A pre-rendering method based on intelligent cockpit attitude prediction

Abstract

The present application belongs to the technical field of virtual reality simulation, and particularly relates to a pre-rendering method based on intelligent cockpit attitude prediction, comprising the following steps: establishing an instruction flow continuous analysis module at the entrance of a rendering pipeline; in response to detecting a turning event trigger point, analyzing the instruction queue after turning, extracting the angular acceleration sequence corresponding to the target attitude sequence, and extracting the time domain parameters including the acceleration segment length, the uniform speed segment length and the deceleration segment length on the angular acceleration sequence; constructing a nonlinear transition coefficient aligned with the phase of the platform real motion trajectory based on the time domain parameters, and performing hierarchical modulation on the nonlinear transition coefficient based on the depth buffer of the current frame to generate a pixel-by-pixel depth perception mixing coefficient. The present application eliminates the systematic visual attitude deviation caused by the linear mixing scheme and reduces the false artifacts of layer sliding at the depth boundary.

Description

Technical Field

[0001] This invention belongs to the field of virtual reality simulation technology, specifically relating to a pre-rendering method based on intelligent cockpit attitude prediction. Background Technology

[0002] In the field of virtual reality simulation, 6-DOF intelligent cockpits have been widely used in flight simulation training, driving simulation and theme park scenarios to achieve an immersive experience. The system needs to keep the user's visual image highly synchronized with the physical posture of the motion platform. The end-to-end rendering latency must be strictly controlled within 20ms. Once this limit is exceeded, the time difference between vision and vestibular sensation will directly induce simulator vertigo.

[0003] Patent document CN119720324B discloses a method and system for intelligent prediction of bridge tower segment attitude based on multi-point collaborative positioning. The method includes: S100, selecting multiple observation points on two adjacent sides of the installed segment, obtaining the spatial coordinates of the observation points using a laser reflector, and performing collaborative measurement using multiple total stations; S200, predicting the attitude change of the segment to be installed based on a rigid body transformation mathematical model, calculating the optimal attitude of the segment to be installed using a rotation matrix and translation vector, optimizing the attitude estimation through multi-point redundant information, and controlling the cumulative error; S300, performing BIM modeling of the bridge tower structure and high-precision 3D rendering, transforming complex monitoring data into an intuitive graphical interface to display the segment status in real time; S400, analyzing the data and combining it with error, and making corresponding adjustments if the error with the design data is large; S500, repeating the above steps until all bridge tower segments are accurately positioned and installed.

[0004] To reduce rendering latency, existing technologies use temporal networks such as LSTM as the core, taking the historical position, velocity, acceleration, and angular velocity sequences of the motion platform as input to predict the pose of the next 1 to 2 frames. The predicted pose is then used as the extrinsic parameter for the rendering camera, generating images in advance and storing them in a buffer. When the actual pose feedback arrives and the prediction deviation exceeds a threshold, pixel displacement field interpolation correction based on optical flow estimation is triggered. To address the historical window lag problem at transition moments, existing technologies employ a hybrid approach combining control command flow pre-parsing and dual-branch pre-rendering. The motion controller command flow has a timing advantage of approximately 8 to 15 ms ahead of IMU physical feedback, enabling advance detection of transition events. The historical branch and the target branch are rendered separately and then processed with a linear transition coefficient. Pixel-by-pixel weighted blending is used to achieve a smooth transition in visual images corresponding to the old and new motion states.

[0005] When the actuators of the 6-DOF motion platform complete a state transition, their actual attitude trajectory is not a nonlinear interpolation, but follows an S-shaped velocity envelope consisting of an acceleration phase, a constant velocity phase, and a deceleration phase. At the beginning of the transition, the platform's angular acceleration rapidly increases, reaching its peak within approximately 30% to 40% of the transition time, then entering a peak angular velocity plateau period, and finally entering the deceleration convergence phase. During the acceleration phase, the actual platform attitude change rate consistently exceeds the linear range. The implicit assumption of uniform velocity adopts a linear approach. By advancing the hybrid weighting, the visual pose presented by the hybrid frames systematically lags behind the platform's actual pose; similarly, during the deceleration phase, linear... Furthermore, the transition progress is overestimated, causing the blended frame to lead the actual posture. The visual consequences of this phase deviation in the 3D scene are amplified layer by layer due to depth. The pixel displacement of near-field scene elements and far-field elements under the same posture angle error can differ by up to two orders of magnitude. Pixel-by-pixel linear blending cannot distinguish depth levels, and near-field and far-field elements exhibit inconsistent motion rates in the blended frame, creating a visual tearing sensation as if the scene is sliding at the depth boundary. This tearing sensation is most intense precisely at the peak angular acceleration point between 30% and 40% of the transition time, which is also the moment when the vestibular otolith sensory line acceleration signal intensity is highest. The superposition of the dual peaks of vision and vestibular perception makes the intensity of simulator dizziness far exceed the effect of a simple frame pause. Summary of the Invention

