An ar scene low-delay triggering system based on track displacement induction

Abstract

The application belongs to the technical field of image data processing, and relates to an AR scene low-delay triggering system based on track displacement induction, which comprises a time sequence feature vector generation module, a physical interaction mode determination module, a dynamic resource stream list generation module, a loading pressure vector construction module, a resource stream list correction module, a resource asynchronous injection module and a delay rendering output module. The application solves the problem of how to predict the micro movement intention of a user and dynamically schedule resources in combination with system bottom layer running load to realize dynamic loading and low-delay triggering of a track type augmented reality scene.

Description

Technical Field

[0001] This invention belongs to the technical field of image data processing and relates to a low-latency triggering system for AR scenes based on orbital displacement sensing. Background Technology

[0002] Augmented reality technology has been applied to track-based interactive scenarios such as museum guided tours and industrial assembly. Users can view overlaid virtual information at specific locations by moving the display devices on the track. However, visual assets such as high-precision 3D models and texture maps are large in size, while the computing power and storage bandwidth of the track display devices are relatively limited. How to present high-resolution virtual scenes with low latency when the device moves to a preset trigger point under limited hardware resources is the main technical challenge currently faced.

[0003] Chinese patent publication number CN121330235A discloses a 3D dynamic scene display system for cultural relics based on virtual-real fusion technology. This prior art generates environmental feature vectors by collecting signals from the cultural relics and their surrounding environment, combines twin simulation with multi-temporal model resource packages to generate scene script sequences, and monitors rendering performance and interactive behavior to dynamically adjust the displayed content. This solution achieves, to a certain extent, the dynamic fusion display of scene materials and real-time adjustment of rendering parameters.

[0004] While the aforementioned existing technologies and similar solutions can dynamically adjust the displayed content, they still have limitations when dealing with the complex and ever-changing physical interactions of users in track-based scenarios. On the one hand, existing resource preloading strategies are usually based on macroscopic spatial region divisions, with coarse-grained predictions, making it difficult to accurately distinguish whether a user is moving quickly or lingering at the region boundaries. This can easily lead to incorrect loading and waste of memory and bandwidth resources. On the other hand, on-demand loading mechanisms based on precise point triggers have inherent response latency. The process from the system detecting the trigger signal to completing resource reading and loading into video memory takes a long time, which can easily cause visual stuttering or screen interruption for the user, weakening the immediacy of the augmented reality experience.

[0005] In summary, how to predict users' micro-movement intentions and combine them with the underlying system load to dynamically schedule resources in order to achieve dynamic loading and low-latency triggering of track-based augmented reality scenarios is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0006] To address the aforementioned problems, this invention provides a low-latency triggering system for AR scenes based on orbital displacement sensing.

[0007] A low-latency triggering system for AR scenes based on orbital displacement sensing includes: The time-series feature vector generation module acquires the original position signal of the track display device, performs differential conversion on the original position signal, and obtains the time-series feature vector. The physical interaction mode determination module performs waveform feature extraction and pattern matching on the temporal feature vector to obtain the target intent vector. The dynamic resource flow list generation module performs asset decomposition and priority calculation on scene asset data based on the target intent vector to obtain an initial resource loading list; Load the pressure vector construction module, obtain the underlying system operation status indicators, perform weighted fusion calculation on the operation status indicators, and obtain the real-time loading pressure vector; The resource loading list correction module performs scheduling strategy adjustment, verification and trimming on the initial resource loading list based on the real-time loading pressure vector to obtain the corrected resource loading list. The asynchronous resource injection module performs asynchronous preloading and cache reclamation based on the revised resource loading list, and writes the corresponding resource assets into the high-speed video memory system; The deferred rendering output module compares the actual physical location with the target scene trigger preset points on-site based on the comparison results, calls the resource assets in the high-speed video memory system to perform rendering operations, and obtains a three-dimensional virtual image.

[0008] A further aspect of this invention involves acquiring the original position signal of the track display device, performing differential conversion on the original position signal to obtain a time-series feature vector, and including the following steps: Acquire real-time pulse-coded data and angle and attitude change parameters sent by the track display device to generate a basic motion dataset; Filtering and noise reduction processing and continuous derivative calculation are performed on the basic motion dataset to obtain velocity data and first-order acceleration data; Perform time difference conversion on the first-order acceleration data to extract the jerk signal, and encapsulate the jerk signal into a time-series feature vector.

[0009] A further aspect of this invention involves performing waveform feature extraction and pattern matching on a time-series feature vector to obtain a target intent vector, including the following steps: Monitor the jerk values ​​contained in the timing feature vector, compare the jerk values ​​with the preset silent noise limit, and confirm the starting point of the instantaneous thrust. When the jerk value exceeds the preset baseline threshold, a single set of changing bands is extracted, and the spectral energy and peak distribution characteristics of the single set of changing bands are extracted. By comparing the peak distribution characteristics with a pre-calibrated preset dynamic feature library, the physical interaction mode is identified, and the target intent vector is generated based on the identification results.

[0010] A further aspect of this invention involves performing asset decomposition and priority calculation on scene asset data based on the target intent vector to obtain an initial resource loading list, including the following steps: Deconstruct the 3D spatial mesh geometry data contained in the scene asset data and divide it into fine-grained resource data blocks; Obtain the arrival probability of the target intent vector marker, and combine it with the occupied volume of the fine-grained resource data block to calculate the expected usage window duration and initial prediction priority of the fine-grained resource data block; The initial resource loading list is generated by concatenating parameters such as fine-grained resource data blocks, initial prediction priorities, and expected usage window duration.

[0011] A further aspect of this invention involves obtaining the arrival probability of the target intent vector marker, combining it with the occupied volume of the fine-grained resource data block, and calculating the expected usage window duration and initial prediction priority of the fine-grained resource data block, including the following steps: Obtain the basic priority and preset display duration of fine-grained resource data blocks, as well as the estimated average resource loading rate; The arrival probability is multiplied by the base priority to generate the initial prediction priority; The loading time is obtained by dividing the occupied volume by the average resource loading rate. The loading time is then added to the preset display duration to generate the expected usage window duration.

[0012] A further aspect of this invention involves obtaining underlying system operation status indicators, performing weighted fusion calculations on these indicators to obtain a real-time loading pressure vector, and includes the following steps: Send a polling command to read the actual read rate of the storage medium at the communication transmission port; Record the physical processing latency of the graphics rendering bus and calculate the latency deviation ratio according to the expected baseline latency deviation. A weighted fusion operation is performed on the ratio of the actual read rate of the storage medium to the latency deviation to generate a real-time load pressure vector.

[0013] A further aspect of this invention involves performing a weighted fusion operation on the actual read rate of the storage medium and the latency deviation ratio to generate a real-time loading pressure vector, including the following steps: Obtain the preset nominal peak read rate and weighting coefficient; The ratio of the actual read rate to the nominal peak read rate of the storage medium is calculated, and the rate component is calculated based on the ratio. The rate component and the latency deviation ratio are weighted and fused using a preset weighted fusion function to generate a real-time load pressure vector.

