A lenticular lens naked eye 3D viewpoint switching method and system

Abstract

The application discloses a kind of cylindrical lens naked eye 3D view point switching method and system, it is related to three-dimensional display technical field.Acquire binocular position, time stamp and observation quality and align according to refreshing cycle time, obtain the central position of viewer;Based on the central position, the viewing state containing speed, acceleration and prediction reliability is constructed, the future position is predicted and the target view point is mapped;According to boundary distance, motion information, direction consistency and prediction reliability, generate boundary risk, schedule multiple view point sequences and transition weight;Interleaved diagram of old and new view point is generated, and transition output frame is synthesized through double buffer, after verification, drive display, when abnormal, the latest stable view point is rolled back and the weight is frozen, so that the boundary crosstalk and the probability of mis-switching are effectively reduced, the display continuity and switching stability are improved, and multiple types of landing scenes such as vehicle-mounted and commercial display are adapted.

Description

Technical Field

[0001] This invention belongs to the field of three-dimensional display technology, specifically relating to a method and system for switching naked-eye 3D viewpoints using a cylindrical lens. Background Technology

[0002] Lens-based naked-eye 3D displays utilize the geometric mapping between a lenticular lens array and the sub-pixel array of the display panel to rearrange and output images from different viewpoints according to a predetermined interlacing rule. This allows the viewer's left and right eyes to receive different images at different spatial positions, thus creating a binocular parallax and stereoscopic visual effect. This type of display does not rely on wearable devices and is suitable for in-vehicle central control systems, exhibition terminals, advertising screens, and interactive professional display equipment.

[0003] In this type of display scheme, the underlying key is not only the generation of multi-view images, but also the stability and continuity of the mapping relationship between the display panel sub-pixel index, the pitch of the cylindrical lens, the viewing position, and the viewpoint number. For static naked-eye 3D displays, the system typically outputs a fixed viewpoint sequence according to a fixed viewing area, and the image rearrangement mapping matrix remains unchanged over a relatively long period of time. Therefore, when the viewer's position is basically stable, a relatively stable stereoscopic display effect can be obtained. However, when the viewer moves laterally, the effective viewing areas of the left and right eyes will cross the boundaries of adjacent viewpoints. If the display end still maintains the original fixed mapping, phenomena such as enhanced crosstalk between the left and right eyes, depth flipping, edge tearing, and stereoscopic discontinuity are likely to occur. The reason is that the viewpoint index field output by the display end is essentially segmented and discrete, while the trajectory of human eye movement and changes in viewing angle are continuous time signals, and the two are discontinuously coupled on the time axis and the spatial axis.

[0004] With the increasing demand for interactive displays, existing systems are beginning to incorporate eye position detection, parallax compensation, and dynamic parameter adjustment to mitigate display mismatch issues caused by changes in viewing position. However, from a processing perspective, many solutions still treat viewpoint selection, image rearrangement, and panel refresh as separate processes: first estimating the user's position, then determining the target viewpoint, and finally refreshing the entire frame. This approach assumes that display switching can be completed instantaneously at frame boundaries, failing to explicitly model the viewpoint boundary crossing issue unique to lenticular lens displays, and also neglecting to handle energy aliasing between previous and subsequent viewpoint subpixels within the same transition time window during global image refresh. This is especially problematic in fixed viewpoint output and global hard switching technologies, where new interleaving... Figure 1Once the entire display panel is replaced, all sub-pixels will synchronously change their assigned viewpoint numbers, causing the same physical sub-pixel to correspond to different viewpoint image sampling positions in two consecutive refresh cycles. When the viewer is in the boundary transition zone, they will simultaneously receive residual light energy from the old mapping and abrupt light energy from the new mapping within a short time window. This spatial discrete mapping, combined with the hard temporal switching, will create instantaneous crosstalk peaks across viewpoints at the microscopic level. Therefore, smooth viewpoint switching in lenticular naked-eye 3D is not simply a matter of image updating, but rather a matter of coordinated control of viewpoint prediction, sequence scheduling, remapping, and transition output.

[0005] Existing technologies have explored several approaches to address the above problems, which can be mainly divided into two categories.

[0006] The first category involves viewing state prediction and adaptive adjustment of display parameters. For example, Chinese patent CN202511849343.X discloses a method for dynamically adjusting the image based on viewing conditions in adaptive naked-eye 3D. It predicts the viewpoint at the next moment based on the viewer's state, the confidence map of the previous rendering cycle, and scene saliency analysis, and then achieves dynamic display through neural light field rendering and active microlens array compensation. Chinese patent CN202510727100.2 discloses an adaptive naked-eye 3D stereoscopic display parameter dynamic adjustment system, which collects user eye position, ambient lighting, and content depth information, and uses a model to predict user movement trends, pre-calculating parallax parameters and multi-view interleaving signals. A common feature of these technologies is that they model the problem as a continuous parameter adaptation problem, and their applicability is more inclined towards systems with active optical adjustment, continuously controllable multi-parameter settings, or continuous rendering compensation.

[0007] The second category involves parallax compensation and image generation optimization. For example, Chinese patent CN202511022827.7 discloses a glasses-free 3D display method based on parallax compensation. This method calculates pixel parallax deviation values ​​by referencing the angular relationship between the viewpoint and the viewing angle, and combines neighborhood search and local pixel optimization to complete image synthesis and viewpoint generation. Dynamic compensation is performed when the parallax deviation is too large or comfort decreases. The focus of this type of technology is on global parallax modeling and local pixel correction in the image content domain; its application boundaries are mainly located at the pixel parallax and image synthesis levels.

[0008] However, the aforementioned existing technologies still have significant limitations in the naked-eye 3D viewpoint switching scenarios using lenticular lenses. The approach of continuous parameter adjustment is ill-suited to the inherent characteristics of discrete viewpoint boundary switching in lenticular lenses. Existing solutions primarily focus on continuous changes in viewpoint position, parallax parameters, or optical compensation, failing to model the causal chain of "discrete viewpoint number switching—panel global refresh—abrupt change in viewpoint affiliation for the same physical sub-pixel—crosstalk peak formation." Even if user movement trends can be predicted, it is difficult to guarantee stable visual results during switching under the lenticular lens structure. Image content domain compensation cannot eliminate the impact of global hard switching in the display mapping domain. Existing optimizations mainly improve the geometric consistency and comfort distribution of the image itself, but the main distortion during viewpoint switching is caused by the overall replacement of the interlaced mapping matrix, which is a spatiotemporal impact in the display mapping domain. Content domain compensation schemes do not provide scheduling and transition mechanisms for such global discrete switching. Meanwhile, existing technologies lack viewpoint pre-switching scheduling and safety degradation control for refresh cycles. Even if position prediction is included, it is mostly used directly for the generation of parameters in the current frame, and is not transformed into a viewpoint sequence scheduling table for future multiple refresh cycles. It is impossible to generate candidate interleaving maps, arrange transition windows and fallback paths in advance before the viewer approaches the boundary. It also lacks a clear backoff mechanism for boundary conditions such as decreased prediction reliability, eye position loss, and insufficient computing power, making it difficult to cope with the risks of interruption and crosstalk.

[0009] Therefore, how to convert continuous eye-tracking trajectories into executable discrete viewpoint switching decisions, and complete the orderly output of target viewpoint, transition viewpoint, and backup viewpoint on the refresh-level time scale of the display panel, while maintaining stable display in the event of observational and computational anomalies, has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0010] The purpose of this invention is to provide a method and system for switching viewpoints in naked-eye 3D using a cylindrical lens, in order to solve the problems in the prior art where continuous parameter adjustment schemes are difficult to handle discrete viewpoint boundary switching in naked-eye 3D using a cylindrical lens, image content domain optimization is difficult to suppress the impact of global hard switching of interleaved mapping, and there is a lack of viewpoint pre-switching and safety degradation mechanisms oriented towards the refresh cycle.

[0011] To achieve the above objectives, this invention provides a method for switching 3D viewpoints using a cylindrical lens without glasses. The method includes the following steps:

[0012] The binocular positions, timestamps, and observation quality output by the eye position acquisition device are obtained. The binocular positions are time-aligned according to the refresh cycle of the display panel to obtain the viewer's center position. The center position is reconstructed based on a stable interpupillary distance estimation when one eye is effective. Based on the time-aligned center position, a viewing state including lateral velocity, lateral acceleration and prediction confidence is constructed. The center position for multiple future refresh cycles is predicted, and the prediction results are mapped to the future target viewpoint according to the viewpoint center table and view area boundary table displayed by the cylindrical lens. Based on the distance from the current location to the nearest view area boundary, the viewer's lateral movement information, the consistency of the movement direction with the nearest view area boundary, and the prediction confidence, a boundary risk is generated. When the pre-switching conditions are met, a viewpoint sequence scheduling result covering multiple future refresh cycles is generated, and the transition weight between the old viewpoint and the new viewpoint is determined. Based on the interleaving mapping tables corresponding to the old viewpoint and the new viewpoint, respectively, an old viewpoint interleaving map and a new viewpoint interleaving map are generated. The old viewpoint interleaving map and the new viewpoint interleaving map are written into a double buffer. When the new viewpoint interleaving map is pre-generated, a transition output frame is synthesized according to the transition weight. The transition output frame is sent to the display controller, which drives the lenticular lens naked-eye 3D display panel to output after the verification is passed. When the prediction confidence is insufficient, the observation continues to fail, the interlacing map pre-generation is not completed, or the output frame verification fails, the interlacing map corresponding to the nearest stable viewpoint is fixed and the transition weight is frozen.