[0006] This invention provides a pre-rendering method based on intelligent cockpit attitude prediction to solve the technical problems in the prior art where linear frame mixing causes near-field and far-field layering visual tearing due to neglecting the platform's S-shaped velocity envelope and scene depth hierarchy differences during the transition of intelligent cockpit motion state, and exacerbates vestibular and visual conflicts at the moment of peak angular acceleration.

[0007] This invention provides a pre-rendering method based on intelligent cockpit attitude prediction, comprising the following steps: S1, establish a continuous instruction stream parsing module at the entrance of the rendering pipeline. In response to the detection of the turning event trigger point, parse the instruction queue after the turning point, extract the angular acceleration sequence corresponding to the target posture sequence, and extract the temporal parameters including the duration of the acceleration segment, the duration of the constant speed segment, and the duration of the deceleration segment from the angular acceleration sequence. S2, based on temporal parameters, constructs nonlinear transition coefficients that are phase-aligned with the platform's actual motion trajectory, and performs layered modulation on the nonlinear transition coefficients based on the depth buffer of the current frame to generate pixel-by-pixel depth-aware hybrid coefficients. S3 uses the pixel-per-pixel depth-aware blending coefficient as weight to perform depth-layered forward mapping synthesis on the dual-branch frames generated by the rendering pipeline and outputs a blended frame. S4, in response to the arrival of actual physical attitude feedback, performs residual verification and correction based on the threshold triggered by the dynamically tightened residual correction during the peak angular acceleration period, and performs directional hidden state reset on the prediction model based on the real inertial measurement unit data after the turning event ends.

[0008] By extracting the three time-domain parameters of the S-shaped velocity envelope from the control command stream, the normalized displacement progress transition coefficient corresponding to the trapezoidal velocity planning standard is derived, so that the visual progress of the hybrid frame and the actual motion trajectory of the platform remain in phase throughout the acceleration, constant speed and deceleration phases, effectively eliminating the systematic visual attitude advance or lag deviation caused by the original linear hybrid scheme in the acceleration and deceleration phases.

[0009] Furthermore, the instruction queue after the transition is parsed to extract the angular acceleration sequence corresponding to the target attitude sequence. Temporal parameters, including the durations of the acceleration, constant velocity, and deceleration phases, are then extracted from the angular acceleration sequence, including: The target attitude sequence after the turning event trigger point is subjected to first-order difference according to the sampling period to obtain the command angular velocity sequence; the command angular velocity sequence is subjected to first-order difference to obtain the angular acceleration sequence. Extreme points and zero-crossing points are detected on the angular acceleration sequence. The time interval when the angular acceleration first reaches the peak value, the duration when the angular acceleration is maintained above the preset proportion of the peak value, and the time interval when the angular acceleration monotonically decays from the peak value to zero are determined respectively. Finally, time-domain parameters including the duration of the acceleration segment, the duration of the uniform velocity segment, and the duration of the deceleration segment are obtained.

[0010] By using two first-order differences of the attitude sequence to obtain the angular velocity and angular acceleration sequences, and relying on the detection of extreme points and zero crossing points to separate the three time-domain parameters of acceleration, uniform speed and deceleration, the S-shaped velocity change characteristics of the motion platform are accurately restored. This provides reliable original time-domain data support for the accurate construction of the subsequent nonlinear transition coefficients, and avoids the phase shift between the transition coefficients and the real motion trajectory caused by the distortion of parameter extraction during the motion stage.