[0014] A further aspect of this invention involves adjusting and verifying the scheduling strategy of the initial resource loading list based on the real-time loading pressure vector to obtain a revised resource loading list, including the following steps: Analyze the system load ratio represented by the real-time load pressure vector and compare the system load ratio with the safe load warning threshold. When the system load ratio exceeds the safe load warning threshold, the extraction priority of fine-grained resource data blocks that meet the conditions identified in the initial resource loading list is reduced. Perform validation and pruning on the initial resource loading list after rearrangement, remove asset reading tasks that exceed the expected usage window duration, and generate a corrected resource loading list.

[0015] A further aspect of this invention involves performing asynchronous preloading and cache reclamation based on the revised resource loading list, writing the corresponding resource assets into the high-speed video memory system, including the following steps: Parse the scheduling instruction vector inside the corrected resource loading list to trigger the block-based fetch instruction; Allocate high-speed video memory buffer blocks to the high-speed video memory system, and write the retrieved fine-grained resource data blocks into the corresponding video memory physical address space. Based on the actual motion displacement of the device and the target intent vector, the failure prediction waypoint is detected, and the video memory buffer occupied by the resources associated with the failure prediction waypoint is released.

[0016] A further aspect of this invention involves comparing the actual physical location with the target intent vector to trigger preset points on-site, then calling upon resource assets within the high-speed video memory system to perform rendering operations to obtain a 3D virtual image, including the following steps: Compare the absolute encoded coordinates of the track equipment sensors corresponding to the actual physical location with the preset points of the target scene trigger site, and extract the position matching trigger signal; Verify the readiness status of fine-grained resource data blocks corresponding to the scene in the physical address space of video memory; The ready fine-grained resource data blocks are fed into the hardware rendering pipeline to perform shading and layout, and output a three-dimensional virtual image.

[0017] In summary, the present invention has the following beneficial technical effects: 1. By acquiring the raw position signal from the track display device, differential conversion is performed to extract the jerk signal to generate a time-series feature vector. Furthermore, the peak distribution characteristics of a single set of changing wavebands are extracted and matched with a pre-calibrated preset dynamic feature library to obtain the target intent vector. This setup can capture the transient characteristics of the user's applied thrust from the dimension of micromechanical changes. Compared to traditional regional prediction methods based on macroscopic displacement or velocity, this invention can effectively distinguish whether the user's push-pull operation has a clear direction or is hesitant and wandering in a local position. Therefore, the system can output target waypoint information with prediction probability weights, providing accurate decision-making basis for subsequent resource preloading, reducing invalid resource loading and memory occupation caused by prediction deviations, and solving the resource waste problem caused by the excessively coarse granularity of traditional macroscopic predictions.

[0018] 2. The target virtual scene is decomposed into fine-grained resource data blocks. Initial prediction priorities are calculated based on the target intent vector. Simultaneously, the ratio of the actual read rate of the storage medium at the communication transmission port to the latency deviation of the rendering bus is obtained. These factors are then fused to generate a real-time loading pressure vector. When the system load ratio exceeds limits, the extraction priority of asset blocks that meet the conditions is dynamically reduced. This setting transforms the overall scene loading task into micro-level, independently sortable asset block loading tasks, and introduces closed-loop feedback on the actual carrying capacity of the underlying system into the scheduling decision. When hardware input / output channels are under high load, the system can proactively implement degradation strategies, prioritizing the loading of the core geometric skeleton. This allows the resource loading strategy to adapt to changes in user behavior and hardware status in real time, improving the core resource loading efficiency under limited bandwidth.

[0019] 3. Based on the revised resource loading list, asynchronous preloading of the high-speed video memory system and reclamation of abandoned predictive waypoint cache are performed. When generating a position matching trigger signal by comparing the actual physical location with the preset trigger point, the readiness of fine-grained resource data blocks is verified before rendering. This setup decouples data reading from the main rendering process through an asynchronous injection mechanism and utilizes a dynamic reclamation strategy to achieve cyclical streaming utilization of limited video memory space, alleviating the response lag problem caused by the traditional precise point-based on-demand loading mechanism. Furthermore, the system verifies the readiness of the basic geometric model and color values ​​before rendering, establishing a basic data integrity verification mechanism for the final output. This maintains low-latency triggering while avoiding clipping or screen interruptions caused by incomplete readiness of fragmented rendering blocks, ensuring the continuity and stability of the orbital augmented reality experience. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. The drawings are used to provide a further understanding of the present invention.

[0021] Figure 1 This discloses a schematic diagram of the framework in the embodiments of this application.

[0022] Figure 2 This discloses a flowchart of an embodiment of this application.

[0023] Figure 3 The present application discloses a bar chart comparing the predicted priorities of different types of assets before and after the resource flow list was revised, as well as a line chart showing the expected usage window duration.

[0024] Figure 4 This application discloses a curve showing the trigger determination logic relationship between distance error, spatial tolerance threshold, and position matching trigger signal status as the absolute coded coordinates of the device changes. Detailed Implementation

[0025] The following is in conjunction with the appendix Figure 1 - Figure 4 A preferred description of the present invention is provided below.

[0026] See attached document Figure 1 - Figure 2 This invention proposes a low-latency triggering system for AR scenes based on orbital displacement sensing, comprising the following modules: The time-series feature vector generation module acquires the original position signal of the track display device, performs differential conversion on the original position signal, and obtains the time-series feature vector. The physical interaction mode determination module performs waveform feature extraction and pattern matching on the temporal feature vector to obtain the target intent vector. The dynamic resource flow list generation module performs asset decomposition and priority calculation on scene asset data based on the target intent vector to obtain an initial resource loading list; Load the pressure vector construction module, obtain the underlying system operation status indicators, perform weighted fusion calculation on the operation status indicators, and obtain the real-time loading pressure vector; The resource loading list correction module performs scheduling strategy adjustment, verification and trimming on the initial resource loading list based on the real-time loading pressure vector to obtain the corrected resource loading list. The asynchronous resource injection module performs asynchronous preloading and cache reclamation based on the revised resource loading list, and writes the corresponding resource assets into the high-speed video memory system; The deferred rendering output module compares the actual physical location with the target scene trigger preset points on-site based on the comparison results, calls the resource assets in the high-speed video memory system to perform rendering operations, and obtains a three-dimensional virtual image.

[0027] In one embodiment of the present invention, the time-series feature vector generation module is configured to perform the following steps: Acquire real-time pulse-coded data and angle and attitude change parameters sent by the track display device to generate a basic motion dataset; Filtering and noise reduction processing and continuous derivative calculation are performed on the basic motion dataset to obtain velocity data and first-order acceleration data; Perform time difference conversion on the first-order acceleration data to extract the jerk signal, and encapsulate the jerk signal into a time-series feature vector.

[0028] The execution unit of the time-series feature vector generation module is a central processing unit deployed inside or connected to the track display device. The central processing unit has a built-in data acquisition module that concurrently calls and continuously acquires real-time pulse-coded data streams sent by the position sensors built into the track display device and synchronously integrated inertial measurement equipment at a preset sampling frequency.

