronic device of the present invention. Detailed Implementation
[0041] This invention proposes a method, device, electronic device, and storage medium for predicting the stability of virtual reality route-level movement speed. The invention will be further described below with reference to the accompanying drawings and specific embodiments.
[0042] Figure 1 The overall flowchart of the virtual reality route-level movement speed stability prediction method of the present invention is as follows:
[0043] 1. Obtain the user's waist center of gravity position sequence, head posture sequence, eye movement direction sequence, and corresponding timestamps during the virtual reality navigation process.
[0044] The virtual scene is constructed by a 3D engine and includes a pre-defined navigation route and multiple turning decision nodes. The user wears a head-mounted display with head and eye tracking capabilities, while a spatial position tracker is positioned at the waist. The system simultaneously records the 3D position of the waist's center of mass, head posture, eye movement direction, scene coordinates, and timestamps.
[0045] The original records were sorted in ascending order of timestamps, and duplicate timestamps, missing timestamps, and non-finite values were removed. Eye-tracking data were filtered based on validity markers, and the eye-tracking direction vectors were normalized. Abnormal jumps in head pose, eye-tracking direction, and waist position were removed before subsequent interpolation.
[0046] The position of the waist center of mass is mapped to a fixed coordinate system in the virtual scene through rigid body transformation. The head pose can be represented by Euler angles, quaternions, or rotation matrices; applying the head pose to the device's forward basis vector yields the head's forward unit direction vector. The eye movement direction is transformed to a coordinate system consistent with the head's forward direction based on the transformation relationship between the device coordinate system and the scene coordinate system.
[0047] Any unit direction vector is u = (u x , u y , u z Its yaw angle φ and pitch angle θ are calculated according to the following formulas:
[0048]
[0049]
[0050] The yaw angle sequence is periodically continuous to eliminate spurious large angles caused by crossing the −π and π boundaries.
[0051] 2. Transform the waist center of mass position sequence to a fixed coordinate system in the virtual scene, and align the waist center of mass position sequence, head pose sequence, and eye movement direction sequence to a unified time grid.
[0052] Interpolation using a unified time grid and adjacent differences yields Δφ(k) = φ[T(k+1)] − φ[T(k)] and Δθ(k) = θ[T(k+1)] − θ[T(k)]. This embodiment uses a 20 ms time step, corresponding to a 50 Hz time resolution.
[0053] 3. Establish a normalized path process based on the path topology and turning decision node information of the virtual scene. Determine turning nodes based on the normalized path process, semantic priors of turning decision nodes, and cross-user turning node templates. Generate turning segments with turning nodes as the center. Define the trajectory between adjacent turning segments as straight segments and filter out route segments that do not meet the validity conditions.
[0054] Figure 2 This is a schematic diagram of the fixed segmentation of the route based on spatial topology prior of the present invention. The system maps the complete waist centroid trajectory to a normalized path process s ∈ [0, 1] to eliminate differences in walking time and local speed among different users. Local speed, heading angle, and cumulative turning angle are calculated on the normalized path process.
[0055] For the k-th resampled trajectory point p(k) = [x(k), y(k), z(k)], ignoring minor changes in the vertical direction, the heading angle can be calculated as ψ(k) = arctan2{x(k+1) − x(k), z(k+1) − z(k)}. The heading angle is then periodically continuous, and the cumulative steering angle is calculated within a sliding window of half-width w:
[0056]
[0057] The predefined steering decision nodes in the virtual scene provide semantic priors m(k). m(k) can be binary or continuously decay according to the distance from the trajectory point to the theoretical turning position. The turning candidate score can be expressed as Q(k) = A(k) · [1 + λm(k)], where λ is the prior weight.