[0011] Furthermore, a nonlinear transition coefficient, phase-aligned with the platform's actual motion trajectory, is constructed based on time-domain parameters, including: Based on the trapezoidal velocity planning standard, the relationship between angular velocity and time in the acceleration, uniform velocity, and deceleration segments is determined. Integrating the angular velocity within each segment with respect to time yields the normalized displacement progress as a function of time. Each segment is removed and normalized to the total displacement, ultimately forming a nonlinear transition coefficient that is phase-aligned with the platform's actual motion trajectory.

[0012] By directly mapping the integral displacement progress of trapezoidal velocity planning to transition coefficients, the hybrid progress is aligned with the platform's actual motion envelope in the time dimension, eliminating the systematic deviation of the linear hybrid scheme in the phase direction.

[0013] Furthermore, the nonlinear transition coefficients are layered and modulated based on the depth buffer of the current frame to generate pixel-by-pixel depth-aware mixing coefficients, including: Get the normalized depth value of each pixel in the depth buffer of the current frame; The near-field disparity ratio is calculated based on the normalized depth of the frustum of the near-field and far-field reference objects, and the near-field conservatism coefficient is determined based on the near-field disparity ratio. The nonlinear transition coefficients are adjusted pixel by pixel using near-field conservative coefficients and normalized depth values, ultimately generating pixel-by-pixel depth-aware hybrid coefficients.

[0014] Furthermore, using pixel-by-pixel depth-aware blending coefficients as weights, depth-layered forward mapping compositing is performed on the dual-branch frames generated by the rendering pipeline to output a blended frame, including: The dual-branch frame includes a continuation branch frame and a target branch frame, and the depth map is obtained synchronously output by the continuation branch frame and the target branch frame during generation; For each pixel in the image, the corresponding pixel-wise depth-aware blending coefficient is queried based on the depth value; The depth level of a pixel is determined based on a depth boundary threshold, and occlusion-aware forward mapping or weighted averaging is performed according to the depth level to finally output a mixed frame.

[0015] By adaptively defining the near-field and far-field boundaries based on depth statistical features, occlusion perception forward mapping is implemented in the near-field region, and the far-field region is directly degraded to a weighted average. This allows high-disparity elements such as near-field instruments and low-disparity elements such as the far-field horizon to undergo synthesis processing that matches their own disparity amplitude within the same mixed frame, thus eliminating layered slip artifacts at the depth boundaries from the geometric source.

[0016] Furthermore, the depth level of a pixel is determined based on a depth boundary threshold, and occlusion-aware forward mapping or weighted averaging is performed according to the depth level, including: For near-field pixels with depth values ​​less than the depth boundary threshold, the corresponding 3D spatial points in the continuing branch frame will be reprojected onto the target view coordinate system according to the camera extrinsic parameters of the target branch frame. If the reprojected coordinates pass the depth test, then the weighted average of the two pixels is taken. If the reprojected coordinates fall within the depth test failure area of ​​the occlusion, then the pixel values ​​of the continuation branch frame are directly used to perform occlusion-aware forward mapping.

[0017] Furthermore, the depth boundary threshold is dynamically determined through the following steps: the normalized depth values ​​of all pixels in the current frame's full-frame depth map are sorted in ascending order; The upper bound of the near-field proportion is determined based on the cumulative proportion corresponding to the first significant valley in the current frame depth histogram. Find the depth upper bound value corresponding to the cumulative number of pixels reaching the upper bound of the near field proportion of the total number of pixels in the whole frame, and use the depth upper bound value as the depth boundary threshold.

[0018] Furthermore, during the peak angular acceleration period, residual verification and correction are performed based on a dynamically tightened residual correction trigger threshold, including: Obtain the maximum tolerable disparity pixel offset corresponding to the normal threshold; The single-frame disparity pixel increment is calculated based on the peak angular velocity in the current instruction stream and the near-field reference depth read from the mean of the current frame depth buffer. Based on the ratio of the maximum tolerable disparity pixel offset to the single-frame disparity pixel increment, the dynamic tightening residual correction trigger threshold is calculated. Based on the dynamically tightened residual correction trigger threshold, residual verification and correction are preferentially performed on the near-field depth boundary pixel band.