[0029] The sampling frequency is set to 100Hz to 500Hz. The setting of the sampling frequency is based on empirical judgment of the mechanical characteristics when humans manually operate the equipment. If the sampling frequency is lower than the range, it may not be able to capture the initial mechanical pulse generated by the intention. If the sampling frequency is higher than the range, redundant data will be generated, increasing the computational load.

[0030] The inertial measurement device can be a microelectromechanical system chip that integrates a three-axis gyroscope and a three-axis accelerometer, used to calculate the attitude angle of the device in real time, and provide a stream of angle and attitude change parameters acquired from the virtual camera perspective for subsequent scene rendering.

[0031] The pulse-coded data stream represents the linear displacement of the device with the pulse count value per unit time. The angle and attitude change parameter stream is usually in quaternion or Euler angle format, used to represent the pitch, yaw and roll attitude of the device during operation. The data acquisition module aligns these two data streams with a unified high-precision timestamp to generate a basic motion dataset containing multi-dimensional spatial position and state information.

[0032] The signal processing module retrieves the basic motion dataset and uses a Kalman filter to fuse and filter the pulse-coded data and angle attitude data in the dataset to eliminate sensor inherent noise and environmental vibration interference, generating a smooth instantaneous displacement sequence.

[0033] Since the denominator of subsequent higher-order time difference operations is a minimum value Therefore, it naturally amplifies the low-amplitude, high-frequency noise in the original signal nonlinearly. This pre-processing Kalman smoothing filter ensures the accelerator... The signal possesses the necessary prerequisites for physical validity and usability. The signal processing module performs continuous numerical differentiation calculations on the smoothed instantaneous displacement sequence to obtain velocity data and first-order acceleration data characterizing the current movement state of the device. The process of continuous differentiation calculation can be expressed by the following formula:

[0034] In the formula, Representative at the The instantaneous velocity at each sampling moment, in mm / s; Representative at the The first-order acceleration at each sampling time is expressed in mm / s². Representative at the The linear displacement values ​​after filtering and calibration at each sampling time point are expressed in mm, with their subscripts... The numbers represent consecutive historical sampling moments preceding the current moment. The raw data of these historical sampling moments comes from the pulse count values ​​collected by the position sensor and is obtained by multiplying them by a pre-calibrated distance coefficient. For example, each pulse represents a displacement of 0.1 mm. The specific value of this distance coefficient is inferred from physical parameters such as the mechanical transmission ratio of the track display device and the nominal line count of the hardware encoder. This represents the sampling time interval, and its value is the reciprocal of the sampling frequency. If the sampling frequency is 100Hz, then... The value is 0.01s, which is the absolute mathematical reciprocal of the selected 100Hz sampling frequency. The setting of the sampling time interval balances the real-time performance of data processing with the accuracy of capturing motion details.

[0035] The feature extraction module receives first-order acceleration data and performs second-order high-frequency time difference transformation on it. Physically, second-order high-frequency time difference transformation is equivalent to differentiating the first-order acceleration data, thereby extracting the acceleration signal reflecting the characteristics of instantaneous force changes. Specifically, it involves obtaining the higher-order difference value characterizing the third derivative of displacement with respect to time, which reflects the rate of change of force acting on the object. The process of second-order high-frequency time difference transformation can be expressed by the following formula:

[0036] In the formula, Representative at the The acceleration signal values ​​at each sampling time are the higher-order differential values. The above differential operation naturally amplifies the rapidly changing components in the signal and suppresses DC and low-frequency components. Therefore, the calculation behavior of the second-order high-frequency time differential conversion itself realizes the function of the high-pass filter module. The process of second-order high-frequency time differential conversion can effectively amplify the signal components of transient force changes caused by user manual interaction or track start and stop.

[0037] The feature extraction module extracts the calculated continuous acceleration signal stream according to a preset window length, for example, the most recent N sampling points. The window length N is a key design parameter, usually between 50 and 100. If the sampling frequency is 100Hz, N is 100, which includes the change history within the past 1 second. The aim is to fully capture a typical push-pull interaction action and avoid excessive computational burden. It encapsulates the signal into a fixed-dimensional vector to generate a time-series feature vector that can characterize the law of change of transient interaction force for subsequent processing.

[0038] For example, suppose the track display device is stationary, and at some point the user begins to push the device by hand. The sampling frequency is set to 100Hz, i.e. The distance coefficient is 0.1 mm per pulse. The data acquisition module operates over 10 consecutive sampling periods (from... arrive The original pulse count sequence obtained is [0,0,0,5,20,45,80,115,150,185].

[0039] Convert pulse count values ​​into displacement sequences The unit is mm, resulting in [0,0,0,0.5,2.0,4.5,8.0,11.5,15.0,18.5] mm. The signal processing module calculates first-order acceleration based on the displacement sequence, for example, in... At any given moment, first-order acceleration mm / s². This allows for the calculation of the first-order acceleration sequence over the entire time period.

[0040] Feature extraction module in Calculate jerk at all times It is necessary to use The displacement value, its value is .calculate Acceleration of moments .exist At any moment, accelerate. You can see at the instant the movement begins to A large-amplitude accelerometer signal peak was generated. If the time-series feature vector dimension N is set to 5, then in The time-series feature vector generated at time step is... The calculation result sequence and the time-series feature vector record the transient interaction force change law from the beginning of the force application to the process of stabilization.

[0041] In one embodiment of the present invention, the physical interaction mode determination module is used to perform the following steps: Monitor the jerk values ​​contained in the timing feature vector, compare the jerk values ​​with the preset silent noise limit, and confirm the starting point of the instantaneous thrust. When the jerk value exceeds the preset baseline threshold, a single set of changing bands is extracted, and the spectral energy and peak distribution characteristics of the single set of changing bands are extracted. By comparing the peak distribution characteristics with a pre-calibrated preset dynamic feature library, the physical interaction mode is identified, and the target intent vector is generated based on the identification results.

[0042] The physical interaction mode determination module is also executed by the central processing unit. The intent recognition module of the central processing unit receives the timing feature vector generated by the step timing feature vector generation module and performs the following physical interaction mode determination and target intent vector. The waveform monitoring unit in the intent recognition module continuously monitors the real-time updated jerk value in the timing feature vector and continuously compares its amplitude with the preset silent noise limit stored at the bottom layer of the system. The purpose of the preset silent noise limit is to filter out benign signal fluctuations caused by minor unevenness of the track or slight vibration of the equipment itself. The value of the preset silent noise limit is usually set to 5000mm / s³ to 15000mm / s³. The lower limit of this range is determined by the basic noise floor of the microelectromechanical system chip, and the upper limit is determined by the extreme value of the small vibration caused by the mechanical play of the track itself. Non-interactive mechanical vibration interference can be filtered out through this range. It is determined by multiplying the standard deviation of the signal collected over a long period of time when the equipment is stationary by a safety factor.

[0043] When the absolute value of the jerk is below the preset silent noise limit at multiple consecutive sampling points, the intent recognition module determines that the state is stationary or uniform. When the jerk value at any sampling point first exceeds the preset silent noise limit, the intent recognition module marks that sampling point as the starting point of the instantaneous thrust. At this time, the waveform capture unit is triggered, and the waveform capture unit checks whether the jerk value has further exceeded a higher preset baseline threshold.