[0058] For training users, candidate peaks are first selected near each theoretical turning position and mapped to the normalized path process. For the r-th topological turning node, the initial positions s(i,r) of each user are aggregated using the median or robust mean to obtain the template position s̄(r). Then, for each user, the local maximum value of Q(k) in the neighborhood of [s̄(r)-δ, s̄(r)+δ] is selected as the actual turning node. This method utilizes the scene topology to ensure cross-user correspondence while preserving individual differences in turning positions.
[0059] Centered on each turning node, turning segments are generated according to the preset process half-width or the actual turning area boundary, and the interval between adjacent turning segments is defined as a straight segment. The validity conditions for segment validity control include at least the following: the segment boundary is within the complete trajectory range, the segment spatial length is within the preset range, the number of valid position samples and the number of valid head-eye samples reach preset values, and the proportion of valid samples reaches the preset proportion. Segments that do not meet the conditions are not included in model training or prediction.
[0060] During the training phase, the prediction model calculates instantaneous velocity sequences based on the waist centroid position sequences within each valid route segment. The ratio of the standard deviation to the mean of the instantaneous velocity sequence is used as the segment-level velocity variation coefficient label. For each valid route segment, at the start time tstart and end time tend, the instantaneous velocity v(k) is calculated based on the Euclidean distance and time difference between adjacent waist centroid positions. Samples with non-finite values, abrupt changes caused by tracking anomalies, and those exceeding the preset natural walking speed range of the scene are removed.
[0061] Piece-level velocity variation coefficient CV v,seg Calculate using the following formula:
[0062]
[0063] Where μv,seg and σv,seg are the mean and standard deviation of the effective instantaneous velocity of the segment, respectively. A larger CVv,seg indicates stronger relative velocity fluctuations and lower velocity stability within the segment; a smaller CVv,seg indicates more stable velocity.
[0064] Fourth, construct a four-channel time-series input within each effective route segment, and extract single-modal statistical features and cross-modal relationship features.
[0065] The forward head direction vector and the eye-tracking unit direction vector are converted into yaw and pitch angles, respectively. The yaw angle sequence is periodically continuous, and the difference between adjacent unified time grid points is calculated to obtain a four-channel timing input. For the i-th valid route segment, the head yaw angle motion, head pitch angle motion, eye-tracking yaw angle motion, and eye-tracking pitch angle motion under a unified time grid are extracted within the corresponding time window to form a four-channel sequence X(i) of size L × 4. In this embodiment, segments with different durations are unified into L = 128 time steps through interpolation, center clipping, or boundary padding.
[0066] The state vector at time step t is represented as x(i, t) = [Δφ H (i, t), Δθ H (i, t), Δφ E (i, t), Δθ E [(i, t)], where H represents head movement and E represents eye movement. The four channels jointly retain horizontal steering and vertical pitch information.
[0067] Single-modal statistical features are calculated for each of the four channels, including mean, standard deviation, minimum, maximum, range, root mean square, and absolute mean of differences. Cross-modal relationship features include the mean, standard deviation, or root mean square of the differences in head and eye movements along the same axis, the ratio of the root mean square of head movement to eye movement, and the Pearson correlation coefficient with zero or low-order delay.
[0068] Explicit features are computed only from the data within the current segment; feature standardization parameters and missing value imputation parameters are determined only from the training user data, and the target user data reuses only the determined parameters.
[0069] 5. Train the prediction model by inputting the four-channel time series input into the sequence coding branch, and inputting the single-modal statistical features and cross-modal relationship features into the feature coding branch.
[0070] Figure 3 This diagram illustrates the self-supervised pre-training and late-stage fusion prediction model of the present invention. In the self-supervised stage, only the head-eye four-channel sequence of the training user is used. For each X(i), two enhancement functions t1 and t2 are independently sampled to obtain two enhanced views X̃(i,1) = t1[X(i)] and X̃(i,2) = t2[X(i)]. Enhancement methods may include small-scale scaling or additive noise, continuous temporal occlusion, local cropping followed by resampling, and single-channel noise perturbation.