[0013] Specifically, the cylindrical lens naked-eye 3D viewpoint switching method of the present invention includes the following steps.

[0014] Step S1: Obtain the binocular positions, timestamps, and observation quality output by the eye position acquisition device; align the binocular positions according to the display panel refresh cycle, so that the binocular observations at the sampling time correspond to the current refresh time; when both eyes are effective, determine the viewer's center lateral position based on the left and right eye lateral positions; when only one eye is effective, reconstruct the center position based on stable interpupillary distance estimation; when both eyes are ineffective or stable interpupillary distance estimation fails, use the prediction result of the previous effective refresh cycle as the current center position. Through this step, asynchronous sampling data is mapped to a unified refresh time base, and the continuity of center position estimation is maintained when one eye fails for a short time.

[0015] Step S2 involves constructing a viewing state based on the time-aligned center position. This viewing state includes at least the current center position, lateral velocity, lateral acceleration, and prediction confidence. The lateral velocity is obtained by differentiating the center positions of the two most recent valid refresh cycles, and the lateral acceleration is obtained by second-order differentiating the center positions of the three most recent valid refresh cycles. Based on this, short-term predictions of the center positions for future refresh cycles are made using the current center position, lateral velocity, and lateral acceleration. Simultaneously, prediction confidence is generated based on the observation quality and the dispersion of the historical band sign prediction residuals, and this prediction confidence is used as a gating factor to determine whether active pre-switching is permitted. This step enables the continuous motion state of the viewer to be proactively characterized and provides a temporal basis for subsequent boundary risk assessment.

[0016] Step S3: Read the viewpoint center table and view zone boundary table obtained through cylindrical lens optical calibration from the memory; for each future predicted center position, determine the viewpoint number corresponding to the nearest viewpoint center, which serves as the target viewpoint for the corresponding future refresh cycle; simultaneously, generate a boundary risk based on the distance from the current position to the nearest view zone boundary, the intensity of lateral movement, the consistency of the movement direction with the nearest boundary direction, and the prediction confidence level. This boundary risk is not determined solely by distance or speed, but reflects the comprehensive state of whether the viewer is approaching and about to cross the view zone boundary in a credible manner.

[0017] Step S4: A candidate viewpoint set is generated based on the current viewpoint, the one-step predicted viewpoint, and the two-step predicted viewpoint. When the boundary risk reaches a preset risk threshold, or when the one-step predicted viewpoint is inconsistent with the current viewpoint and the prediction confidence reaches a preset confidence threshold and the motion direction points to the nearest viewpoint boundary, a pre-switching mode is entered. In the pre-switching mode, the next frame is not directly switched to the target viewpoint. Instead, a viewpoint sequence scheduling result covering multiple future refresh cycles is generated based on future cross-boundary stability and back-and-forth swing. Simultaneously, the transition weight of the new viewpoint relative to the old viewpoint is determined based on the boundary risk and lateral velocity to control the gradual transition from the old viewpoint to the new viewpoint. Through this step, the traditional reactive switching of the current frame is transformed into a controlled scheduling for multiple future refresh cycles.

[0018] Step S5: Read the interlacing mapping tables corresponding to the old viewpoint and the new viewpoint respectively; generate an old viewpoint interlacing map based on the viewpoint image corresponding to the old viewpoint and the old viewpoint interlacing mapping table, and generate a new viewpoint interlacing map based on the viewpoint image corresponding to the new viewpoint and the new viewpoint interlacing mapping table; write the old viewpoint interlacing map to the old cache, and write the new viewpoint interlacing map to the new cache; when all pixel blocks of the new viewpoint interlacing map have been written, the mapping task completion status is returned, and the frame size and pixel format verification passes, it is determined that the pre-generation of the new viewpoint interlacing map is complete. Through this step, the new interlacing map required for switching is prepared in advance before the display frame switching, reducing the risk of timeout caused by temporary recalculation during the refresh cycle.

[0019] Step S6: When the new viewpoint interlacing map is pre-generated, the old and new viewpoint interlacing maps are weighted and synthesized according to the transition weights to obtain a transition output frame. The transition output frame is sent to the display controller, which performs at least one of the following checks: frame size consistency, pixel format consistency, buffer completion flag, frame integrity, and pixel range. After passing the checks, the transition output frame is written to the panel driver buffer and driven to output by the lenticular lens naked-eye 3D display panel. This step reduces the energy jump caused by the direct overall replacement of the old and new maps through controlled gradation of the display mapping domain.

[0020] Step S7: After each refresh cycle, update the prediction error, residual window, and prediction confidence based on the new eye position observations; fine-tune the risk threshold online within a small range according to the pre-switching mis-triggered situation and missed switching situation, and roll back to the previous stable parameter version when the comprehensive score drops beyond the allowable range; when the prediction confidence continues to be lower than the preset threshold, the binocular observations continue to fail for more than the preset duration, the new viewpoint interlacing map is not pre-generated before the end of the current refresh cycle, or the output frame verification fails, enter the degradation mode; in the degradation mode, fix the output interlacing map corresponding to the most recently stable viewpoint and freeze the transition weights. Through this step, the system maintains acceptable display continuity even when there are observational or computational anomalies.

[0021] Furthermore, the stable pupillary distance estimation in step S1 is maintained as follows: when multiple consecutive sampling times satisfy the condition that both eyes are effective and the observation quality is not lower than a preset quality threshold, the stable pupillary distance estimation is updated according to the lateral position difference between the left and right eyes at the corresponding sampling times; when either eye is ineffective, the observation quality is lower than a preset quality threshold, or the instantaneous pupillary distance deviates from the most recent stable value by more than a preset deviation threshold, the stable pupillary distance estimation is frozen; when the duration of continuous failure exceeds a preset time, the stable pupillary distance estimation is disabled and monocular reconstruction is stopped. This method enables monocular reconstruction to be based on relatively stable physiological distance information, rather than directly relying on instantaneous abnormal observations.

[0022] Furthermore, the prediction confidence in step S2 is determined jointly by the sample variance of the historical signed prediction residuals and the observation quality; the signed prediction residual is the difference between the actual center position and the one-step prediction position of the previous refresh cycle to the current refresh cycle; the prediction confidence decreases as the residual variance increases and increases as the observation quality increases; when the average observation quality is lower than a preset threshold, the prediction confidence is set to zero. This setting ensures that active pre-switching is only allowed when the observations are reliable and the predictions are stable.

[0023] Furthermore, the viewpoint center table and view area boundary table in step S3 are obtained through cylindrical lens optical calibration. During calibration, multiple positions are sampled along the horizontal direction of the display at a standard viewing distance. A single-viewpoint test image is output for each position, and the energy distribution of the main viewpoint at that position is recorded. When the main energy proportion of a certain viewpoint reaches a local maximum, the corresponding horizontal position is recorded as the viewpoint center. When the main energy proportions of two adjacent viewpoints are approximately equal, the corresponding horizontal position is recorded as the view area boundary between the two viewpoints. This method ensures that the view area division matches the cylindrical lens pitch, tilt angle, panel resolution, and sub-pixel arrangement.

[0024] Further, the stable cross-boundary movement of future viewpoints in step S4 can be defined as: a one-step predicted viewpoint is different from the current viewpoint, and two-step predicted viewpoints are the same as the one-step predicted viewpoint; the back-and-forth swing of the target viewpoint in the next two frames can be defined as: a one-step predicted viewpoint is different from the current viewpoint, and two-step predicted viewpoints return to the current viewpoint, or there has been a record in the recent multiple refresh cycles of switching from the current viewpoint to an adjacent viewpoint and then returning to the current viewpoint. In the case of stable cross-boundary movement, a sequence of future multi-frame viewpoints containing the current viewpoint and the target viewpoint is generated; in the case of back-and-forth swing, the switching is delayed and the current viewpoint is maintained. This setting can suppress invalid flips caused by short-term jitter near the boundary.

[0025] Furthermore, the interlacing mapping table in step S5 at least includes physical sub-pixel coordinates, corresponding viewpoint numbers, source image coordinates, and color channel indices. The interlacing map generation process writes the pixel sampling results in the viewpoint image to the corresponding physical sub-pixel positions according to the mapping table. This mapping relationship directly corresponds to the actual output structure of the lenticular lens display panel.

[0026] Furthermore, the transition output in step S6 employs a pixel-by-pixel weighted synthesis of the old viewpoint interlacing map and the new viewpoint interlacing map. For implementations where the pixel data has a fixed bit width, it is preferable to first mix them in a high-precision format, then round them down and apply physical output range limitations. This method ensures that the transition output can reflect the continuous transition between the old and new viewpoints while also being compatible with the physical driving range of the display panel.

[0027] Furthermore, the most recently stable viewpoint in step S7 refers to a viewpoint that, before entering the degradation mode, has continuously output for at least a preset period, whose corresponding period prediction reliability meets the stability condition, has not experienced any output frame verification failures, and whose output viewpoint number remains unchanged. If no most recently stable viewpoint meets the conditions, the system reverts to the default center viewpoint. This method prioritizes display states that have recently been verified as stable for degradation output.

[0028] The present invention also provides a cylindrical lens naked-eye 3D viewpoint switching system, including an acquisition module, a state construction module, a scheduling module, an interleaving generation module and an output control module.