[0044] The preset baseline threshold is set to 30,000 mm / s³ to 80,000 mm / s³. This baseline threshold range is derived from statistical distribution charts of numerous real human push-pull interaction tests. Its lower limit is higher than the system's maximum resting amplitude, and its upper limit is slightly lower than the average initial impulse of normal adult push-pull operations, aiming to distinguish between unconscious slight touches and intentional push-pull operations. Once the accelerometer value exceeds the preset baseline threshold, a valid interaction event is confirmed, and a single set of changing waveforms of a fixed duration is extracted from the starting point.

[0045] The fixed duration is usually set to 300ms to 800ms. According to statistics based on ergonomic characteristics, a typical manual push-pull intention, from muscle exertion and equipment acceleration to the smoothing of the movement, usually falls within this time range. This fully covers the main stages of a typical push-pull movement from exertion and acceleration to uniform speed or deceleration, forming a temporary acceleration waveform segment.

[0046] The intent recognition module calls the built-in wavelet analysis module, which can use basis functions such as Daubechies wavelet or Morlet wavelet to perform multi-scale decomposition on the temporarily stored accelerometer waveform segment, thereby obtaining its spectral energy in different frequency bands. The wavelet energy calculation process can be expressed by the following formula:

[0047] In the formula, The waveform segment representing acceleration is in the first... Spectral energy at each decomposition scale; This represents the discrete wavelet transform performed on the accelerometer waveform segment, at the th... The first decomposition scale Wavelet coefficients; This represents the total number of wavelet coefficients at each decomposition scale. By calculating and normalizing the energy at different decomposition scales, the energy distribution vector can be obtained. The energy distribution vector, along with dimensions such as the scale of the principal energy, the number of significant peaks, the peak values ​​and their positions, together constitute the key peak distribution characteristics.

[0048] The pattern matching unit extracts the peak distribution features and performs similarity calculations with various standard fingerprint templates stored in a pre-calibrated preset dynamic feature library. For example, it uses Euclidean distance or cosine similarity algorithms to identify the physical interaction mode that the current device is most likely to belong to.

[0049] The pre-calibrated preset dynamic feature library is a database established through offline data collection before system deployment. The physical interaction modes include at least automatic motor-driven states and manual interaction states. Among them, the motor-driven state includes standard features such as start / stop and acceleration / deceleration; the manual interaction state includes typical features such as continuous acceleration push / pull and reciprocating trial push / pull.

[0050] The preset dynamic feature library also associates a set of target waypoint arrival probability distributions with each mode. The mode matching unit combines the first and second matching results to extract the corresponding probability distributions and generate a target intent vector containing probability weights. The target intent vector is usually presented in an array structure, such as [target waypoint ID_1, probability_1, target waypoint ID_2, probability_2,...] format, which clarifies the targets and their priorities that should be focused on in subsequent resource scheduling.

[0051] For example, following the example in the timing feature vector generation module, an acceleration signal with an amplitude of 500,000 mm / s³ appears in the timing feature vector. A preset silent noise limit is set to 10,000 mm / s³, and a preset baseline threshold is set to 50,000 mm / s³. The waveform monitoring unit detects that 500,000 mm / s³ far exceeds the preset baseline threshold, thus triggering waveform capture. Assuming a fixed duration of 400 ms, at a sampling rate of 100 Hz, 40 consecutive acceleration data points are captured to form a temporary acceleration waveform segment.

[0052] The wavelet analysis module processes the temporarily stored accelerometer waveform segment. The calculation results show that its energy is highly concentrated in the high-frequency scale, and a single strong peak appears in the first 1 / 4 of the time window. The extracted peak distribution features are encoded into a vector, such as [0.1,0.8,0.1,1,500000,0.1], which represent the proportion of low, medium and high frequency energy, the number of peaks, the maximum peak value and the peak position, respectively.

[0053] The pattern matching unit compares the peak distribution features with a pre-calibrated preset dynamic feature library and finds that its cosine similarity with the "Quick Start Manual Interaction" template in the table is the highest, reaching 0.98. Therefore, the intent recognition module determines that the device is in a manual interaction state. According to the preset dynamic feature library, the probability distribution associated with the "Quick Start Manual Interaction" template is as follows: the next nearest orbital target point, such as TP-05, has an arrival probability of 0.95; the second nearest target point, TP-06, has an arrival probability of 0.05. The target intent vector generated and output by the intent recognition module is [TP-05, 0.95, TP-06, 0.05].

[0054] In one embodiment of the present invention, the dynamic resource flow list generation module is used to perform the following steps: Deconstruct the 3D spatial mesh geometry data contained in the scene asset data and divide it into fine-grained resource data blocks; Obtain the arrival probability of the target intent vector marker, and combine it with the occupied volume of the fine-grained resource data block to calculate the expected usage window duration and initial prediction priority of the fine-grained resource data block; The initial resource loading list is generated by concatenating parameters such as fine-grained resource data blocks, initial prediction priorities, and expected usage window duration.

[0055] The dynamic resource flow list generation module is executed by the central processing unit (CPU). The CPU's built-in resource management module receives the target intent vector generated by the physical interaction mode determination module and performs the following operations to generate an initial resource loading list associated with specific timeframes. The resource management module accesses a pre-built scene asset database, which stores information about all visible 3D spatial mesh geometry data and their associated resources within the target virtual scene in list form.

[0056] The resource management module traverses the scene asset database, parses each resource, and extracts its texture detail (such as the resolution of the texture map) and modeling face count (i.e. the number of polygons contained in the 3D model). Based on these two key indicators, the scene assets are decomposed and classified into fine-grained resource data blocks with different basic priorities.

[0057] Fine-grained resource data blocks refer to the smallest independent loading unit that constitutes a complete rendering scene, such as mesh data, texture maps, shader programs, etc. The basic priority of fine-grained resource data blocks is based on the above indicators. For example, the basic priority of asset blocks with more than 50,000 polygons or textures of 4096×4096 can be set to a higher value close to 1, while those with fewer polygons and solid color fills can be set to a lower value.

[0058] The resource management module extracts the endpoint coordinates of each path from the received target intent vector, i.e., each target waypoint and its corresponding arrival probability. For a set of fine-grained resource data blocks associated with each target waypoint, the resource management module calculates the loading time required to load the fine-grained resource data block into the cache by combining the volume occupied by each fine-grained resource data block itself (i.e., the file size on the storage medium) and the estimated system loading rate. It then combines this with the preset display duration of the fine-grained resource data block in the scene to calculate the expected usage window duration. The logic for setting the expected usage window duration aims to ensure that assets are fully loaded before display begins and remain in the cache during the display period.

[0059] The resource management module multiplies the base priority value of fine-grained resource data blocks by their arrival probability to generate an initial predicted priority. The calculation process of the initial predicted priority and the expected usage window duration can be expressed by the following formula:

[0060] In the formula, Representative and the The first target waypoint associated with the Initial prediction priority for each fine-grained resource data block; Represents the record in the intent vector. The probability of reaching each waypoint; Representative asset block The normalized base priority, ranging from 0 to 1; This represents the expected duration of the window, in seconds. Representative asset block The occupied volume, in MB; This represents the system's estimated average resource loading rate, expressed in MB / s. This represents the preset effective display duration, measured in seconds (s).