[0071] Temporal Convolutional Network (TCN) encoder with two enhanced view inputs sharing parameters θ and projection head qω The normalized representation u(i, m) = norm{q is obtained. ω [f θ (X̃(i,m))]}。 Two augmented views of the same original segment constitute a positive sample pair, while views of other segments in the same batch constitute negative samples.
[0072] For a batch containing B original segments, training is performed using normalized temperature-scaled cross-entropy loss:
[0073]
[0074] Where sim represents cosine similarity and τ is the temperature coefficient. After self-supervised training, the projection head is removed, and the temporal convolutional encoder parameters are retained as the initialization of the sequence branches of the supervised model.
[0075] 6. The high-level representations output by the sequence coding branch and the feature coding branch are concatenated and input into the regression head. The regression head outputs the predicted value of the velocity variation coefficient, which characterizes the stability of the segmented movement speed of the target route.
[0076] Before supervised training, self-supervised contrastive pre-training is performed using only the four-channel temporal input of the training user. Two random temporal augmentations are applied to the same four-channel temporal input to form positive sample pairs. The two augmented views are input into the temporal convolutional encoder and the projection head with shared parameters to obtain normalized representations. Normalized temperature-scaled cross-entropy loss is used to update the temporal convolutional encoder. After pre-training, the projection head is removed, and the parameters of the temporal convolutional encoder are used as the initialization parameters of the sequence coding branch.
[0077] During the supervision phase, the sequence branch encodes X(i) as z. seq Explicit feature branching encodes the feature vector s(i) into z. feat The fusion is represented as z = [z seq z feat The regression head outputs ŷ(i). The supervised loss is as follows:
[0078]
[0079] Where y(i) is the true segmented velocity variation coefficient.
[0080] In this embodiment, the sequence encoding branch is a temporal convolutional encoder with dilated convolution and residual connections, and the feature encoding branch is a multilayer perceptron. The input sequence size is 128 × 4. The temporal convolutional encoder outputs a 32-dimensional representation, the projection head outputs a 64-dimensional normalized representation during the self-supervised stage, the explicit feature branch outputs a 32-dimensional representation, and the two branches are concatenated to form a 64-dimensional representation. The fusion regression head adopts a 64 → 32 → 1 fully connected structure. The above dimensions can be adjusted according to the sample size and computational resources.
[0081] To evaluate and deploy generalization capabilities for unknown users, data is isolated according to user identity. In each round, the entire route of one target user is segmented and used as test data, while the data of the remaining users is used as training data; the target user's data is not used in self-supervised pre-training, supervised training, feature normalization parameter calculation, missing value imputation parameter calculation, and model selection.
[0082] In this embodiment, 14 turning decision nodes are set up in a virtual supermarket scenario, and navigation data from 36 users are collected using a head-mounted display device with head tracking and eye tracking capabilities, as well as a waist tracker. The model prediction target is segmented computer vision (CV). v,seg Training and testing were conducted using a leave-one-user cross-validation method based on user identity isolation.
[0083] Under full yaw and pitch four-channel input, the late-stage fusion model using self-supervised pre-training temporal convolutional model achieved a MAE of 0.052, a Pearson correlation coefficient r of 0.762, and a coefficient of determination R0. 2 The prediction result is 0.579. The performance of the model that retains only the yaw direction is lower than that of the four-channel model, indicating that pitch direction information can provide supplementary information for route-level local motion speed stability prediction.
[0084] The above results are used to illustrate the feasibility and predictive effect of the technical solution of the present invention in the embodiments, and do not constitute a limitation on the type of virtual scene, acquisition device, number of turning nodes, number of network layers, feature dimension or training parameters.
[0085] Figure 4 This is a structural diagram of the virtual reality route-level movement speed stability prediction device of the present invention. It includes:
[0086] The data acquisition module is used to acquire the user's waist center of gravity position sequence, head posture sequence, eye movement direction sequence and corresponding timestamps during the virtual reality navigation process;
[0087] The data transformation module is used to transform the waist center of mass position sequence to a fixed coordinate system of the virtual scene, and to align the waist center of mass position sequence, head posture sequence and eye movement direction sequence to a unified time grid.