[0029] The acquisition module is used to obtain the positions of both eyes, timestamps, and observation quality output by the eye position acquisition device, align the positions of both eyes according to the refresh cycle of the display panel to obtain the center position of the viewer, and reconstruct the center position based on stable interpupillary distance estimation when one eye is effective.

[0030] The state construction module is used to construct a viewing state containing lateral velocity, lateral acceleration and prediction confidence based on the time-aligned center position. It predicts the center position for multiple future refresh cycles and maps the prediction results to the future target viewpoint according to the viewpoint center table and view area boundary table displayed by the cylindrical lens.

[0031] The scheduling module is used to generate boundary risks based on the distance from the current position to the nearest view area boundary, the viewer's lateral movement information, the consistency of the movement direction with the nearest view area boundary, and the prediction confidence. When the pre-switching conditions are met, it generates a viewpoint sequence scheduling result covering multiple future refresh cycles and determines the transition weight between the old viewpoint and the new viewpoint.

[0032] The interleaving generation module is used to generate old viewpoint interleaving maps and new viewpoint interleaving maps based on the interleaving mapping tables corresponding to the old viewpoint and the new viewpoint, respectively, and write the old viewpoint interleaving maps and new viewpoint interleaving maps into the double buffer.

[0033] The output control module is used to synthesize a transition output frame according to the transition weight when the new viewpoint interlacing map pre-generation is completed, and send the transition output frame to the display controller. The display controller drives the lenticular lens naked-eye 3D display panel to output the frame after the verification is passed. When the prediction confidence is insufficient, the observation continues to fail, the interlacing map pre-generation is not completed, or the output frame verification fails, the interlacing map corresponding to the nearest stable viewpoint is fixed and the transition weight is frozen.

[0034] Compared with existing technologies, this invention possesses multi-dimensional technical and economic advantages. Addressing the instantaneous crosstalk problem caused by the overall replacement of interlaced maps when the discrete viewpoint boundary of a cylindrical lens crosses in naked-eye 3D, a collaborative mechanism of boundary risk assessment, multi-frame viewpoint sequence scheduling, and dual-interlaced map transition output is adopted. This transforms full-frame hard switching into controlled pre-switching. By predicting cross-boundary trends in advance, pre-generating new viewpoint interlaced maps, and gradually increasing the proportion of new viewpoints according to weights, the amplitude of viewpoint abrupt changes in the same physical sub-pixel in adjacent refresh cycles is reduced, the energy abrupt change slope at the boundary position is lowered, effectively suppressing crosstalk peaks and improving viewing smoothness. To address the jitter-induced false switching problem caused by short-term back-and-forth movement near the boundary, stable cross-boundary determination, directional consistency constraints, and prediction reliability gating are introduced in the scheduling stage. Instead of directly switching based on single-frame prediction results, a viewpoint sequence is generated based on multi-frame prediction. When the viewpoint swings back and forth between the next two frames, the switching is delayed while maintaining the current viewpoint. This effectively suppresses invalid viewpoint flipping amplified by boundary noise, significantly reducing the false switching rate and visual fatigue. For abnormal operating conditions such as short-term eye position failure, sudden drop in reliability, and timeout of interlacing graph generation, a degradation control path centered on the nearest stable viewpoint is set up. When the observation or calculation conditions do not meet the reliable switching requirements, the verified stable interlacing graph output is maintained and the transition weights are frozen, thus isolating the direct impact of abnormal input on the display end, improving the effective display continuity rate in abnormal scenarios, reducing operation and maintenance troubleshooting costs, and adapting to complex commercial deployment scenarios. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the embodiments are briefly described below. The drawings are only used to assist in understanding the technical concept and process relationship of the present invention and do not constitute a limitation on the scope of protection of the present invention.

[0036] Figure 1 This is a flowchart of the overall method in an embodiment of the present invention.

[0037] Figure 2 This is a flowchart of the data acquisition and status construction unit in an embodiment of the present invention.

[0038] Figure 3 This is a flowchart of the view area boundary risk assessment and viewpoint sequence scheduling unit in an embodiment of the present invention.

[0039] Figure 4 This is a flowchart of the interlacing graph pre-generation and transition output unit in an embodiment of the present invention.

[0040] Figure 5 This is a flowchart of the online feedback update and abnormal degradation unit in an embodiment of the present invention. Detailed Implementation

[0041] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0042] In this invention, the lenticular lens naked-eye 3D display panel is the actual physical display object. The lenticular lens array is set on the light-emitting side of the display panel. Different physical sub-pixels are projected to different lateral viewing positions after being refracted or guided by the lenticular lenses. The viewer's left and right eyes are located in the viewing space in front of the display panel. The eye position acquisition device is used to measure the position of the viewer's eyes in the display coordinate system. The processor reads the viewpoint center table, view area boundary table, and interlacing mapping table obtained during the optical calibration stage. It first calculates the viewing state based on the eye position changes, and then converts the state into target viewpoints and transition output control quantities for multiple future refresh cycles. That is to say, this invention first establishes the entity mapping relationship between the display panel, lenticular lenses, viewing area, and viewpoint number, and then uses prediction and scheduling algorithms to control the switching process of this entity mapping relationship, rather than having the final display action directly determined by a black-box model.

[0043] The display coordinate system can be referenced to the plane where the center of the display panel is located, with the horizontal axis along the width of the display panel and the vertical axis along the height of the display panel. The left-eye horizontal position, right-eye horizontal position, corresponding timestamp, and observation quality output by the eye position acquisition device are all defined in this coordinate system. The viewpoint center table records the center position of each logical viewpoint in the horizontal observation space, the view area boundary table records the boundary position between adjacent logical viewpoints, and the interleaving mapping table records the correspondence between the physical sub-pixels of the display panel and the image pixels of each logical viewpoint.

[0044] Example 1: Overall flow of a method and system for switching naked-eye 3D viewpoints using a cylindrical lens

[0045] like Figures 1 to 5 As shown in the figure, the cylindrical lens naked-eye 3D viewpoint switching method and system provided in this embodiment includes seven collaborative steps: viewing data acquisition and time alignment, eye movement state construction and future position prediction, view area mapping and boundary risk assessment, viewpoint sequence scheduling, interlacing map pre-generation and dual buffer management, transition output, and online feedback update and abnormal degradation.

[0046] Step S1: Observe the data acquisition and time alignment S101. Collect raw viewing data.

[0047] This sub-step is executed collaboratively by the computer device's processor and the eye position acquisition device. The eye position acquisition device acquires the position of the viewer's left and right eyes and the observation quality according to a sampling period, forming the original observation sequence. In this embodiment, the sampling period can be... This value can be determined by the sampling capability of the eye position acquisition device and the real-time requirements of the display system. The observation vector at each sampling moment can be expressed as: ,

[0048] in, Indicates the first The observation vector at each sampling time; and These represent the horizontal and vertical positions of the left eye in the monitor's coordinate system, respectively, both in millimeters; and These represent the horizontal and vertical positions of the right eye in the monitor's coordinate system, respectively, both in millimeters; It represents the observation quality score at the current sampling time. It is a dimensionless quantity and can range from 0 to 1.

[0049] When any eye position is occluded and cannot be effectively identified, the corresponding position component is marked as invalid. This sampling result does not directly enter the main prediction link but is transferred to the subsequent missing data compensation process. When the observation quality is lower than a preset quality threshold, the current observation is preferably used only for anomaly statistics and not for main state updates. In this embodiment, the preset quality threshold can be exemplified as follows: This threshold can be obtained through offline testing statistics in occluded scenarios.

[0050] S102, Perform time alignment according to the refresh cycle and calculate the center position.

[0051] This sub-step is executed by the processor. Since the refresh rate of the display panel is typically different from the eye position sampling period, it is necessary to align the observations from the most recent sampling periods to the current refresh time. In this embodiment, the display panel refresh rate can be... This is used to correspond to a 60 Hz refresh rate; the processor can align the eye position observations from the three most recent sampling periods to the current refresh time. Linear interpolation can be used for continuous position quantities, and last value preservation can be used for observation quality to avoid unnecessary smoothing of discrete quality scores.

[0052] When both eyes are effective, the processor calculates the lateral position of the center of both eyes based on the lateral position of the left and right eyes. The calculation relationship is as follows: ,

[0053] in, Indicates the first The horizontal position of the center of both eyes at each sampling time is shown in millimeters. and The meaning is the same as in the previous text.

[0054] To maintain the center position estimation even in the case of monocular loss, the processor maintains a stable interpupillary distance estimate. When multiple consecutive sampling times satisfy the conditions of binocular validity and observation quality not lower than the update threshold, the lateral position difference between the left and right eyes at the corresponding sampling times is used to update the stable pupillary distance estimate. In this embodiment, the number of consecutive samples can be 4, and the update threshold can be [missing information]. All of these can be obtained offline through a large number of observation samples under stable gaze scenarios. If the conditions are met for four consecutive sampling times, the stable pupillary distance estimate can be updated according to the moving average. Conversely, when either eye is invalid, the observation quality is below the threshold, or the instantaneous pupillary distance deviates from the most recent stable value by more than the deviation threshold, the stable pupillary distance estimate is frozen. In this embodiment, the deviation threshold can be set to... This value can be determined based on the statistical results of short-term fluctuations in adult interpupillary distance.