[0019] By dynamically calculating and tightening the threshold based on the current peak angular velocity and near-field depth during the peak angular acceleration period, the sensitivity of residual correction triggering is directly linked to the actual visual risk level at the current moment. Priority correction is implemented on the near-field depth boundary region during the critical time period when the visual and vestibular conflict is most intense, which effectively suppresses near-field visual artifacts induced by small residuals during this high-risk period.

[0020] Furthermore, after the transition event, a directional hidden state reset is performed on the prediction model based on real inertial measurement unit data, including: The transition event is considered to have completely ended when the nonlinear transition coefficient reaches its maximum normalized value. Extract the actual inertial measurement unit data corresponding to the deceleration and constant velocity phases during the new state establishment phase; The extracted real inertial measurement unit data is input into the encoder of the prediction model to update the model's hidden state and clear the residual leave state memory vector, thus completing the orientation hidden state reset.

[0021] Furthermore, an instruction flow continuous parsing module is established at the entry point of the rendering pipeline, including: Determine the motion controller's output cycle; The instruction stream continuous parsing module is initialized at the entrance of the rendering pipeline according to the issuance cycle.

[0022] The beneficial effects of this invention are: by dynamically tightening the residual threshold based on the real-time peak angular velocity and near-field depth during the peak angular acceleration period, and by resetting the LSTM hidden state in a directional manner with real IMU data during the deceleration and constant velocity periods after the turning event has completely ended, the trigger sensitivity of residual correction is matched with the actual visual risk level of the current frame in real time.

[0023] While ensuring that the prediction model can recover its full prediction capability by starting with high-quality real motion data after the new steady state is established, it effectively avoids the accumulation of prediction bias in the early stage of the new steady state caused by hidden state contamination, thus improving the overall stability and accuracy of pre-rendering. Attached Figure Description

[0024] Figure 1 This is a flowchart of the pre-rendering method based on intelligent cockpit attitude prediction in this invention.

[0025] Figure 2 This is a schematic diagram of the dynamic residual correction trigger threshold tightening mechanism during the peak angular acceleration period in this invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] An embodiment of the pre-rendering method based on intelligent cockpit attitude prediction provided by this invention: like Figure 1 As shown, the pre-rendering method based on intelligent cockpit attitude prediction includes the following steps: S1, parse the instruction stream to extract time-domain parameters.

[0028] In an optional embodiment, the intelligent cockpit motion controller issues a sequence of target attitude commands to the actuators at fixed intervals. This sequence of commands implicitly contains complete motion planning information for the current state transition. A continuous command flow parsing module is established at the rendering pipeline inlet, responding to the detection of a transition event trigger point. The system immediately parses the instruction queue after the turn and extracts the S-shaped velocity envelope parameters, instead of simply outputting a single turn intensity scalar. For example, the motion controller can send commands at a frequency of 1 kHz, updating the target attitude to the actuator every 1 ms.

[0029] The instruction stream continuous parsing module has The target attitude sequence is then first-diffused according to the sampling period to obtain the commanded angular velocity sequence for each degree of freedom. The commanded angular acceleration sequence is then obtained by performing another first-diffusion on the obtained commanded angular velocity sequence. The sigmoid envelope on the angular acceleration sequence is characterized by a monotonically increasing angular acceleration to a peak value. Then, maintaining an approximately constant value, and finally monotonically decreasing to 0, the instruction stream continuous parsing module detects extreme points and zero-crossing points on the angular acceleration sequence, extracting the following three key time-domain parameters: acceleration segment duration. (from When the angular acceleration first reached (Time interval), duration of constant speed segment (angular acceleration maintained at) (the duration of the preset ratio and above) and the duration of the deceleration phase (angular acceleration from) The total duration of the transition period is obtained by simply adding the three durations (the time interval that monotonically decays to 0). :

[0030] Furthermore, three time-domain parameters , , Together Together, they constitute the complete input for the hybrid control of this frame, all derived from observable data in the instruction queue. It should be noted that there are many other ways in the art to detect transition events from the instruction stream. Those skilled in the art can choose different parsing methods based on the control protocol type of the actual motion platform. For example, performing frame-level parsing of servo bus messages or frame-by-frame sliding window comparison of CAN bus instruction byte streams are all within the scope of this invention.

[0031] The rendering pipeline obtains a complete temporal description of the S-shaped velocity envelope at the start of platform motion, eliminating the information lag caused by historical window inference for the accurate construction of subsequent transition coefficients.