[0088] The route segmentation module is used to establish a normalized path process based on the path topology and turning decision node information of the virtual scene, determine turning nodes based on the normalized path process, semantic priors of turning decision nodes and cross-user turning node templates, generate turning segments with turning nodes as the center, define the trajectory between adjacent turning segments as straight segments, and filter out route segments that do not meet the validity conditions.
[0089] The feature extraction module is used to construct a four-channel time series input within each effective route segment and extract single-modal statistical features and cross-modal relationship features;
[0090] The model training module is used to train the prediction model by inputting four-channel time-series inputs into the sequence coding branch, and inputting single-modal statistical features and cross-modal relational features into the feature coding branch of the prediction model.
[0091] The prediction output module is used to concatenate the high-level representations output by the sequence coding branch and the feature coding branch and input them into the regression head. The regression head outputs the predicted value of the velocity variation coefficient, which characterizes the stability of the segmented movement speed of the target route.
[0092] In summary, this invention enables behavioral segments of different users at the same path location to have consistent spatial semantics, reducing the loss of temporal structure information caused by using only statistics; in addition, it can improve model initialization and cross-user prediction capabilities under limited annotation conditions, and the output segmented velocity variation coefficient can be used as input parameters for virtual reality training evaluation, rehabilitation monitoring, and navigation interaction adaptation.
[0093] This embodiment also includes an electronic device and a storage medium. Figure 5 This is a schematic diagram of the electronic device of the present invention. An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements various steps in the virtual reality route-level movement speed stability prediction method. A storage medium stores a computer program thereon. When the computer program is executed by a processor, it implements various steps in the virtual reality route-level movement speed stability prediction method.
[0094] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented in various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.
[0095] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0096] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0097] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0098] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.
[0099] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.
Claims
1. A method for predicting the stability of movement speed at the route level in virtual reality, characterized in that, include: Acquire the user's waist center of gravity position sequence, head posture sequence, eye movement direction sequence, and corresponding timestamps during virtual reality navigation; Transform the waist center of mass position sequence to a fixed coordinate system in the virtual scene, and align the waist center of mass position sequence, head pose sequence, and eye movement direction sequence to a unified time grid; A normalized path process is established based on the path topology and turning decision node information of the virtual scene. Turning nodes are determined based on the normalized path process, the semantic prior of the turning decision nodes, and the cross-user turning node template. Turning segments are generated with the turning nodes as the center. The trajectory between adjacent turning segments is defined as a straight segment. Route segments that do not meet the validity conditions are filtered out. A four-channel time-series input is constructed within each effective route segment to extract single-modal statistical features and cross-modal relationship features; The four-channel time series input is input into the sequence coding branch of the prediction model, and the single-modal statistical features and cross-modal relationship features are input into the feature coding branch of the prediction model for training. The high-level representations output by the sequence coding branch and the feature coding branch are concatenated and input into the regression head. The regression head outputs the predicted value of the velocity variation coefficient, which characterizes the stability of the segmented movement speed of the target route.
2. The virtual reality route-level movement speed stability prediction method according to claim 1, characterized in that, The process for determining the turning point is as follows: For the resampled waist centroid trajectory, the heading angle is calculated based on the planar displacement direction of adjacent trajectory points. The absolute value of the period difference between adjacent heading angles is accumulated within the sliding window to obtain the cumulative steering angle. Then, the turning candidate score is constructed based on the cumulative steering angle and the semantic prior of the steering decision node. The local maximum value of the turning candidate score is selected as the turning node in the preset search neighborhood of each cross-user template position.