[0061] The resource management module concatenates all the fine-grained resource data blocks associated with the target waypoints, their corresponding initial prediction priorities, and the calculated expected usage window durations to form an unordered list containing all resources to be scheduled. Then, the unordered list is sorted in descending order according to the initial prediction priorities and compiled to generate an initial resource loading list in instruction format. The initial resource loading list is represented as an ordered array of objects in data structure, providing a priority-ordered job sequence for the subsequent asynchronous loading engine.

[0062] For example, following the output of the physical interaction mode determination module, the target intent vector received by the resource management module is [TP-05, 0.95, TP-06, 0.05]. Assume that the target waypoint TP-05 corresponds to a complex virtual engine display scenario: Asset Block A {Engine Casing, Basic Priority}. Occupying volume Display duration }. Asset Block B {Internal Piston, Basic Priority} Occupying volume Display duration }. The target waypoint TP-06 corresponds to a simple text sign, including asset block C {sign panel, basic priority}. Occupying volume Display duration }

[0063] System estimated loading rate The speed is 20MB / s. The resource management module first calculates the initial prediction priority for each asset block: asset block A's... Asset Block B Asset block C .

[0064] Calculate the expected usage window duration for each: Asset Block A Asset Block B Asset block C Then, the resource management module concatenates the calculation results into a list and sorts them in descending order according to the initial prediction priority.

[0065] The generated initial resource loading list is logically represented as an ordered sequence. The first item is {Asset Block A identifier, priority 0.95, window duration 15.5s}, the second item is {Asset Block B identifier, priority 0.76, window duration 15.2s}, and the third item is {Asset Block C identifier, priority 0.025, window duration 8.025s}.

[0066] In one embodiment of the present invention, the loading pressure vector construction module is used to perform the following steps: Send a polling command to read the actual read rate of the storage medium at the communication transmission port; Record the physical processing latency of the graphics rendering bus and calculate the latency deviation ratio according to the expected baseline latency deviation. A weighted fusion operation is performed on the ratio of the actual read rate of the storage medium to the latency deviation to generate a real-time load pressure vector.

[0067] The load pressure vector construction module is executed by the Central Processing Unit (CPU). The CPU's built-in system monitoring module runs continuously in parallel with other modules to construct a load pressure vector characterizing the current system load state. Through interaction with the underlying operating system drivers, the system monitoring module balances the need for real-time monitoring data with the need to avoid excessive CPU resource consumption at fixed intervals. It sends a polling instruction to the data exchange microkernel controlling the hardware storage medium—the software module in the operating system kernel responsible for managing storage input / output. The fixed interval can be set to every 100ms, which satisfies the physiological threshold for human visual perception of screen stuttering, typically between 100ms and 200ms, while avoiding unnecessary overhead on the underlying CPU from higher-frequency polling.

[0068] Polling commands are used to query and read the transient data throughput of the communication transmission port over the past cycle, thereby obtaining the actual read rate of the storage medium. Simultaneously, the system monitoring module continuously records the physical processing latency of the graphics rendering bus from the moment it receives an externally received resource loading request stream until the data block pointed to by the resource loading request is successfully loaded into the cache. It then compares the physical processing latency with a preset expected baseline latency to calculate the dimensionless latency deviation ratio.

[0069] In this system, when the actual processing delay equals the baseline delay, a delay deviation ratio of 0 indicates no congestion; the longer the delay and the larger the positive value of the delay deviation ratio, the more severe the congestion. The fusion calculation unit within the system monitoring module calculates the periodically acquired media data read rate and the delay deviation ratio recorded at the corresponding time using a preset weighted fusion function. The purpose of the weighted fusion function is to combine two physical quantities with different dimensions; the calculation formula is expressed as:

[0070] And the time delay deviation ratio:

[0071] In the formula, Representative at The scalar value of the real-time loading pressure vector calculated at each moment; The minimum value function is used to normalize the pressure value to the range of 0 to 1; the larger the value, the higher the pressure. Representative at Real-time read speed, in MB / s; Represents the hardware's nominal peak read rate; Representative at The physical processing delay recorded at any given time, in milliseconds (ms). This represents the expected baseline processing latency obtained from tests conducted in the idle state. The weighting coefficient is set between 0 and 1. For example, setting it to 0.6 is intended to assign the reading rate of the underlying physical medium to represent the proportion of a slightly higher hardware bottleneck, while assigning the remaining weight of 0.4 to the task processing latency to represent the scheduling status of the upper software layer, so as to truly reflect the dominant role of data path throughput on the overall system pressure. The upper limit for preventing overflow of the set delay deviation ratio is, for example, set to 2.0. This upper limit means that when the actual processing delay reaches 3 times the baseline normal delay, that is, when the deviation rate reaches 200%, the system is already at the physical congestion tolerance limit, which can effectively prevent occasional extreme outliers from causing the pressure calculation model to diverge.

[0072] By using a weighted fusion function, the combined effects of data path bandwidth and task queue congestion are reflected, thereby generating a real-time loading pressure vector that represents the overall system load congestion. This real-time loading pressure vector is continuously updated and provided to other modules as a basis for decision-making.

[0073] For example, suppose that at a certain moment the system monitoring module starts monitoring, the nominal peak read rate of the hardware solid-state drive... The system's preset baseline processing latency is 500 MB / s. The weighting factor is 5ms. Set to 0.6, maximum. Set to 2.0.

[0074] exist At any given time, the system monitoring module obtains the actual read rate of the storage medium through polling. The speed is 450MB / s, and the physical processing latency of the current loading task is also recorded. The processing time is 6ms. (This refers to the computational speed of the fusion processing unit.) Then calculate the delay deviation ratio. After weighted fusion, the loading pressure value at time step [time] is calculated. A low load pressure value indicates that the system is currently under very light load.

[0075] Another 500ms passed, at At that moment, due to the influx of numerous resource loading requests, the system monitoring module detected... Reduced to 150MB / s while handling latency It increased dramatically to 25ms. At this point, the rate component was... Delay deviation ratio Calculate the loading pressure value. Reaching the upper limit of 1.0 indicates that the system is currently under full load or overload load pressure, and the real-time load pressure vector is updated to 1.0.

[0076] In one embodiment of the present invention, the resource flow list correction module is configured to perform the following steps: Analyze the system load ratio represented by the real-time load pressure vector and compare the system load ratio with the safe load warning threshold. When the system load ratio exceeds the safe load warning threshold, the extraction priority of fine-grained resource data blocks that meet the conditions identified in the initial resource loading list is reduced. Perform validation and pruning on the initial resource loading list after rearrangement, remove asset reading tasks that exceed the expected usage window duration, and generate a corrected resource loading list.