[0032] S2, construct the transition coefficients and generate the mixing coefficients.

[0033] In an optional embodiment, based on the three-segment time-domain parameters extracted by S1, a nonlinear transition coefficient aligned with the phase of the platform's actual motion trajectory is constructed. The linear substitution exhibits a systematic phase deviation from the actual motion envelope. This is to eliminate the lead and lag mismatch in the acceleration and deceleration phases of the mixed frames.

[0034] Define relative time ,in, For the system's global absolute time, based on the trapezoidal velocity programming standard commonly used in industrial motion control, the angular velocity increases linearly with time during the acceleration phase, i.e. ( ), Indicates The instantaneous angular velocity function of the platform is the independent variable. This represents the peak angular velocity; during the uniform velocity phase, the angular velocity is constant. During the deceleration phase, the angular velocity decreases linearly to 0, i.e. ( ).

[0035] Based on trapezoidal velocity planning, the angular velocity within each segment is integrated over time to obtain the normalized displacement progress versus time relationship, i.e., the transition coefficient. The integral result needs to be normalized to... Therefore, each segment is removed to obtain the total displacement. ,in:

[0036] This leads to the improved piecewise expression for the transition coefficient, which corresponds to the standard formula for trapezoidal velocity programming.

[0037] In the above formula , , , and All values ​​are extracted by S1 from the instruction stream or directly calculated from the extracted values. Represents time with relative motion A real-time varying nonlinear transition coefficient function. (In the acceleration phase...) The growth rate and angular velocity increase synchronously, ensuring that the mixing progress and the actual displacement of the platform remain in phase; the deceleration phase converges symmetrically, making... When approaching the target state, the platform's angular velocity returns to zero for synchronization. In the above formula, A larger value indicates that the platform's movement rate is higher at the current moment and the mixing progress is faster; a smaller value indicates that the platform is approaching the target state and its movement is slowing down, and the mixing progress is slowing down accordingly, thereby eliminating systematic phase mismatch.

[0038] After completing the nonlinear transition coefficient After construction, the blending coefficients are further modulated in layers based on the depth buffer of the current frame. The rendering pipeline synchronously outputs a depth map while generating dual-branch frames, containing the normalized depth value for each pixel in the depth map. .Will Depth-layered modulation into pixel-by-pixel depth-sensing mixing coefficients :

[0039] Here The near-field conservative coefficient is derived from the near-far field disparity ratio. It is the normalized depth of the near-field reference object's frustum obtained from statistics in the current frame's depth buffer. Based on this, the normalized depth of the far-field reference is... Then, under the same attitude angle error Below, near-field pixel displacement Far-field pixel displacement The ratio is approximately equal to To ensure that the near-field mixing progress does not precede the actual near-field displacement rate, a near-field conservatism coefficient is applied. The constraints are:

[0040] Pick This allows the near-field end to be mixed with a conservative coefficient that matches the disparity ratio, while the far-field end ( )hour It is completely restored to the original transition coefficient. The depth is dynamically calculated in each frame by the depth buffer statistics, thus adapting to dynamic situations where the depth distribution changes with the scene.

[0041] The hybrid progress maintains phase consistency with the platform's actual motion envelope in the time dimension, while matching the parallax characteristics of each depth layer in the spatial dimension, thus eliminating the systematic deviations of the linear scheme in both phase and depth directions.

[0042] S3, perform forward mapping to synthesize the output mixed frame.

[0043] In an optional embodiment, the pixel-wise depth-aware blending coefficients output by S2 are used. Using weights, depth-layered forward mapping synthesis is performed on the dual-branch frames to replace the pixel-by-pixel weighted average that cannot distinguish depth levels, thereby eliminating near-field and far-field misalignment artifacts; the dual-branch frames include the continuation branch frame and the target branch frame.

[0044] Continued branch frames and target branch frame Each layer synchronously outputs its own perspective depth map during generation. and During the compositing stage, for each pixel in the image, based on its depth value... Query the corresponding and based on depth threshold Determine the composition path for this pixel.