[0055] When only the left eye is effective and the stable interpupillary distance (IPD) estimate is valid, the processor reconstructs the binocular center positions based on the effective eye position and the stable IPD estimate. When only the right eye is effective, a symmetrical reconstruction method is used. If the duration of consecutive failures exceeds the failure time threshold, the stable IPD estimate is deemed to have failed, and monocular reconstruction is disabled. In this embodiment, the failure time threshold can be set to... The reliability of monocular reconstruction can be determined offline based on continuous occlusion conditions. If both eyes are invalid or the stable interpupillary distance estimation fails, the center position of this cycle will use the prediction result of the previous valid refresh cycle, and an anomaly flag will be set for subsequent reliability calculations.

[0056] S103, Median Filtering and Outlier Removal.

[0057] This sub-step is executed by the processor. To eliminate instantaneous jump points, a sliding window median filter is applied to the center position sequence. In this embodiment, the window length can be 5, a value determined by a trade-off between time delay and anomaly resistance. When the number of valid samples in the window is less than 3, it can degenerate into taking the median value of the valid samples. The velocity estimation does not directly perform median processing on the filtered sequence again, but preferably uses the difference of adjacent valid refresh samples to reduce phase lag. When the displacement between two consecutive sampling points exceeds a preset displacement threshold and the observation quality decreases beyond a preset quality reduction threshold, the sampling point can be marked as an anomaly and removed from the difference calculation. In this embodiment, the displacement threshold can be taken as... The quality degradation threshold can be set to 0.25. After this processing, the refresh level center position is obtained. This serves as input for constructing the subsequent viewing state.

[0058] Step S2: Eye-tracking state construction and future location prediction S201, Construct state variables.

[0059] This sub-step is executed by the processor. The processor constructs three types of state variables: directly observable states, estimated physical states, and risk assessment-related states. Directly observable states include the current eye center positions. The current average longitudinal position of both eyes and the average observation quality; estimated physical state including lateral velocity. and lateral acceleration Risk assessment-related statuses include boundary proximity factor and false trigger rate. Leakage switching rate and prediction credibility The prediction confidence level represents the reliability of the current prediction result for proactive pre-switching decisions; it is derived from the quality score within the current observation window and the one-step prediction residual in the historical refresh cycle. False trigger rate. The missed handover rate is used to characterize the proportion of systems that, after entering a pre-handover phase, do not actually cross the boundary effectively and instead return to the original viewpoint. This is used to characterize the proportion of systems that did not enter the pre-switching phase but subsequently experienced actual boundary crossings and target viewpoint lag. These state variables are collectively used for subsequent boundary risk assessment, online fine-tuning, and parameter rollback decisions.

[0060] S202, Predict the position of multiple future refresh cycles.

[0061] This sub-step is executed by the processor. The processor estimates the lateral velocity and lateral acceleration based on the difference results of the most recent valid samples in the refresh level sequence. In this embodiment, let the three most recent valid refresh times be... , and The corresponding center positions are respectively , and Furthermore, the interval between adjacent refreshes is approximately equal to the refresh cycle. Lateral velocity can be calculated based on the difference between the center positions of the two most recent effective refresh cycles, and lateral acceleration can be calculated based on the second-order difference between the center positions of the three most recent effective refresh cycles.

[0062] In one alternative implementation, the future... The center position of each refresh cycle is predicted using a short-term acceleration model, and the calculation relationship is as follows: ,

[0063] in, This indicates that the prediction will proceed backward from the current refresh cycle. The horizontal position of the center of both eyes per refresh cycle, in millimeters; This indicates the horizontal position of the center of both eyes during the current refresh cycle, in millimeters. This indicates the estimated horizontal velocity during the current refresh cycle, expressed in millimeters per second. This indicates the estimated lateral acceleration during the current refresh cycle, expressed in millimeters per square second. Indicates the refresh cycle, in seconds; This represents the number of prediction steps, which is a positive integer. It represents the proportionality coefficient of the constant acceleration displacement term.

[0064] This formula corresponds to the processing step S2, which involves "short-term prediction of the center position for multiple future refresh cycles." Its inputs are the current center position, lateral velocity, lateral acceleration, and refresh cycle. The output is the predicted center position for several future refresh cycles, which is then used in the viewpoint mapping step.

[0065] When the effective observation length is insufficient to calculate velocity or acceleration, the corresponding value is set to a preset value. In this embodiment, when the effective observation length is less than two refresh cycles, the lateral velocity can be set to zero; when the effective observation length is less than three refresh cycles, the lateral acceleration can be set to zero. If the sample participating in the difference is a hold value obtained only by extrapolation from both eyes and is invalid, then the sample does not participate in the acceleration estimation. When the prediction result exceeds the allowable observation range of the display, it is cropped to the boundary. In this embodiment, the effective lateral display width can be exemplified as follows: .

[0066] In this embodiment, the prediction time domain length can be 3 frames, corresponding to a look-ahead window of approximately 3 frames. This length can cover one interleaving graph preload and two transition outputs.

[0067] S203. Calculate the prediction reliability.

[0068] This sub-step is executed by the processor. The processor first calculates the signed prediction residual, which is the difference between the actual center position and the predicted position from the previous refresh cycle to the current refresh cycle. Then, it calculates the variance of this residual over the most recent multiple refresh cycle windows and combines this variance with the average observation quality to calculate the prediction confidence. In one optional implementation, the following relationship can be used: ,

[0069] in, Indicates the first The prediction reliability of each refresh cycle is a dimensionless quantity. This represents the sample variance of the signed predicted residuals in the most recent residual window, in square millimeters. This represents the variance of the baseline residuals, in square millimeters. This indicates a smoothing term to prevent the denominator from being zero; This represents the average observation quality of the current window, and is a dimensionless quantity. This indicates exponentiation.

[0070] This formula corresponds to the processing step S2, "generating prediction confidence and using it as an active pre-switching gating value." The output is the prediction confidence, which is then directly used in boundary risk calculation and pre-switching trigger determination.

[0071] When there are insufficient valid residual samples within the window, the residual variance can be set to the baseline value; when more than half of the residual samples come from monocular reconstruction data, the prediction confidence can be attenuated; when the average observation quality is below the threshold, the prediction confidence is directly set to zero. In this embodiment, the average observation quality threshold can be set to 0.35, the attenuation coefficient can be set to 0.85, and the baseline residual variance can be set to... Smoothing terms are acceptable. .

[0072] After this step, the system obtains the current state vector, the future position prediction result, and the prediction confidence level, and subsequent modules can perform view area boundary judgment on the same time base.

[0073] Step S3: Viewport Mapping and Boundary Risk Assessment S301. Read and calibrate the associated view area parameters.

[0074] This sub-step involves the processor reading the lenticular lens viewing area parameter dictionary from memory, including a viewpoint center table and a viewing area boundary table. This parameter table is obtained through factory optical calibration, which first determines the physical entity and observation position: the display panel is fixedly assembled with the lenticular lens array, and a calibration camera or optical probe is positioned laterally along the display at a standard observation distance. Subsequently, single-viewpoint test images are sequentially output at multiple lateral sampling positions, recording the percentage of primary viewpoint energy received by the left and right eyes.

[0075] When the principal energy percentage of a certain viewpoint reaches a local maximum, that lateral position is recorded as the center of that viewpoint; when the principal energy percentages of two adjacent viewpoints are approximately equal, that lateral position is recorded as the boundary between the two viewpoints. To reduce the impact of noise, each sampling position can be repeatedly measured and averaged. In this embodiment, the standard observation distance for the in-vehicle central control scene can be taken as... The standard viewing distance for vertical screen displays is acceptable. The absolute value threshold for the difference in principal energy percentage between adjacent viewpoints can be set to 0.02.

[0076] In this embodiment, the number of logical viewpoint areas can be 8, which is determined by the pitch of the cylindrical lens, the tilt angle, the panel resolution, and the target viewing range.

[0077] S302, Mapping target viewpoint number.

[0078] This sub-step is executed by the processor. For any predicted center location, the processor searches the viewpoint center table for the viewpoint center with the closest lateral distance and sets the corresponding number as the target viewpoint. In one alternative implementation, this mapping relationship can be represented as: ,

[0079] in, Indicates the prediction of the first The target viewpoint number that should be output during each refresh cycle; Indicates the index that minimizes the target value within the parentheses; Indicates the logical viewpoint number index; Indicates the total number of logical viewpoints; Indicates the prediction of the first The center horizontal position of each refresh cycle, in millimeters; Indicates the first The horizontal coordinates of the center of each logical viewpoint are in millimeters. This represents absolute value operations.

[0080] This formula corresponds to the processing step in step S3, "mapping the prediction result to future target viewpoints." The output result is the target viewpoint number for multiple future refresh cycles, which is then used in the scheduling step.

[0081] When the distances between the centers of two viewpoints and the predicted position are exactly equal, it is preferable to select the same number as the current output viewpoint; when the predicted position exceeds the range of the leftmost or rightmost viewpoint center, it is clamped to the boundary viewpoint respectively.

[0082] S303, Calculate boundary proximity, directional consistency, and boundary risk.

[0083] This sub-step is executed by the processor. The processor first determines the nearest view area boundary and the distance from the current position to that boundary based on the distance between the current position and the boundaries of each view area. Next, it determines whether the lateral movement direction points to the nearest boundary, thus obtaining directional consistency. Finally, the boundary proximity, velocity intensity, directional consistency, and prediction confidence are combined to form the boundary risk.