3. The virtual reality route-level movement speed stability prediction method according to claim 1 or 2, characterized in that, The construction process of the four-channel timing input is as follows: The forward head direction vector and the eye movement direction vector are converted into yaw angle and pitch angle, respectively. The yaw angle sequence is periodically continuous and the difference is calculated between adjacent unified time grid points to obtain a four-channel timing input.
4. The virtual reality route-level movement speed stability prediction method according to claim 3, characterized in that, The cross-user turning node template is obtained by aggregating the median or robust mean of the normalized path process positions of multiple training users at the same topology turning decision node. The validity conditions include: the segment boundary is within the complete trajectory range, the segment spatial length is within a preset range, the number of valid position samples and the number of valid head and eye samples reach preset values, and the proportion of valid samples reaches a preset proportion.
5. The virtual reality route-level movement speed stability prediction method according to claim 1, characterized in that, During the training phase, the prediction model calculates the instantaneous velocity sequence based on the position sequence of the center of mass at the waist within each effective route segment, and uses the ratio of the standard deviation to the mean of the instantaneous velocity sequence as the label of the segment-level velocity variation coefficient.
6. The virtual reality route-level movement speed stability prediction method according to claim 1, characterized in that, The single-modal statistical features include the mean, standard deviation, minimum, maximum, range, root mean square, and absolute mean difference extracted from the yaw and pitch motion sequences of the head and eye movements; the cross-modal relationship features include statistics on the difference between head and eye movements along the same axis, the ratio of head movement to eye movement amplitude, and zero-delay or low-order delay correlation coefficients; the sequence encoding branch is a temporal convolutional encoder with dilated convolution and residual connections. The feature encoding branch is a multilayer perceptron.
7. The virtual reality route-level movement speed stability prediction method according to claim 5, characterized in that, Before supervised training, self-supervised contrastive pre-training was performed using only the four-channel temporal input of the training user. Two random temporal augmentations were applied to the same four-channel temporal input to form positive sample pairs. The two augmented views were then input into the temporal convolutional encoder and the projection head with shared parameters to obtain normalized representations. Normalized temperature-scaled cross-entropy loss was used to update the temporal convolutional encoder. After pre-training, the projection head was removed, and the parameters of the temporal convolutional encoder were used as the initialization parameters for the sequence coding branch. Notably, the self-supervised pre-training, supervised training, feature normalization parameter calculation, missing value imputation parameter calculation, and model selection did not use the target user's data.
8. A virtual reality route-level movement speed stability prediction device, characterized in that, include: The data acquisition module is used to acquire the user's waist center of gravity position sequence, head posture sequence, eye movement direction sequence and corresponding timestamps during the virtual reality navigation process; The data transformation module is used to transform the waist center of mass position sequence to a fixed coordinate system of the virtual scene, and to align the waist center of mass position sequence, head posture sequence and eye movement direction sequence to a unified time grid. The route segmentation module is used to establish a normalized path process based on the path topology and turning decision node information of the virtual scene, determine turning nodes based on the normalized path process, semantic priors of turning decision nodes and cross-user turning node templates, generate turning segments with turning nodes as the center, define the trajectory between adjacent turning segments as straight segments, and filter out route segments that do not meet the validity conditions. The feature extraction module is used to construct a four-channel time series input within each effective route segment and extract single-modal statistical features and cross-modal relationship features; The model training module is used to train the prediction model by inputting four-channel time-series inputs into the sequence coding branch, and inputting single-modal statistical features and cross-modal relational features into the feature coding branch of the prediction model. The prediction output module is used to concatenate the high-level representations output by the sequence coding branch and the feature coding branch and input them into the regression head. The regression head outputs the predicted value of the velocity variation coefficient, which characterizes the stability of the segmented movement speed of the target route.
9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements each step of the virtual reality route-level movement speed stability prediction method as described in any one of claims 1 to 7.
10. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements each step of the virtual reality route-level movement speed stability prediction method as described in any one of claims 1 to 7.