[0045] For depth value For near-field pixels, occlusion-aware forward mapping is used to reproject the corresponding 3D spatial point in the continuation frame onto the target viewpoint coordinate system according to the camera extrinsic parameters of the target branch, thereby obtaining its position in the target viewpoint coordinate system. The corresponding pixel coordinates And take the weighted average of the two pixels; if the coordinates fall within the range after reprojection... If the depth test fails in a certain near-field occlusion area, then directly select... The pixel values ​​are used to avoid occlusion and penetration artifacts. For depth values... For far-field pixels, the parallax projection difference is much less than one pixel, directly degenerating into a standard weighted average, thus saving GPU computational overhead. This synthesis rule is expressed as:

[0046] in The current pixel coordinates, Let be the reprojection coordinates of the pixel from the target's viewpoint. The final fused output image features are for pixel p. To preserve the original pixel value at pixel p in the branch frame, Match pixels in reprojection for the target branch frame The original pixel value at that location.

[0047] Furthermore, the depth boundary threshold The depth is dynamically determined by the objective statistical characteristics of the current frame's depth buffer. The depth upper bound corresponding to the sum of the normalized depth values ​​of all pixels in the current frame's full-frame depth map, sorted in ascending order, and the percentage of these accumulated pixels reaching the upper bound of the near-field proportion, is used as the depth upper bound for this frame. Among them, the upper limit of the near-field proportion. The cumulative proportion corresponding to the first significant valley in the current frame depth histogram is used to characterize the natural depth fault between near-field occluders and mid-to-far-field background. It reads and computes directly from the depth buffer in each frame, thus adapting to the differences in cockpit geometry across different platforms such as flight simulators and driving simulators.

[0048] This synthesis operation is implemented using a GPU Compute Shader, with the near-field reprojection branch and the far-field weighted branch executed in parallel within different thread bundles. Far-field pixels are directly degenerated into a weighted average, and near-field reprojection only applies to depth boundary pixel bands, significantly reducing the total computational cost compared to full-frame optical flow correction and avoiding the iterative convergence process of optical flow estimation.

[0049] Near-field high-parallax elements and far-field low-parallax elements each undergo synthesis processing that matches their own parallax amplitude, eliminating layer slip artifacts at depth boundaries from the geometric source with acceptable GPU computational overhead.

[0050] S4, perform residual correction and hidden state reset.

[0051] In an optional embodiment, in response to the arrival of actual physical attitude feedback from the inertial measurement unit (IMU) and encoder, the system performs residual verification and correction. The peak angular acceleration period of the S-shaped envelope (i.e., During the time window corresponding to the critical time period when the platform angular acceleration rises from the half-peak to the peak, the phase deviation and depth parallax effects reach their maximum values ​​simultaneously. This phenomenon can cause layer artifacts to easily exceed the perception threshold. Therefore, this step implements priority-enhanced verification for this period.

[0052] During the current peak angular acceleration period, the near-field elements are affected by the peak angular velocity. The resulting increase in single-frame parallax pixels is due to the camera focal length. z-coordinates of the view space relative to the near-field reference object Joint decision, Read from the instruction stream in real time. Normalized depth readings from the current frame's near-field reference in the depth buffer Distance from the near cutting plane of the cone of vision and distance of cutting surface The following can be obtained by calculation using the standard linear back projection relation:

[0053] Single-frame disparity pixel increment The calculation formula is as follows:

[0054] Near-field reference The smaller the value, the closer the object is to the camera, and the greater the pixel displacement produced at the same angular velocity. This increases accordingly, consistent with the physical laws governing high parallax in the near field.

[0055] Furthermore, in calculation before, if , To determine the maximum tolerable parallax pixel offset, directly... Set to the minimum positive threshold to trigger correction with the highest sensitivity; otherwise, the tightened dynamic residual correction trigger threshold is determined by the ratio of the tolerable offset to the risk increment:

[0056] The higher or The smaller, The larger, the better Compared to The greater the tightening, the more strictly the correction triggering timing is linked to actual visual risk, avoiding wasting correction resources during low angular velocity transitions. The correction range remains limited to a local area, meaning it is processed preferentially based on the depth map. The near-field depth boundary pixel bands. These pixel bands correspond to the highest parallax magnification, and if they are not processed first, they will produce layer artifacts; while in the far-field region, even if there are a small amount of residuals, the pixel displacement is extremely small, so they are not included in the correction calculation.