[0077] The resource flow list correction module is executed by the central processing unit (CPU). The CPU's built-in scheduling strategy adjustment module receives real-time load pressure vectors continuously reported by the step-based load pressure vector construction module and dynamically corrects the initial resource load list generated by the step-based dynamic resource flow list generation module based on these real-time load pressure vectors. The pressure parsing unit within the scheduling strategy adjustment module analyzes the system load ratio represented by the real-time load pressure vector in real time. The value is then compared with a safety load warning threshold embedded in the control chip or software configuration.

[0078] The safety load warning threshold is usually set as an empirical threshold, such as 0.7, meaning that intervention begins when the system load reaches 70%, balancing the need to avoid lag with maintaining visual quality. When the system load ratio remains below the safety load warning threshold, the scheduling strategy adjustment module considers the load normal and directly issues the initial resource loading list. However, once the system load ratio is detected to exceed the limit, the scheduling strategy adjustment module immediately adjusts the underlying scheduling strategy.

[0079] Specifically, the priority adjustment unit traverses the current initial resource loading list and identifies fine-grained resource data blocks that simultaneously meet two conditions: First, the fine-grained resource data block itself has a large size, for example, the file size exceeds a certain size threshold, such as 5MB. This size threshold is based on asset size distribution statistics and is used to distinguish high-precision models or textures. Second, the type of fine-grained resource data block belongs to complex particle and high-definition texture data that enhances the realism of the scene rather than the core structural composition, such as high-definition texture maps larger than 2048×2048 or post-processing scripts with high computational costs.

[0080] For the identified fine-grained resource data blocks, regardless of how high the initial prediction priority of the fine-grained resource data block is, the priority adjustment unit will forcibly lower its extraction priority. Specifically, the initial prediction priority is multiplied by a penalty coefficient, such as 0.1, so that its order in the loading queue is significantly delayed.

[0081] The control logic of the priority adjustment unit can be abstractly expressed as: if For each asset block in the list If the conditions are met and {High-resolution textures, complex particles}, then the new priority ;otherwise .in, Represents the safety load warning threshold; Representative asset block The volume occupied; This represents a preset volume threshold. Representative asset block Type identifier; This represents the adjusted revised forecast priority; Represents the initial prediction priority; This is the penalty coefficient.

[0082] After rearranging according to the new priority, the verification and pruning unit performs final verification. This unit checks the expected usage window duration of each fine-grained resource data block in the list and removes asset tasks that might time out due to lowering their priority, as well as all read tasks of predicted waypoints that have expired due to changes in the actual movement trajectory of the device. After this pruning and optimization, a revised resource loading list with a consistent structure but reordered according to the actual load is generated to guide subsequent loading execution.

[0083] For example, following the conclusion of the step of loading the pressure vector construction module, in At any given time, the real-time load pressure vector value is 1.0. This is the set safe load warning threshold. The value is 0.7. The scheduling strategy adjustment module detected this. This triggers an adjustment strategy. At this point, the initial resource loading list to be processed is generated by the dynamic resource flow list generation module.

[0084] The first item is {Asset Block A, priority 0.95}, the second item is {Asset Block B, priority 0.76}, and the third item is {Asset Block C, priority 0.025}. Assume that Asset Block A (engine casing) is tagged as high-resolution texture data, and its size of 10MB exceeds the set size threshold of 5MB. Asset Block B (internal piston) is tagged as core geometry. Asset Block C (signboard panel) is too small to meet the requirements.

[0085] Therefore, the priority adjustment unit only adjusts the priority of asset block A, and its new priority... The priorities of asset blocks B and C remain unchanged, i.e. The scheduling strategy adjustment module adjusts the scheduling strategy according to the new priority. The list is reordered, resulting in the following new order: Asset Block B (Priority 0.76), Asset Block A (Priority 0.095), Asset Block C (Priority 0.025).

[0086] The validation and trimming unit checks this list. Assuming that the expected usage window duration for all assets remains valid and has not expired after this adjustment, the loading order of the final revised resource loading list is changed: the piston (asset block B), which is the core geometry, is loaded first, followed by the downgraded high-definition shell (asset block A), and finally the low-priority signboard (asset block C). The revised resource loading list ensures that even under high system load, the core content of the scene is prioritized for rendering.

[0087] See appendix Figure 3 Based on the data from the above embodiments, it can be seen that the initial prediction priority (0.76) and the corrected prediction priority (0.76) of asset block B, which is a core component, remain unchanged; asset block A, which is a high-definition texture, has its initial prediction priority significantly reduced from 0.95 to 0.095 due to the system overload triggering a degradation strategy; and the initial prediction priority of lightweight asset block C remains unchanged at 0.025. The orange broken lines in the figure reflect the expected usage window duration of each asset at the target waypoint, which are 15.2s, 15.5s, and 8.025s, respectively. This figure illustrates the dynamic scheduling strategy of prioritizing the loading of the core basic model under high load pressure.

[0088] In one embodiment of the present invention, the asynchronous resource injection module is configured to perform the following steps: Parse the scheduling instruction vector inside the corrected resource loading list to trigger the block-based fetch instruction; Allocate high-speed video memory buffer blocks to the high-speed video memory system, and write the retrieved fine-grained resource data blocks into the corresponding video memory physical address space. Based on the actual motion displacement of the device and the target intent vector, the failure prediction waypoint is detected, and the video memory buffer occupied by the resources associated with the failure prediction waypoint is released.

[0089] The execution body of the asynchronous resource injection module is the central processing unit. The memory management module of the central processing unit receives the corrected resource loading list generated by the resource flow list correction module and drives the underlying memory pool control engine, i.e. the set of video memory scheduling APIs provided by the operating system or graphics driver, to realize the orderly and asynchronous preloading of resource assets to the high-speed video memory system. Here, the high-speed video memory system refers to the high-bandwidth synchronous dynamic random access memory on the GPU board.

[0090] The inventory parser within the memory management module parses the scheduling instruction vectors contained within the corrected resource loading list line by line, processing them one by one in priority order. For the first asset block with the highest priority, the inventory parser issues a block-by-block fetch instruction to the controller of the low-speed storage space. The block-by-block fetch instruction contains the physical / logical storage address and length information of the asset block, asynchronously initiating the data reading process.

[0091] Simultaneously, the resource receiving and control unit calls the video memory management interface in the target receiving end's video memory to allocate a high-speed video memory buffer block whose size matches the volume of the first asset block. Once the block-by-block fetching instruction is completed and the data arrives in system memory, the data transmission module immediately writes the retrieved fine-grained resource data blocks step by step and sequentially into the previously allocated video memory physical address space—that is, the area directly addressable by the GPU for high-performance rendering. After processing the current item, the list parser immediately processes the next item in the list, repeating the parallel process of fetching, allocating, and writing.

[0092] Meanwhile, the cache reclamation unit compares the actual displacement of the device with the positions of each predicted waypoint in the target intent vector, detecting failed predicted waypoints that have already been traversed by the device or whose arrival probability has dropped to zero according to the latest intent prediction. Once a failed predicted waypoint is detected, the cache reclamation unit triggers a reclamation mechanism, querying and releasing the video memory buffers occupied by all resources associated with the failed predicted waypoint. By following a mechanism of asynchronous loading and parallel reclamation of video memory according to priority, video memory no longer statically stores the entire scene, but acts as a dynamic transfer station that is constantly dynamically overwritten and released, achieving cyclical streaming use and supporting the smooth operation of large-scale virtual scenes.