[0084] In one alternative implementation, the boundary risk index can be expressed as: ,

[0085] in, Indicates the first The boundary risk index for each refresh cycle is a dimensionless quantity. Indicates the boundary proximity factor; This represents the absolute value of the estimated lateral velocity, expressed in millimeters per second. This indicates the normalized upper limit of velocity, expressed in millimeters per second. Indicates the smoothing term; Indicates the weight of directional consistency; Indicates the directional consistency factor; This indicates the reliability of the prediction.

[0086] This formula corresponds to the "generate boundary risk" step in step S3. The output is the boundary risk index, which is directly used for pre-switching determination in step S4.

[0087] When the current position is far from the boundary, the boundary proximity factor tends to a lower value; when the velocity is close to zero, the velocity term tends to a lower value; when the prediction confidence is zero, the overall risk index is directly zero. In this embodiment, the cutoff distance for boundary risk calculation can be taken as... The upper limit of speed normalization can be taken as follows: The directional consistency weight can be taken as 0.35, and the smoothing term can be taken as... .

[0088] Step S4: Viewpoint sequence scheduling S401, Generate a set of candidate viewpoints.

[0089] This sub-step is executed by the processor. The processor generates a set of candidate viewpoints using the current viewpoint, the one-step predicted viewpoint, and the two-step predicted viewpoint, and removes duplicate numbers.

[0090] In this embodiment, "stable cross-boundary future viewpoint" can be defined as: the one-step predicted viewpoint is different from the current viewpoint, and the two-step predicted viewpoint is the same as the one-step predicted viewpoint; "back and forth swing" can be defined as: the one-step predicted viewpoint is different from the current viewpoint, and the two-step predicted viewpoint returns to the current viewpoint, or a situation has been recorded in the last 3 refresh cycles where the viewpoint switched from the current viewpoint to an adjacent viewpoint and then returned to the current viewpoint.

[0091] S402, Entering the pre-switching determination.

[0092] This sub-step is executed by the processor. The system enters pre-switching mode when one of the following conditions is met: the boundary risk reaches a preset risk threshold; or the predicted viewpoint in one step is inconsistent with the current viewpoint, the prediction confidence reaches a preset confidence threshold, and the direction of motion points to the nearest boundary; or the future viewpoint stably crosses the boundary and there is a record close to the boundary in the most recent adjacent refresh cycle. If none of the conditions are met, the current viewpoint output is maintained.

[0093] In this embodiment, the risk threshold can be set to 0.42, and the confidence threshold can be set to 0.60.

[0094] S403, Generate future multi-frame viewpoint sequences.

[0095] This sub-step is executed by the processor. After entering pre-switching mode, the system does not directly switch the next frame to a one-step predicted viewpoint, but instead generates a sequence of future multi-frame viewpoints based on the boundary risk level and the stability of future viewpoints.

[0096] For example, in one alternative implementation, when the risk is at a medium level, a sequence of "current viewpoint - target viewpoint - target viewpoint" can be used; when the risk is high and the future viewpoint is stable, a more forward target viewpoint sequence can be used; when the predicted viewpoint swings back and forth in the next two frames, the sequence degenerates into a delayed switching structure based on the current viewpoint; when a one-step prediction is still equal to the current viewpoint but the two-step prediction has undergone a stable change and the confidence is high enough, a candidate interleaving map can be pre-generated and not switched in the next frame.

[0097] S404, Calculate the transition weights.

[0098] This sub-step is executed by the processor. When the current refresh cycle is scheduled to transition from the old viewpoint to the new viewpoint, the processor determines the new viewpoint weights based on boundary risk and lateral velocity. In one alternative implementation, the following relationship can be used: ,

[0099] in, This represents the output weight of the new viewpoint in the current refresh cycle. It is a dimensionless quantity with a value range of 0 to 1. Indicates range-restricted operations; This represents the boundary risk weighting coefficient; Indicates the first Boundary risk index for each refresh cycle; This represents the speed weighting coefficient; This represents the absolute value of the lateral velocity, expressed in millimeters per second. This indicates the normalized upper limit of velocity, expressed in millimeters per second. This indicates the smoothing term.

[0100] This formula corresponds to the processing step "determining the transition weights between the old and new viewpoints" in step S4. The output is the transition weights, which are directly used for the synthesis of the transition output frames in step S6.

[0101] When the system is not in pre-switching mode Set to zero; when speed is unavailable, the speed term is set to zero, and only the risk term dominates; truncation is performed when the result exceeds the range. In this embodiment, the risk weight coefficient can be 0.70, and the speed weight coefficient can be 0.30.

[0102] After step S4, the system obtains the scheduling results of future multi-frame viewpoint sequences and the transition weight of the current frame.

[0103] Step S5: Interleaving graph pre-generation and dual buffer management S501, Read the mapping table between the current viewpoint and the target viewpoint.

[0104] This sub-step involves the processor reading the interlacing map table from memory. The interlacing map table is a structured parameter table obtained by calibrating the cylindrical lens pitch, tilt angle, panel resolution, and RGB subpixel arrangement. This table includes at least the physical subpixel coordinates, corresponding viewpoint numbers, source image coordinates, and color channel indices.

[0105] S502, Generate old viewpoint interlacing diagram and new viewpoint interlacing diagram.

[0106] This sub-step is executed by the processor. For any viewpoint to be output, the processor reads the corresponding view frame from the viewpoint image buffer, and then rearranges the source pixels to the target physical sub-pixel positions according to the interleaving mapping table to form an interleaved output frame. If the panel resolution is... The data dimension of the interleaved output frame is then... , where 3 represents the number of color channels.

[0107] The process takes a viewpoint image and a corresponding interleaving map as input and outputs an interleaving map that can be directly sent to the display controller. The old viewpoint interleaving map corresponds to the currently stable output viewpoint, while the new viewpoint interleaving map corresponds to the candidate switching target viewpoint. Since the new viewpoint interleaving map is pre-generated before switching, the display system does not need to re-compile the entire interleaving calculation within the critical refresh cycle when officially switching to the transition output.

[0108] In this embodiment, for example: the front The scheduling decision is completed within [timeframe], and subsequent [timeframes] will not exceed [number]. Complete the generation of the new interlacing diagram, and reserve no less than [amount missing]. Used for driving transmission.

[0109] S503, write to the double cache and check the cache completion flag.

[0110] This sub-step is executed by the processor. The processor maintains an old cache and a new cache; the old cache stores the old viewpoint interleaving map, and the new cache stores the new viewpoint interleaving map. When a potential crossover between the next two frames is predicted, the processor pre-generates the new cache during the current frame's display.

[0111] To determine whether the pre-generation of the new viewpoint interlacing graph is complete, a cache completion flag can be set. When all pixel blocks of the new interlaced map have been written to the target video memory address, the mapping task queue returns a completion status, and the frame size and format verification pass, the buffer completion flag is set. If the current state is not completed, it is considered complete; otherwise, it is considered incomplete. If the current state is not completed by the end of the current refresh cycle, it is considered that the pre-generation is incomplete and a degradation strategy is triggered in the subsequent output stage.

[0112] Step S6: Transition Output S601, Generate transition output frames by weighted mixing.

[0113] This sub-step is executed by the processor. Given that the new viewpoint interlacing map has already been pre-generated, the processor performs pixel-by-pixel weighted synthesis of the old and new viewpoint interlacing maps according to the transition weights obtained in step S4, forming the final output frame. In one optional implementation, this transition output relationship can be expressed as: ,

[0114] in, Indicates the first Each refresh cycle actually outputs interleaved frames to the display controller; This indicates an old viewpoint interleaved frame, which corresponds to the old viewpoint number currently being displayed; This indicates a new viewpoint interleaving frame, which corresponds to the new viewpoint number to be interleaved; Indicates the weight of the new viewpoint in the current refresh cycle; This indicates the weight of the old viewpoint.

[0115] This formula corresponds to the processing step "synthesizing the transition output frame according to the transition weight" in step S6. The output result is a transition output frame, which is directly sent to the display controller.

[0116] When implementing 8-bit pixels, it is preferable to first perform pixel-by-pixel blending in floating-point or fixed-point high-precision format, then round to the nearest integer according to the rounding rules, and finally limit the result to the range allowed by the panel. In this embodiment, the pixel range can be 0 to 255 for example. If the old and new viewpoint frame sizes are inconsistent, the system prohibits blending and enters degradation mode.

[0117] S602, Perform mixed verification and write to the display controller driver refresh.

[0118] This sub-step is executed collaboratively by the display controller and the processor. After receiving the transition output frame, the display controller performs a check, which may include at least one of the following: frame size consistency check, pixel format consistency check, buffer completion flag check, frame integrity check, and pixel range check.

[0119] Frame integrity verification can employ at least one of the following methods: line count comparison, block completion flag comparison, or CRC check. If CRC check is used, the checksum of the entire frame byte stream must match that of the generating end. Pixel range verification can use random sampling or full-frame statistics to determine whether the output value is within the allowed range. Failure of any check indicates a failed output frame verification.

[0120] Upon successful verification, the display controller writes the transition output frame to the panel driver buffer and drives the lenticular lens naked-eye 3D display panel to output according to the refresh cycle. If the next refresh cycle is still in the transition state, the updated transition weights are used to generate the next transition output frame; otherwise, it switches to a pure target viewpoint frame.

[0121] Step S7: Online feedback updates and abnormal degradation S701. Calculate the prediction error and update the residual window.