[0057] Furthermore, in response to the end of the peak angular acceleration period, the residual threshold is recovered. The system continues the mixing and synthesis process according to the standard procedure until... Reaching 1. When At this point, the transition event has completely ended, and the system performs a directed hidden state reset on the LSTM prediction model, establishing the new state in the next phase. The real IMU time-step data collected within the system is input into the encoder to update the model's hidden state and clear residual out-of-state memory vectors. During the acceleration phase, the platform is still in a strongly nonlinear dynamic process, and its IMU data has low representativeness for the new steady state; forcibly inputting it would introduce transient disturbances. However, the data from the deceleration and constant-speed phases fully reflect the motion characteristics of the target state. Therefore, directional reset is performed by inputting only the real data from the deceleration and constant-speed phases to improve the accuracy of the initial prediction starting point for the new steady state. After the reset, the prediction window and lead are fully restored, and the system enters a new steady-state pre-rendering cycle.

[0058] The trigger sensitivity of residual correction is matched in real time with the actual visual risk level of the current frame. After the new steady state is established, the prediction model recovers its full prediction capability starting from high-quality real motion data.

[0059] like Figure 2 As shown, it can be observed that during the period when the angular velocity envelope is close to its peak, the dynamic tightening threshold shifts the largest relative to the normal threshold, indicating that the method of the present invention triggers residual correction with the highest sensitivity during the critical time period when the visual and vestibular conflict is most intense, thus ensuring visual stability at the near-field depth boundary.

[0060] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A pre-rendering method based on intelligent cockpit attitude prediction, characterized in that, Includes the following steps: S1, establish a continuous instruction stream parsing module at the entrance of the rendering pipeline. In response to the detection of the turning event trigger point, parse the instruction queue after the turning point, extract the angular acceleration sequence corresponding to the target posture sequence, and extract the temporal parameters including the duration of the acceleration segment, the duration of the constant speed segment, and the duration of the deceleration segment from the angular acceleration sequence. S2, based on temporal parameters, constructs nonlinear transition coefficients that are phase-aligned with the platform's actual motion trajectory, and performs layered modulation on the nonlinear transition coefficients based on the depth buffer of the current frame to generate pixel-by-pixel depth-aware hybrid coefficients. S3 uses the pixel-per-pixel depth-aware blending coefficient as weight to perform depth-layered forward mapping synthesis on the dual-branch frames generated by the rendering pipeline and outputs a blended frame. S4, in response to the arrival of actual physical attitude feedback, performs residual verification and correction based on the threshold triggered by the dynamically tightened residual correction during the peak angular acceleration period, and performs directional hidden state reset on the prediction model based on the real inertial measurement unit data after the turning event ends.

2. The pre-rendering method based on intelligent cockpit attitude prediction according to claim 1, characterized in that, The command queue after the transition is parsed, and the angular acceleration sequence corresponding to the target attitude sequence is extracted. Temporal parameters, including the durations of the acceleration, constant velocity, and deceleration phases, are then extracted from the angular acceleration sequence, including: The target attitude sequence after the turning event trigger point is subjected to first-order difference according to the sampling period to obtain the command angular velocity sequence; the command angular velocity sequence is subjected to first-order difference to obtain the angular acceleration sequence. Extreme points and zero-crossing points are detected on the angular acceleration sequence. The time interval when the angular acceleration first reaches the peak value, the duration when the angular acceleration is maintained above the preset proportion of the peak value, and the time interval when the angular acceleration monotonically decays from the peak value to zero are determined respectively. Finally, time-domain parameters including the duration of the acceleration segment, the duration of the uniform velocity segment, and the duration of the deceleration segment are obtained.

3. The pre-rendering method based on intelligent cockpit attitude prediction according to claim 1, characterized in that, Nonlinear transition coefficients, constructed based on time-domain parameters and aligned with the phase of the platform's actual motion trajectory, include: Based on the trapezoidal velocity planning standard, the relationship between angular velocity and time in the acceleration, uniform velocity, and deceleration segments is determined. Integrating the angular velocity within each segment with respect to time yields the normalized displacement progress as a function of time. Each segment is removed and normalized to the total displacement, ultimately forming a nonlinear transition coefficient that is phase-aligned with the platform's actual motion trajectory.