[0093] For example, following the output of the resource stream list correction module, the memory management module receives the corrected resource loading list, in the following order: 1. Asset Block B (internal piston), 2. Asset Block A (engine casing), 3. Asset Block C (signboard panel). The list parser processes the first item, issuing an instruction to the solid-state drive to read asset block B, and simultaneously allocating a 4MB buffer block in video memory. When the data stream of asset block B arrives, it is written to the video memory buffer block.

[0094] The manifest parser processes the second item, triggering an asynchronous read of asset block A and requesting a 10MB buffer block in video memory. While asset block A is being transferred, the manifest parser may have already begun processing the third item, triggering a read of asset block C. This process is parallel.

[0095] Meanwhile, assuming the device's actual position on the track is 2500mm, and the target intent vector generated by the physical interaction mode determination module also includes a failure prediction waypoint TP-04 located at 1200mm on the track, the resource asset block D associated with this waypoint has already been loaded into video memory. The cache reclamation unit detects that the current position of 2500mm is far beyond the 1200mm of waypoint TP-04, and the probability of TP-04 in the latest target intent vector is zero, therefore determining TP-04 as a failure prediction waypoint. It then issues a command to the video memory manager to reclaim and release the video memory buffer occupied by asset block D. This frees up video memory space for the upcoming asset block A or subsequent higher-priority resources, achieving efficient video memory recycling.

[0096] In one embodiment of the present invention, the deferred rendering output module is configured to perform the following steps: Compare the absolute encoded coordinates of the track equipment sensors corresponding to the actual physical location with the preset points of the target scene trigger site, and extract the position matching trigger signal; Verify the readiness status of fine-grained resource data blocks corresponding to the scene in the physical address space of video memory; The ready fine-grained resource data blocks are fed into the hardware rendering pipeline to perform shading and layout, and output a three-dimensional virtual image.

[0097] The deferred rendering output module is executed by the central processing unit (CPU). The CPU's built-in rendering scheduling module reads displacement information extracted from the basic motion dataset generated by the temporal feature vector generation module in real time and converts it into absolute coded coordinates of the track equipment sensors. This means that the incremental displacement pulses are associated with the physical starting point of the track and accumulated to obtain the corresponding absolute spatial values. At the same time, it performs periodic real-time comparison calculations with the target scene trigger preset points extracted from the target intent vector of the physical interaction mode determination module, which are the calibration point data established along the track.

[0098] When the single-dimensional spatial distance between the absolute encoded coordinates and a preset point is less than a preset spatial tolerance threshold Upon arrival, the rendering scheduling module immediately determines that the physical arrival has occurred. The calculation process for triggering the physical arrival determination can be expressed by the formula:

[0099] when hour, Otherwise, it is 0, where, Represents spatial distance error, in mm; Represents the current absolute encoded coordinates; Represents the preset point value; The spatial tolerance threshold is set in conjunction with the upper limit of the screen refresh rate of the device speed calibration. Its range is usually set to 5mm to 20mm to ensure that the screen does not produce a visual disconnect when the device is forced to wake up. The formula uses absolute value distance calculation, which is naturally compatible with the bidirectional low latency trigger judgment requirements of the track display device when sliding forward and backward. It is a Boolean indicator.

[0100] When the determination is reached, that is When the value is 1, the rendering scheduling module captures instantaneous data to generate a position matching trigger signal. This position matching trigger signal is identified as a high-priority hardware interrupt triggering scene construction, and the rendering scheduling module checks the current readiness status of each asset block in the corresponding scene in the video memory. The readiness status verification process executes the basic model readiness standard.

[0101] The basic model readiness standard serves as a data integrity threshold to prevent visual damage caused by fragmented rendering blocks. It requires the rendering scheduling module to confirm that the skeleton structure that serves as the basic geometric model in the target scene and the core asset blocks that provide basic color values ​​have been fully filled into the high-speed buffer block. Even if the surface high-definition texture is lacking, the scene rationality can be maintained by using the shading algorithm, thereby avoiding the loss of model topology.

[0102] After the verification is passed, the rendering scheduling module sends all the fine-grained resource data blocks that are ready and resident in the buffer block into the hardware rendering pipeline according to the graphics instruction sequence. That is, the dedicated computing circuit in the GPU is responsible for coordinate transformation, depth testing, rasterization and image generation. After vertex shading calculation, polygon rasterization and pixel arrangement operations, the system instantly outputs a complete 3D virtual image and drives the track display device to display it with low latency.

[0103] For example, following the scheduling state of the asynchronous resource injection module, the memory management module has previously prioritized scheduling the core asset category block corresponding to target point TP-05, namely the internal piston model, based on the revised resource loading list. Furthermore, asset block B is fully ready to be reserved in the physical address space of video memory. Meanwhile, asset block A, which belongs to high-definition texture data, namely the high-definition texture part of the engine shell, is still in a continuous transmission and loading state due to bandwidth constraints.

[0104] Assuming the user pushes the track device into the target display area with a constant force, the system's set spatial tolerance threshold... The target scenario trigger point TP-05's track coordinate record is 3000mm. When the system performs a high-frequency comparison task and finds that the absolute encoded coordinates of the current track equipment sensor reach 2992mm, the calculated distance difference is 8mm, which is less than the set tolerance threshold of 10mm. The relative distance condition activates the physical detection and judgment logic, and the system instantly generates a position matching trigger signal with a control value of 1.

[0105] Upon receiving the position matching trigger signal, the rendering scheduling module quickly checked the data status of the high-speed video memory buffer block. The verification module found that the data volume of the piston component of asset block B, which serves as the basic geometric model, had been fully allocated and filled, and its basic three-dimensional skeleton structure and metallic color value parameter information were complete, thus meeting the basic model readiness standard set by the program in advance.

[0106] Since the high-definition large texture of asset block A, which is being asynchronously transmitted, has not yet been declared ready, the system temporarily suspends the incomplete shell rendering request based on a degradation rendering strategy. After stripping the failed instructions, it directly pulls the prepared asset block B and sends it into the hardware rendering pipeline to start rasterization shading arrangement calculations. Finally, upon completion of the calculations, the track display device continuously outputs a complete 3D virtual image of the internal piston mechanical structure of the underlying engine with low latency at the instant it slides into the actual scene point.

[0107] See appendix Figure 4 In the diagram, the solid orange line represents the V-shaped trend of the absolute spatial distance error as the equipment coordinates change, with the center zero point being the preset target point of 3000mm; the dashed gray line represents the set spatial tolerance threshold. The value is 10mm. When the device slides and the absolute spatial distance error drops below this tolerance threshold, i.e., when the device's current absolute encoding coordinates fall within the range of 2990mm to 3010mm, the condition is met. Under the condition of ≤10mm, the position matching trigger signal status represented by the green solid line instantly jumps from 0 to 1. This figure illustrates the physical triggering mechanism of this system that utilizes spatial tolerance to achieve early wake-up and reduce rendering latency.

