Customs declaration process teaching method and system based on virtual reality

Abstract

The invention provides a customs declaration process teaching method and system based on virtual reality, and belongs to the technical field of teaching methods.The customs declaration process teaching method and system based on virtual reality comprises a virtual reality scene generation module used for constructing and outputting three-dimensional virtual scene data; the role distribution module is in communication connection with the virtual reality scene generation module; the document simulation module is used for storing and outputting the electronic form data to the student terminal; the process decision module is in communication connection with the document simulation module; the behavior acquisition module is used for acquiring operation track data generated by the student terminal in the virtual scene in real time; and the teaching management background is in communication connection with the behavior acquisition module, and solves the source denoising problems of high-frequency distortion of an operation track and inaccurate evaluation caused by vibration, rotation and misoperation in virtual reality-based customs declaration teaching.

Description

Technical Field

[0001] This invention belongs to the field of teaching methods technology, specifically, it relates to a teaching method and system for customs declaration process based on virtual reality. Background Technology

[0002] In the field of port logistics and customs talent training, the teaching of customs declaration procedures has long relied on physical documents, paper cases, and scattered on-site demonstrations. With the maturity of virtual reality technology, digitizing real-world scenarios such as ports, regulatory zones, and customs halls and integrating them into VR headsets has become a direction that various universities are vying to explore. Compared to traditional classrooms, VR solutions can present dangerous or high-cost scenarios within the classroom and can also achieve repeated training through playback functions, thus being regarded as an important breakthrough in customs practice teaching.

[0003] With the integration of virtual reality into customs declaration training, teachers can bring port scenarios into the classroom, allowing students to complete electronic declarations and cooperate with inspections simply by wearing headsets. However, existing systems generally treat "scene recreation" as the endpoint, neglecting the reliability of the data link: the inertial measurement units (IMU) of the controllers and headset, and the eye-tracking sensors output high-frequency glitches during rapid rotations, wrist vibrations, or accidental shaking; if simple low-pass filtering is used at the software layer, phase lag is significant, and trajectory playback still jumps, making it difficult for teachers to determine whether students are truly focusing on key areas; the competency profile is thus distorted, reducing the credibility of teaching. More importantly, after vibration noise is mixed in, the trajectory stored in the background deviates from the actual movement, and it cannot be restored afterward regardless of the algorithm used, creating a dilemma of "inaccuracy at the source, failure throughout the process." The industry urgently needs a hardwired-level solution that can suppress noise in real time at the data acquisition end without conflicting with real-time requirements; otherwise, VR customs declaration training will remain in a rudimentary stage that is visible but unreliable for a long time. Summary of the Invention

[0004] In view of this, the present invention provides a teaching method and system for customs declaration process based on virtual reality, which solves the problem of noise reduction at the source in customs declaration teaching based on virtual reality, which causes high-frequency distortion of operation trajectory and inaccurate evaluation due to vibration, rotation and misoperation.

[0005] This invention is implemented as follows:

[0006] This invention provides a virtual reality-based customs declaration process teaching system, comprising:

[0007] Virtual reality scene generation module, used to construct and output three-dimensional virtual scene data;

[0008] The role assignment module is communicatively connected to the virtual reality scene generation module and is used to send role identification data to the student terminals that enter the virtual scene.

[0009] The document simulation module is used to store and output electronic form data to the student's terminal;

[0010] The process decision module is communicatively connected to the document simulation module and is used to output deployment instruction data based on the electronic form data.

[0011] The behavior acquisition module is used to collect real-time operation trajectory data generated by the student's terminal within the virtual scene; and

[0012] The teaching management backend is connected to the behavior acquisition module and is used to receive and store the operation trajectory data.

[0013] The technical effects of the virtual reality-based customs declaration process teaching system provided by this invention are as follows: Through the closed-loop connection of six hardware modules, namely "scene-role-document-deployment-behavior-backend", the teaching elements that were originally scattered in the classroom, customs hall and port are compressed into the same VR terminal, which significantly reduces the cost of site occupation and equipment transportation. At the same time, by utilizing the plug-and-play feature of the standard interface between modules, schools can complete system expansion or upgrade without changing the existing local area network.

[0014] Based on the above technical solution, the virtual reality-based customs declaration process teaching system of the present invention can be further improved as follows:

[0015] The virtual reality scene generation module includes:

[0016] Point cloud data interface, used to receive external laser point cloud data;

[0017] A 3D reconstruction unit, connected to the point cloud data interface, is used to convert the laser point cloud data into triangular mesh data; and

[0018] The rendering unit, connected to the 3D reconstruction unit, is used to render the triangular mesh data into a frame sequence in real time and output it to the student terminal.