[0122] This sub-step is executed by the processor after each refresh cycle. The processor reads the new round of eye position observations, calculates the error between the current actual center position and the predicted position in the previous step, and updates the residual window and prediction confidence. Here, the absolute prediction error is used for online statistics and threshold fine-tuning, while the signed residual sample variance is still used for prediction confidence calculation to preserve the system's sensitivity to prediction offset direction and fluctuation stability.

[0123] S702, statistical analysis of false triggers and missed switches, and fine-tuning of thresholds.

[0124] This sub-step is executed by the processor. Within the continuous evaluation window, the system statistically analyzes the proportion of false pre-switching and missed switching, and makes minor online adjustments to the risk threshold based on the statistical results.

[0125] In this context, a false pre-switching occurs when the system enters pre-switching mode and no actual boundary crossing occurs within a certain number of consecutive refresh cycles, and the final output viewpoint returns to the original viewpoint. A missed switch occurs when the system does not enter pre-switching mode, but an actual boundary crossing occurs subsequently, resulting in a lag in the target viewpoint. In this embodiment, if the false triggering rate exceeds 12% within 20 consecutive refresh cycles, the risk threshold can be increased by 0.03; if the missed switch rate exceeds 10% consecutively, the risk threshold can be decreased by 0.03; the total adjustment range is preferably limited to... Inside.

[0126] To maintain the publicly disclosed training and deployment closed-loop chain during the disclosure phase, this embodiment further employs a method that separates offline parameter candidate range determination from online fine-tuning. Specifically, in the offline phase, the candidate parameter range is determined through historical trajectory playback and grid scanning. These candidate parameters include at least a risk threshold, a confidence threshold, a boundary risk weight coefficient, a velocity weight coefficient, and a cutoff distance and normalization upper limit related to boundary risk calculation. In the online phase, only minor adjustments to the risk threshold are allowed, while other parameters remain within the stable candidate set determined offline. This avoids stability degradation caused by simultaneous drift of multiple parameters during operation.

[0127] In addition, the system maintenance online comprehensive score This is used to determine whether the current parameter version has significantly degraded compared to the previous stable parameter version. The comprehensive score... The normalized crosstalk peak statistic and the false trigger ratio can be considered comprehensively. Leakage switching ratio And effective stereoscopic display continuity. The change in the score of the current parameter version relative to the reference stable version can be expressed as: ,

[0128] in, Indicates the percentage decrease in relative score; This indicates the reference comprehensive score corresponding to the previous stable parameter version; This indicates the overall score corresponding to the current parameter version.

[0129] When the overall score within the continuous evaluation window decreases by more than a preset percentage relative to the previous stable version, it automatically rolls back to the previous stable parameter version. In this embodiment, this decrease percentage can be 8%. Through the closed-loop mechanism of offline calibration—online fine-tuning—version rollback described above, consistency is maintained between parameter sources, online adaptation, and safe recovery.

[0130] S703, fix the output of the nearest stable viewpoint and freeze the transition weights.

[0131] This sub-step is executed by the processor. Degradation mode is entered when any of the following occurs: prediction confidence is below the threshold for multiple consecutive refresh cycles; binocular observations continue to fail for more than the duration threshold; the new viewpoint interleaving map is not pre-generated before the end of the current refresh cycle; or output frame verification fails.

[0132] In this embodiment, the prediction confidence threshold can be set to 0.35, and the number of consecutive periods can be set to 2; the threshold for the duration of continuous failure of binocular observation can be set to... .

[0133] In degradation mode, the system outputs the interleaving graph corresponding to the most recently stable viewpoint and freezes the transition weights to zero. The most recently stable viewpoint is defined as the viewpoint that has had continuous output for at least a preset period before entering degradation mode, whose corresponding period prediction reliability meets the stability condition, has not experienced any output frame verification failures, and whose output viewpoint number remains unchanged. In this embodiment, the continuous output period can be at least two refresh periods, and the stability condition threshold can be 0.60. If no most recently stable viewpoint meets the conditions, the system reverts to the default center viewpoint.

[0134] Example 2: Horizontal cross-viewing zone pre-switching in naked-eye 3D display using a cylindrical lens in vehicle central control

[0135] This embodiment is applied to a naked-eye 3D display screen in a vehicle's central control system. The display screen has an effective horizontal display width of 600 mm, a panel resolution of 1920×1080, and forms 8 logical viewpoint areas after lenticular lens optical calibration. The eye position acquisition device is installed on the upper edge of the central control system, with a sampling period of 8 ms; the panel refresh period is 16.67 ms; the prediction time domain length is 3 frames; the median filter window length is 5; the boundary risk calculation cutoff distance is 18 mm; the velocity normalization upper limit is 220 mm / s; the risk threshold is 0.42; the confidence threshold is 0.60; and the transition weight coefficients are 0.70 and 0.30, respectively. All of the above parameters can be obtained through offline calibration, where the viewpoint center and boundary parameters are obtained through optical calibration at a standard viewing distance of 700 mm, and the velocity upper limit is derived from the 95th percentile velocity statistics of lateral head movement under vehicle seating conditions.

[0136] In this embodiment, the computer device may employ a quad-core automotive-grade processor, with the GPU used for interlaced graph mixing output, and dual buffers used to store old viewpoint interlaced frames and new viewpoint interlaced frames respectively.

[0137] Within one refresh cycle, assuming the time-aligned left eye position is 286 mm and the right eye position is 350 mm, with an average observation quality of 0.92, the lateral position of the binocular center is 318.00 mm. Further, assuming the center positions of the two most recent effective refresh cycles are 315.90 mm and 313.92 mm respectively, the processor calculates the current lateral velocity as approximately 125.97 mm / s and the lateral acceleration as approximately 431.81 mm / s² based on the velocity-acceleration relationship in Example 1. Based on this state, the center positions for the next three frames are predicted to be approximately 320.16 mm, 322.44 mm, and 324.84 mm for the first, second, and third frames, respectively.

[0138] If the center of the 4th viewpoint is 312 mm and the center of the 5th viewpoint is 336 mm, and the boundary between them is 324 mm, then the current viewpoint is 4, and the nearest viewpoint in the next two frames will both tend to be 5. The distance from the current position to the nearest boundary is 6.00 mm, corresponding to a relatively high boundary proximity. Further, if the four most recent signed prediction residuals are -6.0 mm, -4.0 mm, 1.0 mm, and 3.0 mm, then according to the prediction confidence calculation logic in Example 1, the current prediction confidence is approximately 0.70. Since the nearest boundary is located to the right of the current position and the lateral velocity is positive, the direction of motion points to the nearest boundary, and the direction consistency is established. Therefore, the processor calculates the current boundary risk as approximately 0.27 based on the boundary risk calculation relationship.

[0139] Although the boundary risk has not yet reached the risk threshold of 0.42, the second trigger condition for pre-switching is met because the predicted viewpoint is different from the current viewpoint, the prediction confidence is above 0.60, and the direction of motion points to the nearest boundary. The system then enters pre-switching mode. At this time, the viewpoint sequence for the next three frames generated by the processor can be "4-5-5". Furthermore, based on the transition weight calculation relationship, the weight of the new viewpoint in the current refresh cycle is approximately 0.36.

[0140] Under this weighting, the processor reads the interleaving maps corresponding to viewpoints 4 and 5 from memory, generates old viewpoint interleaving maps and new viewpoint interleaving maps respectively, and writes them to the double buffer. Then, the processor synthesizes a transition output frame according to the old viewpoint weight of 0.64 and the new viewpoint weight of 0.36. After completing size, format, integrity, and pixel range checks, the display controller sends this transition output frame to the display panel. If the predicted center position in the next refresh cycle moves further to the right or exceeds the boundary, the new viewpoint weight continues to increase; when the future viewpoint stabilizes at 5, the system switches to pure viewpoint 5 output.

[0141] Compared with the traditional approach of "switching the entire frame once the current position crosses the boundary", this embodiment does not abruptly change the viewpoint affiliation of all physical sub-pixels from viewpoint 4 to viewpoint 5 within a single critical frame. Instead, it first confirms the cross-boundary trend through prediction of the next three frames, and then gradually increases the output proportion of the target viewpoint through double buffering and transition weights. Therefore, it can reduce the instantaneous energy jump received at the boundary observation position.

[0142] This embodiment employs a digital twin simulation for verification: the simulation object is an 8-viewpoint in-vehicle central control cylindrical lens display system; the simulation duration is 60 seconds; the observer's lateral movement range is ±42 mm, with an average speed of 98 mm / s, and multiple rapid cross-boundary movements are superimposed in the middle stage; the control scheme is a fixed-viewpoint global hard switching scheme. Evaluation metrics include normalized crosstalk peak value, average switching settling time, and the number of visible discomfort recordings. Results show that the mean normalized crosstalk peak value of the control scheme is 0.31 with a standard deviation of 0.03, while the mean value of the present invention's scheme is 0.18 with a standard deviation of 0.02; the average switching settling time of the control scheme is 83 ms, while that of the present invention is 49 ms; the average number of visible discomfort recordings within 60 seconds is 14 in the control scheme and 8 in the present invention. These results demonstrate that, in cross-viewpoint scenarios, the present invention can reduce the instantaneous visual impact caused by hard switching.

[0143] Subsequently, based on the above calculation and processing results, the processor continues to execute the subsequent steps as described in Embodiment 1, such as online feedback updates, abnormal degradation judgment, and parameter rollback control. The specific execution logic is the same as in Embodiment 1, and will not be repeated here.