[0108] Each of the modules can be implemented in whole or in part through software, hardware, or a combination thereof. It supports hardware embedded in or independent of the processor in the computer device, and also supports software stored in the memory of the computer device, so that the processor can call and execute the operations corresponding to each of the above modules.

[0109] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A low-latency triggering system for AR scenes based on orbital displacement sensing, characterized in that, include: The time-series feature vector generation module acquires the original position signal of the track display device, performs differential conversion on the original position signal, and obtains the time-series feature vector. The physical interaction mode determination module performs waveform feature extraction and pattern matching on the temporal feature vector to obtain the target intent vector. The dynamic resource flow list generation module performs asset decomposition and priority calculation on scene asset data based on the target intent vector to obtain an initial resource loading list; Load the pressure vector construction module, obtain the underlying system operation status indicators, perform weighted fusion calculation on the operation status indicators, and obtain the real-time loading pressure vector; The resource loading list correction module performs scheduling strategy adjustment, verification and trimming on the initial resource loading list based on the real-time loading pressure vector to obtain the corrected resource loading list. The asynchronous resource injection module performs asynchronous preloading and cache reclamation based on the revised resource loading list, and writes the corresponding resource assets into the high-speed video memory system; The deferred rendering output module compares the actual physical location with the target scene trigger preset points on-site based on the comparison results, calls the resource assets in the high-speed video memory system to perform rendering operations, and obtains a three-dimensional virtual image.

2. The low-latency triggering system for AR scenes based on orbital displacement sensing according to claim 1, characterized in that, The process involves acquiring the raw position signal from the track display device, performing differential conversion on the raw position signal to obtain a time-series feature vector, and including the following steps: Acquire real-time pulse-coded data and angle and attitude change parameters sent by the track display device to generate a basic motion dataset; Filtering and noise reduction processing and continuous derivative calculation are performed on the basic motion dataset to obtain velocity data and first-order acceleration data; Perform time difference conversion on the first-order acceleration data to extract the jerk signal, and encapsulate the jerk signal into a time-series feature vector.

3. The low-latency triggering system for AR scenes based on orbital displacement sensing according to claim 1, characterized in that, Perform waveform feature extraction and pattern matching on the time-series feature vector to obtain the target intent vector, including the following steps: Monitor the jerk values ​​contained in the timing feature vector, compare the jerk values ​​with the preset silent noise limit, and confirm the starting point of the instantaneous thrust. When the jerk value exceeds the preset baseline threshold, a single set of changing bands is extracted, and the spectral energy and peak distribution characteristics of the single set of changing bands are extracted. By comparing the peak distribution characteristics with a pre-calibrated preset dynamic feature library, the physical interaction mode is identified, and the target intent vector is generated based on the identification results.

4. The AR scene low-latency triggering system based on track displacement sensing according to claim 1, characterized in that, Based on the target intent vector, asset decomposition and priority calculation are performed on the scene asset data to obtain an initial resource loading list, including the following steps: Deconstruct the 3D spatial mesh geometry data contained in the scene asset data and divide it into fine-grained resource data blocks; Obtain the arrival probability of the target intent vector marker, and combine it with the occupied volume of the fine-grained resource data block to calculate the expected usage window duration and initial prediction priority of the fine-grained resource data block; The initial resource loading list is generated by concatenating parameters such as fine-grained resource data blocks, initial prediction priorities, and expected usage window duration.

5. The low-latency triggering system for AR scenes based on orbital displacement sensing according to claim 4, characterized in that, The arrival probability of the target intent vector is obtained, and combined with the occupied volume of the fine-grained resource data block, the expected usage window duration and initial prediction priority of the fine-grained resource data block are calculated, including the following steps: Obtain the basic priority and preset display duration of fine-grained resource data blocks, as well as the estimated average resource loading rate; The arrival probability is multiplied by the base priority to generate the initial prediction priority; The loading time is obtained by dividing the occupied volume by the average resource loading rate. The loading time is then added to the preset display duration to generate the expected usage window duration.

6. The AR scene low-latency triggering system based on track displacement sensing according to claim 1, characterized in that, Obtain the underlying system operation status indicators, perform weighted fusion calculations on the operation status indicators to obtain the real-time loading pressure vector, including the following steps: Send a polling command to read the actual read rate of the storage medium at the communication transmission port; Record the physical processing latency of the graphics rendering bus and calculate the latency deviation ratio according to the expected baseline latency deviation. A weighted fusion operation is performed on the ratio of the actual read rate of the storage medium to the latency deviation to generate a real-time load pressure vector.

7. A low-latency triggering system for AR scenes based on orbital displacement sensing according to claim 6, characterized in that, The real-time load pressure vector is generated by performing a weighted fusion calculation on the ratio of the actual read rate of the storage medium to the latency deviation, including the following steps: Obtain the preset nominal peak read rate and weighting coefficient; The ratio of the actual read rate to the nominal peak read rate of the storage medium is calculated, and the rate component is calculated based on the ratio. The rate component and the latency deviation ratio are weighted and fused using a preset weighted fusion function to generate a real-time load pressure vector.

8. The low-latency triggering system for AR scenes based on orbital displacement sensing according to claim 1, characterized in that, Based on the real-time loading pressure vector, the initial resource loading list is adjusted and verified by performing scheduling strategy adjustments and pruning to obtain a revised resource loading list, including the following steps: Analyze the system load ratio represented by the real-time load pressure vector and compare the system load ratio with the safe load warning threshold. When the system load ratio exceeds the safe load warning threshold, the extraction priority of fine-grained resource data blocks that meet the conditions identified in the initial resource loading list is reduced. Perform validation and pruning on the initial resource loading list after rearrangement, remove asset reading tasks that exceed the expected usage window duration, and generate a corrected resource loading list.

9. The low-latency triggering system for AR scenes based on track displacement sensing according to claim 1, characterized in that, Based on the revised resource loading list, asynchronous preloading and cache reclamation are performed, and the corresponding resource assets are written to the high-speed video memory system, including the following steps: Parse the scheduling instruction vector inside the corrected resource loading list to trigger the block-based fetch instruction; Allocate high-speed video memory buffer blocks to the high-speed video memory system, and write the retrieved fine-grained resource data blocks into the corresponding video memory physical address space. Based on the actual motion displacement of the device and the target intent vector, the failure prediction waypoint is detected, and the video memory buffer occupied by the resources associated with the failure prediction waypoint is released.

10. A low-latency triggering system for AR scenes based on orbital displacement sensing according to claim 1, characterized in that, Based on the comparison results of the target scene trigger preset points on-site according to the association between the actual physical location and the target intent vector, the resource assets in the high-speed video memory system are called to perform rendering operations to obtain a three-dimensional virtual image, including the following steps: Compare the absolute encoded coordinates of the track equipment sensors corresponding to the actual physical location with the preset points of the target scene trigger site, and extract the position matching trigger signal; Verify the readiness status of fine-grained resource data blocks corresponding to the scene in the physical address space of video memory; The ready fine-grained resource data blocks are fed into the hardware rendering pipeline to perform shading and layout, and output a three-dimensional virtual image.

Images