[0019] The beneficial effects of adopting the above-mentioned improvement scheme are as follows: the virtual reality scene generation module replaces the traditional manual modeling with a hard pipeline of "point cloud interface-reconstruction unit-rendering unit", shortening the cycle from real port laser scanning to roamable VR from 7 days to 4 hours. Moreover, the triangular mesh data is generated once and reused on multiple devices, ensuring that the size error of the inspection area seen by all trainees is less than 2cm, avoiding the inconsistency of assessment standards caused by "different scenes in the same class".

[0020] Furthermore, the document simulation module includes:

[0021] A form database for storing at least one electronic form template;

[0022] The input interface is used to receive field data entered by students' terminals; and

[0023] The verification unit, connected to the input interface, is used to compare the field data with preset rule data and output verification result data.

[0024] The beneficial effects of adopting the above-mentioned improvement scheme are as follows: The document simulation module uses an embedded architecture of "form database - input interface - verification unit" to solidify the 337 cross-reference relationships of 47 columns of customs documents in the hardware regular expression engine, so that the trainee can get feedback that is completely consistent with the real document review system within 200ms after entering a column, which solves the lag problem of "filling out and then judging, and difficulty in tracking errors" in traditional paper simulation.

[0025] Furthermore, the behavior acquisition module includes:

[0026] Handle gesture sensor, used to output the spatial coordinate data of the handle;

[0027] Eye-tracking sensors are used to output data on the direction of eye movement; and

[0028] The timing integration unit, connected to the handle posture sensor and the eye-tracking sensor, is used to align the spatial coordinate data and the gaze direction data according to the timestamp and generate the operation trajectory data.

[0029] The beneficial effects of adopting the above-mentioned improvement scheme are as follows: the behavior acquisition module uses a hard clock alignment scheme of "handle posture sensor + eye movement sensor + timing integration unit" to compress the error between spatial coordinates and line of sight to within 12ms. When the teacher plays back the video, he can reproduce the detailed actions of the student "first looking down at the document and then looking up to find the storage location" without missing a single frame, providing a quantitative basis for accurately assessing operating habits.

[0030] This invention provides a virtual reality-based teaching method for customs declaration procedures, including the aforementioned virtual reality-based customs declaration procedure teaching system, wherein the method includes the following steps:

[0031] S10: Construct a 3D virtual scene and distribute it to student terminals;

[0032] S20: Send role identification data to the student's terminal;

[0033] S30: Output electronic form data to the student's terminal;

[0034] S40: Receives field data returned by the student terminal and generates deployment instruction data according to preset rule data;

[0035] S50: Collect the operation trajectory data generated by the student terminal in the virtual scene; and store the operation trajectory data in the teaching management backend.

[0036] The steps for obtaining each data item in the method are as follows:

[0037] Role identification data: When the teacher clicks "Start Course" on the terminal, an HTTP POST packet arrives at the teaching management backend; the backend RPC calls the role allocation module; the allocation strategy unit reads the UID table in the EEPROM, selects a UID by modulo algorithm, encapsulates it into 8B data, and unicasts it via Wi-Fi to the student's head-mounted display with the corresponding MAC address.

[0038] Electronic form data: After the document simulation module powers on and performs a self-test, it loads the xsd template into the DDR in one go; the student terminal sends a "request form" UDP packet; the module fills in the template XML with default values, length L≤1400B, and sends it back as "electronic form data".

[0039] Field data: Students input fields on the VR keyboard; the headset locally packages the data into JSON, 802.11n unicast, and document simulation modules; after module verification, the original JSON is saved as "field data".

[0040] Control command data: The process decision module receives the field data, the decision CPU runs the scoring function, outputs 8B instructions, and copies them to the RS-485 FIFO for transmission; the RS-485 receive interrupt of the student terminal stores the 8B into RAM, which is the "control command data".

[0041] Behavior Acquisition Module: Handheld Attitude Sensor: Built-in Bosch BMI160 6-axis IMU, outputting quaternions via 1kHz I²C; Eye Tracking Sensor: Tobii EyeChip, outputting (x,y) normalized coordinates via a 90 Hz USB 2.0 isochronous pipeline; Timing Integration Unit: Lattice LFE5U-25 FPGA, running a 125 MHz metronome, adding 64-bit timestamps (48-bit ns + 16-bit sequence number) to both types of data, merging them into 32 bytes / sample of operation trajectory data.

[0042] Based on the above technical solution, the virtual reality-based customs declaration process teaching method of the present invention can be further improved as follows:

[0043] Step S50 further includes:

[0044] S51: To address the erroneous operation, vibration, or rotation noise contained in the operation trajectory data generated by the student terminal in the virtual scene, a constant velocity state model of the handle spatial coordinates and eye line direction is established. Using the theoretical prediction variance and sensor measurement noise variance as inputs, the Kalman gain is solved iteratively through the Riccati equation, so that the theoretical prediction value and the measured value are dynamically weighted and averaged under the meaning of minimum mean square error to obtain the filtered trajectory data.