[0144] Example 3: Boundary jitter and abnormal degradation control in naked-eye 3D display terminals

[0145] This embodiment is applied to a commercial display-based naked-eye 3D vertical screen terminal. In this scenario, viewers often make short-distance lateral movements in front of the screen, and may experience temporary loss of vision in one or both eyes due to multiple people obstructing the view. Although the traditional frame-by-frame nearest-viewpoint direct switching scheme can respond quickly to position changes, it is prone to frequent switching when jittering near the boundary; if switching is still based on unstable observation results when the eye position fails briefly, intermittent interruptions are likely to occur.

[0146] This embodiment follows the overall processing flow and main parameter system of Embodiment 1. However, due to the tighter budget for graphics computing resources in the exhibition scene, the double-buffered pre-generated window is fixed at 2 frames, the standard viewing distance is set to 900 mm, and the viewpoint center and view area boundary table are recalibrated accordingly. The observation quality threshold is set to 0.35, and the threshold for continuous binocular failure duration is set to 33 ms, which were calibrated through occlusion experiments and continuous frame drop scenario tests, respectively.

[0147] Assume that at a certain moment, the center positions of both eyes in three consecutive refresh cycles are 323.00 mm, 325.00 mm, and 323.50 mm, respectively, with a boundary position of 324.00 mm. If a frame-by-frame nearest viewpoint direct switching scheme is adopted, the system will continuously flip between viewpoint 4 and viewpoint 5. However, in this embodiment, the system does not switch immediately based solely on the current nearest viewpoint number, but instead makes a delay decision by combining prediction confidence, directional consistency, and consistency between the next two frames.

[0148] For example, within an abnormal period, assuming the current center position is 323.40 mm, the estimated velocity is 38.00 mm / s, the estimated acceleration is -120.00 mm / s², the average observation quality is 0.41, and the last four signed residuals are -8.0 mm, 5.0 mm, 7.0 mm, and 2.0 mm, respectively. According to the prediction confidence calculation logic in Example 1, the current prediction confidence is approximately 0.21. Since this value is below 0.35, the system directly prohibits active pre-switching. Even if the predicted position has crossed the boundary in one step, the boundary risk will be suppressed to a low level due to confidence gating, and the system continues to maintain the current stable viewpoint 4 output, with the transition weight frozen to zero.

[0149] If both eyes fail simultaneously for more than 33 ms in the next refresh cycle, a degradation mode is triggered, and the interleaving map corresponding to the most recent stable viewpoint 4 is fixedly output, without attempting to generate a new transitional output. Furthermore, if the new interleaving map pre-generation task is not completed on time due to GPU congestion during the degradation period, the system continues to maintain stable viewpoint output according to the degradation conditions in Example 1 until the observation recovers and the prediction confidence recovers above the stable threshold for two consecutive refresh cycles, the target viewpoints in the next two frames are consistent, and the hybrid verification passes. Only then is the degradation lifted and the system re-enters the pre-switching process.

[0150] This embodiment also underwent digital twin simulation verification: the simulation duration was 10 seconds; the normal working condition was a single person's slow, lateral probing movement; the impact working condition was back-and-forth jitter near the boundary, superimposed with two binocular occlusions and one GPU pre-generation delay; the control scheme was frame-by-frame direct switching of the nearest viewpoint without abnormal degradation; the evaluation indicators included the number of viewpoint switches within 10 seconds, the false switch rate, the effective stereoscopic display continuity rate, and the number of visible interruptions. The results showed that the average number of viewpoint switches in the control scheme was 23, while the present invention's scheme was 9; the false switch rate in the control scheme was 17.6%, while the present invention's scheme was 6.8%; the effective stereoscopic display continuity rate in the control scheme was 82.4%, while the present invention's scheme was 95.1%; the average number of visible interruptions in the control scheme was 11, while the present invention's scheme was 3. These results indicate that the present invention can maintain more stable naked-eye 3D display continuity even when boundary jitter and abnormal observations occur simultaneously.

[0151] Subsequently, based on the above calculation and processing results, the processor continues to execute the subsequent steps as described in Example 1, such as online false trigger statistics, risk threshold fine-tuning, abnormal degradation retention and recovery exit judgment, etc. The specific execution logic is the same as in Example 1, and will not be repeated here.

[0152] Example 4: Electronic Equipment

[0153] This embodiment provides an electronic device for implementing a lenticular lens naked-eye 3D viewpoint switching method and system as described in any of the above embodiments. The electronic device includes a processor, a memory, a communication interface, and a computer program stored in the memory and executable on the processor.

[0154] The processor, when executing the computer program, performs the following operations: acquiring the binocular positions, timestamps, and observation quality output by the eye position acquisition device; aligning the binocular positions according to the display panel refresh cycle to obtain the viewer's center position, and reconstructing the center position based on a stable interpupillary distance estimate when one eye is active; constructing a viewing state including lateral velocity, lateral acceleration, and prediction confidence based on the time-aligned center position; predicting the center position for multiple future refresh cycles; and mapping the prediction results to the future target viewpoint according to the viewpoint center table and view zone boundary table displayed by the cylindrical lens; and based on the distance from the current position to the nearest view zone boundary, the viewer's lateral motion information, the consistency of the motion direction with the nearest view zone boundary, and the prediction confidence. A boundary risk is generated, and a viewpoint sequence scheduling result covering multiple future refresh cycles is generated when the pre-switching conditions are met. The transition weights between the old and new viewpoints are determined. An old viewpoint interleaving map and a new viewpoint interleaving map are generated based on the interleaving mapping tables corresponding to the old and new viewpoints, respectively. The old viewpoint interleaving map and the new viewpoint interleaving map are written into a double buffer, and a transition output frame is synthesized according to the transition weights when the new viewpoint interleaving map is pre-generated. The transition output frame is sent to the display controller, which drives the lenticular lens naked-eye 3D display panel to output after the verification is passed. When the prediction confidence is insufficient, the observation continues to fail, the interleaving map pre-generation is not completed, or the output frame verification fails, the interleaving map corresponding to the most recent stable viewpoint is fixed and the transition weights are frozen.

[0155] The memory can be used to store program instructions, viewpoint center table, view area boundary table, interleaving mapping table, historical parameter versions, and cache completion flags generated during operation. Overall score False trigger rate Leakage switching rate And statistical results. The communication interface is used to interact with the eye position acquisition device, display controller, or external calibration device.

[0156] Example 5: Computer-readable storage medium

[0157] This embodiment provides a non-transitory computer-readable storage medium storing a computer program or computer instructions. When the computer program or computer instructions are executed by a processor, they implement the steps of the cylindrical lens naked-eye 3D viewpoint switching method and system described in any one of embodiments one to three above.

[0158] The computer-readable storage medium may include a disk storage device, an optical storage device, a flash memory, a read-only memory, a random access memory, or other media capable of storing program instructions.

[0159] It should be noted that relational terms such as "first" and "second" in the text are only used to distinguish different objects, steps, or states, and do not indicate an actual order, primary or secondary relationship, or necessary dependency. Terms such as "include" and "contain" are open-ended qualifiers and do not exclude the existence of other elements, steps, or components not explicitly listed.

[0160] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the principles and spirit of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for switching naked-eye 3D viewpoints using a cylindrical lens, characterized in that, include: The binocular positions, timestamps, and observation quality output by the eye position acquisition device are obtained. The binocular positions are time-aligned according to the refresh cycle of the display panel to obtain the viewer's center position. The center position is reconstructed based on a stable interpupillary distance estimation when one eye is effective. Based on the time-aligned center position, a viewing state including lateral velocity, lateral acceleration and prediction confidence is constructed. The center position for multiple future refresh cycles is predicted, and the prediction results are mapped to the future target viewpoint according to the viewpoint center table and view area boundary table displayed by the cylindrical lens. Based on the distance from the current location to the nearest view area boundary, the viewer's lateral movement information, the consistency of the movement direction with the nearest view area boundary, and the prediction confidence, a boundary risk is generated. When the pre-switching conditions are met, a viewpoint sequence scheduling result covering multiple future refresh cycles is generated, and the transition weight between the old viewpoint and the new viewpoint is determined. Based on the interleaving mapping tables corresponding to the old viewpoint and the new viewpoint, respectively, an old viewpoint interleaving map and a new viewpoint interleaving map are generated. The old viewpoint interleaving map and the new viewpoint interleaving map are written into a double buffer. When the new viewpoint interleaving map is pre-generated, a transition output frame is synthesized according to the transition weight. The transition output frame is sent to the display controller, which drives the lenticular lens naked-eye 3D display panel to output after the verification is passed. When the prediction confidence is insufficient, the observation continues to fail, the interlacing map pre-generation is not completed, or the output frame verification fails, the interlacing map corresponding to the nearest stable viewpoint is fixed and the transition weight is frozen.

2. The method for switching naked-eye 3D viewpoints using a cylindrical lens according to claim 1, characterized in that, The process of aligning the positions of both eyes according to the refresh cycle of the display panel to obtain the viewer's center position, and reconstructing the center position based on a stable interpupillary distance estimation when one eye is effective, includes: Time alignment is performed on the eye positions from the most recent multiple sampling periods so that the eye positions correspond to the current refresh time; When both eyes are effective, calculate the horizontal position of the center of both eyes based on the horizontal position of the left eye and the horizontal position of the right eye; When only one eye is effective and the stable interpupillary distance estimation is effective, the transverse position of the binocular center is reconstructed based on the effective eye position and the stable interpupillary distance estimation. When both eyes are invalid or the stable interpupillary distance estimation fails, the prediction result of the previous effective refresh cycle is used as the current center position.