4. The pre-rendering method based on intelligent cockpit attitude prediction according to claim 1, characterized in that, Based on the depth buffer of the current frame, the nonlinear transition coefficients are layered and modulated to generate pixel-by-pixel depth-aware mixing coefficients, including: Get the normalized depth value of each pixel in the depth buffer of the current frame; The near-field disparity ratio is calculated based on the normalized depth of the frustum of the near-field and far-field reference objects, and the near-field conservatism coefficient is determined based on the near-field disparity ratio. The nonlinear transition coefficients are adjusted pixel by pixel using near-field conservative coefficients and normalized depth values, ultimately generating pixel-by-pixel depth-aware hybrid coefficients.

5. The pre-rendering method based on intelligent cockpit attitude prediction according to claim 1, characterized in that, Using pixel-wise depth-aware blending coefficients as weights, depth-layered forward mapping compositing is performed on the dual-branch frames generated by the rendering pipeline to output a blended frame, including: The dual-branch frame includes a continuation branch frame and a target branch frame, and the depth map is obtained synchronously output by the continuation branch frame and the target branch frame during generation; For each pixel in the image, the corresponding pixel-wise depth-aware blending coefficient is queried based on the depth value; The depth level of a pixel is determined based on a depth boundary threshold, and occlusion-aware forward mapping or weighted averaging is performed according to the depth level to finally output a mixed frame.

6. The pre-rendering method based on intelligent cockpit attitude prediction according to claim 5, characterized in that, The depth level of a pixel is determined based on a depth boundary threshold, and occlusion-aware forward mapping or weighted averaging is performed according to the depth level, including: For near-field pixels with depth values ​​less than the depth boundary threshold, the corresponding 3D spatial points in the continuing branch frame will be reprojected onto the target view coordinate system according to the camera extrinsic parameters of the target branch frame. If the reprojected coordinates pass the depth test, then the weighted average of the two pixels is taken. If the reprojected coordinates fall within the depth test failure area of ​​the occlusion, then the pixel values ​​of the continuation branch frame are directly used to perform occlusion-aware forward mapping.

7. The pre-rendering method based on intelligent cockpit attitude prediction according to claim 5, characterized in that, The depth boundary threshold is dynamically determined by the following steps: sort the normalized depth values ​​of all pixels in the current frame's full-frame depth map in ascending order; The upper bound of the near-field proportion is determined based on the cumulative proportion corresponding to the first significant valley in the current frame depth histogram. Find the depth upper bound value corresponding to the cumulative number of pixels reaching the upper bound of the near field proportion of the total number of pixels in the whole frame, and use the depth upper bound value as the depth boundary threshold.

8. The pre-rendering method based on intelligent cockpit attitude prediction according to claim 1, characterized in that, During the peak angular acceleration period, residual verification and correction are performed based on a dynamically tightened residual correction trigger threshold, including: Obtain the maximum tolerable disparity pixel offset corresponding to the normal threshold; The single-frame disparity pixel increment is calculated based on the peak angular velocity in the current instruction stream and the near-field reference depth read from the mean of the current frame depth buffer. Based on the ratio of the maximum tolerable disparity pixel offset to the single-frame disparity pixel increment, the dynamic tightening residual correction trigger threshold is calculated. Based on the dynamically tightened residual correction trigger threshold, residual verification and correction are preferentially performed on the near-field depth boundary pixel band.

9. The pre-rendering method based on intelligent cockpit attitude prediction according to claim 1, characterized in that, After the transition event, a directional hidden state reset is performed on the prediction model based on real inertial measurement unit data, including: The transition event is considered to have completely ended when the nonlinear transition coefficient reaches its maximum normalized value. Extract the actual inertial measurement unit data corresponding to the deceleration and constant velocity phases during the new state establishment phase; The extracted real inertial measurement unit data is input into the encoder of the prediction model to update the model's hidden state and clear the residual leave state memory vector, thus completing the orientation hidden state reset.

10. The pre-rendering method based on intelligent cockpit attitude prediction according to claim 1, characterized in that, Establish a continuous instruction stream parsing module at the entrance of the rendering pipeline, including: Determine the motion controller's output cycle; The instruction stream continuous parsing module is initialized at the entrance of the rendering pipeline according to the issuance cycle.

Images