3. The method for switching naked-eye 3D viewpoints using a cylindrical lens according to claim 2, characterized in that, The stable interpupillary distance estimation reconstruction includes: When a preset number of consecutive sampling times all satisfy the condition that both eyes are effective and the observation quality is not lower than a preset quality threshold, the stable pupillary distance estimate is updated based on the difference in the lateral position of the left and right eyes at the corresponding sampling time. When either eye is invalid, the observation quality is below a preset quality threshold, or the instantaneous pupillary distance deviates from the most recent stable value by more than a preset deviation threshold, the stable pupillary distance estimation is frozen. When the duration of continuous failure exceeds a preset time, the stable interpupillary distance estimation is rendered invalid, and monocular reconstruction is disabled.

4. The method for switching naked-eye 3D viewpoints using a cylindrical lens according to claim 1, characterized in that, The construction includes a viewing state comprising lateral velocity, lateral acceleration, and prediction confidence, and predicts the center position for multiple future refresh cycles, including: Calculate the lateral velocity based on the difference in the center positions of the two most recent effective refresh cycles; The lateral acceleration is calculated using the second-order difference based on the center position of the three most recent effective refresh cycles. Based on the current center position, lateral velocity, and lateral acceleration, short-term predictions are made of the center position for multiple future refresh cycles according to the refresh cycle. When there are insufficient valid samples to calculate lateral velocity or lateral acceleration, the corresponding values ​​will be set to preset values.

5. The method for switching naked-eye 3D viewpoints using a cylindrical lens according to claim 4, characterized in that, The prediction reliability is determined by the observation quality and the sample variance of the historical band sign prediction residuals, and serves as a gating factor for whether active pre-switching is allowed. The signed prediction residual is the difference between the actual center position and the one-step prediction position of the previous refresh cycle to the current refresh cycle. The prediction reliability decreases as the sample variance of the historical signed prediction residual increases and increases as the observation quality increases. When the average observation quality is below a preset threshold, the prediction confidence level is set to zero.

6. The method for switching naked-eye 3D viewpoints using a cylindrical lens according to claim 1, characterized in that, The process of mapping the prediction results to future target viewpoints and generating boundary risks based on the viewpoint center table and view area boundary table displayed by the cylindrical lens includes: Read the viewpoint center table and view zone boundary table obtained through cylindrical lens optical calibration from memory; For any predicted center location, determine the viewpoint number corresponding to the viewpoint center closest to it as the target viewpoint; Calculate the distance from the current position to the nearest viewpoint boundary, and determine directional consistency based on whether the viewer's movement direction points to the nearest viewpoint boundary; Boundary risks are generated based on the distance, lateral velocity, directional consistency, and prediction confidence.

7. The method for switching naked-eye 3D viewpoints using a cylindrical lens according to claim 6, characterized in that, The process of generating viewpoint sequence scheduling results covering multiple future refresh cycles and determining the transition weights between old and new viewpoints includes: Generate a candidate viewpoint set using the current viewpoint, the one-step predicted viewpoint, and the two-step predicted viewpoint; When the boundary risk reaches the preset risk threshold, or when the one-step predicted viewpoint is inconsistent with the current viewpoint and the prediction confidence reaches the preset confidence threshold and the movement direction points to the nearest viewpoint boundary, the pre-switching mode is entered. After entering the pre-switching mode, when the future viewpoint stably crosses the boundary, a future multi-frame viewpoint sequence containing the current viewpoint and the target viewpoint is generated; When the target viewpoint swings back and forth between the next two frames, delay the switch and maintain the current viewpoint; Among them, a stable crossover of future viewpoints means that the one-step prediction viewpoint is different from the current viewpoint, and the two-step prediction viewpoint is the same as the one-step prediction viewpoint. The target viewpoint swings back and forth between the next two frames if the predicted viewpoint is different from the current viewpoint and the predicted viewpoint returns to the current viewpoint after two steps, or if there has been a record of switching from the current viewpoint to an adjacent viewpoint and then returning to the current viewpoint in the most recent refresh cycles. The new viewpoint weights are determined based on boundary risk and lateral velocity, serving as the transition weights for the transition from the old viewpoint to the new viewpoint.

8. The method for switching naked-eye 3D viewpoints using a cylindrical lens according to claim 1, characterized in that, The process of generating old viewpoint interleaving maps and new viewpoint interleaving maps based on the interleaving mapping tables corresponding to the old and new viewpoints respectively, writing the old and new viewpoint interleaving maps into a double buffer, and synthesizing a transition output frame according to the transition weights when the new viewpoint interleaving map is pre-generated includes: Read the interleaving mapping table corresponding to the old viewpoint and the interleaving mapping table corresponding to the new viewpoint from the memory. The interleaving mapping table contains at least physical sub-pixel coordinates, corresponding viewpoint numbers, source image coordinates, and color channel indexes. An old viewpoint interlacing map is generated based on the viewpoint image corresponding to the old viewpoint and the interlacing mapping table corresponding to the old viewpoint; a new viewpoint interlacing map is generated based on the viewpoint image corresponding to the new viewpoint and the interlacing mapping table corresponding to the new viewpoint. Write the old viewpoint interlacing map into the old cache, write the new viewpoint interlacing map into the new cache, and determine that the pre-generation of the new viewpoint interlacing map is complete when all pixel blocks have been written, the mapping task completion status is returned, and the frame size and pixel format verification pass. The old viewpoint interlacing map and the new viewpoint interlacing map are weighted and synthesized pixel by pixel according to the transition weight to obtain the transition output frame.

9. The method for switching naked-eye 3D viewpoints using a cylindrical lens according to claim 8, characterized in that, The process involves sending the transition output frame to the display controller, which then drives the lenticular lens naked-eye 3D display panel to output the image after successful verification. In cases of insufficient prediction confidence, persistent observation failures, incomplete interlacing graph pre-generation, or output frame verification failure, the process fixes the interlacing graph corresponding to the nearest stable viewpoint and freezes the transition weights, including: The display controller performs at least one of the following checks on the transition output frame: frame size consistency check, pixel format consistency check, buffer completion flag check, frame integrity check, and pixel range check; When the verification passes, the transition output frame is written to the panel driver buffer and driven to output by the lenticular lens naked-eye 3D display panel; When the prediction reliability is lower than the preset threshold for consecutive preset periods, the binocular observation fails for more than the preset duration, the new viewpoint interlacing map is not pre-generated before the end of the current refresh period, or the output frame verification fails, the degradation mode is entered. In degradation mode, the interlacing graph corresponding to the nearest stable viewpoint is fixed, and the transition weights are frozen to zero. Among them, the most recently stable viewpoint is the viewpoint that has continuously output for no less than a preset period before entering the degradation mode, whose prediction confidence of the corresponding period meets the stability condition, has not experienced any output frame verification failure, and whose output viewpoint number remains unchanged. If no nearest stable viewpoint meets the criteria, then fall back to the default center viewpoint.

10. A cylindrical lens naked-eye 3D viewpoint switching system, characterized in that, It includes an eye position acquisition device, a processor, a memory, a display controller, and a lenticular lens naked-eye 3D display panel; The eye position acquisition device is used to acquire the viewer's eye position, timestamp, and observation quality; the memory is used to store the viewpoint center table, view zone boundary table, interlacing mapping table, and program instructions executed by the processor. The processor is connected to the eye position acquisition device, the memory and the display controller, and is used to execute the program instructions to implement the cylindrical lens naked-eye 3D viewpoint switching method as described in any one of claims 1 to 9, including: aligning the positions of the two eyes according to the display panel refresh cycle to obtain the viewer's center position, and reconstructing the center position based on a stable interpupillary distance estimation when one eye is effective; Based on the time-aligned center position, a viewing state including lateral velocity, lateral acceleration, and prediction confidence is constructed. The center position for multiple future refresh cycles is predicted, and the prediction results are mapped to future target viewpoints according to the viewpoint center table and view area boundary table displayed by the cylindrical lens. Based on the distance from the current position to the nearest view area boundary, the viewer's lateral motion information, the consistency of the motion direction with the nearest view area boundary, and the prediction confidence, a boundary risk is generated. When the pre-switching conditions are met, a viewpoint sequence scheduling result covering multiple future refresh cycles is generated, and the transition weight between the old viewpoint and the new viewpoint is determined. An old viewpoint interlacing map and a new viewpoint interlacing map are generated based on the interlacing mapping tables corresponding to the old viewpoint and the new viewpoint, respectively. The old viewpoint interlacing map and the new viewpoint interlacing map are written into a double buffer, and a transition output frame is synthesized according to the transition weight when the new viewpoint interlacing map is pre-generated. The display controller is connected to the processor and the lenticular lens naked-eye 3D display panel, and is used to receive the transition output frame. After the verification is passed, it drives the lenticular lens naked-eye 3D display panel to output. When the prediction confidence is insufficient, the observation continues to fail, the interlacing map pre-generation is not completed, or the output frame verification fails, it cooperates with the processor to fix the output of the interlacing map corresponding to the nearest stable viewpoint and freeze the transition weight.

Images