[0045] S52: The filtered trajectory data is used as the operation trajectory data for subsequent storage and capability profile generation.

[0046] Furthermore, the steps for constructing the three-dimensional virtual scene include:

[0047] Receive external laser point cloud data;

[0048] Convert the laser point cloud data into triangular mesh data; and

[0049] The triangular mesh data is rendered in real time to generate a frame sequence and distributed to the student terminals.

[0050] The beneficial effects of adopting the above-mentioned improved solution are as follows: the three-step hard process of "laser point cloud - triangular mesh - real-time rendering" transforms the port replication work, which originally required professional modelers, into a one-click operation of "scanning - inserting - automatic running". This allows schools to continuously update teaching scenarios as long as they have publicly available port point clouds, ensuring that course content is synchronized with the latest regulatory environment.

[0051] Furthermore, the step of generating deployment command data includes:

[0052] The field data is compared with the rule database;

[0053] If the comparison result meets the first condition, the first deployment command data is output; and if the comparison result meets the second condition, the second deployment command data is output.

[0054] The beneficial effects of adopting the above-mentioned improvement scheme are as follows: the "field-rule-dual-condition output" hard decision mechanism uses 8B instruction codes to replace manual judgment, which can complete risk control within 10ms. This allows trainees to experience the same "second-level response" as the real customs system, while avoiding the differences caused by subjective scoring by teachers.

[0055] Furthermore, the step of collecting operation trajectory data includes:

[0056] Collect the spatial coordinate data of the handle at the first time interval;

[0057] Data on eye gaze direction was collected at the second time interval; and

[0058] Align the spatial coordinate data with the line-of-sight data according to the timestamp to generate operation trajectory data.

[0059] Settings for the first and second time intervals:

[0060] First time interval (handle): Determined by the BMI160's built-in 1kHz sampling rate, fixed and not adjustable;

[0061] Second time interval (eye movement): Tobii EyeChip defaults to 90Hz, which can be changed to 60Hz or 120Hz via the bInterval field during USB enumeration;

[0062] Alignment strategy: FPGA metronome 125MHz, every The period (=64ms) forces the two latest samples to be packaged together to ensure that the maximum synchronization error is <12ms.

[0063] Furthermore, the method further includes:

[0064] After storing the operation trajectory data, capability profile data is generated based on the operation trajectory data; and the capability profile data is output to the teacher terminal.

[0065] The beneficial effects of adopting the above-mentioned improvement scheme are as follows: by adding a "capability profile" output step in the background, the original trajectory data is automatically mapped into a four-dimensional radar chart with compliance, timeliness, etc. Teachers can locate the common shortcomings of the class within 5 seconds without additional statistics, thus achieving precise teaching.

[0066] Furthermore, the method further includes:

[0067] The operation trajectory data is packaged into data blocks; and the hash value of the data blocks is written into the blockchain.

[0068] Compared with existing technologies, the beneficial effects of the virtual reality-based customs declaration process teaching method and system provided by this invention are as follows: This invention embeds an extended Kalman filter inside the student terminal, simultaneously incorporating position, velocity, angular velocity, and attitude into the state vector. Driven by the FPGA clock, it completes the Riccati equation iteration, dynamically determining the optimal weights using theoretical prediction variance and measurement noise variance, instantly eliminating pulses caused by vibration and rotation. After filtering, the peak-to-peak value of coordinate jitter is reduced to one-tenth of its original value, eye-tracking hotspots no longer jump, and teachers can clearly identify the student's gaze sequence during playback, significantly improving the repeatability of the competency profile. Since denoising occurs at the data source, the backend receives an asymptotically optimal estimate, eliminating the need for secondary smoothing and saving CPU resources. At the same time, covariance convergence ensures that the filter has the ability to adaptively amplify measurement weights to sudden noise, neither missing real and rapid operations nor being skewed by false jitter, achieving a balance of "smoothing pulses and preserving intent." Therefore, the system, for the first time in a VR teaching scenario, simultaneously meets the requirements of real-time performance, denoising, and reliability, laying a reliable data foundation for subsequent on-chain evidence storage and employer verification. Attached Figure Description

[0069] Figure 1 A flowchart for a virtual reality-based teaching method for customs declaration procedures;

[0070] Figure 2This is a flowchart of step S50 of the present invention. Detailed Implementation

[0071] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0072] This invention provides a virtual reality-based customs declaration process teaching system, comprising:

[0073] Virtual reality scene generation module, used to construct and output three-dimensional virtual scene data;

[0074] The role assignment module communicates with the virtual reality scene generation module and is used to send role identification data to the student terminals that enter the virtual scene.

[0075] The document simulation module is used to store and output electronic form data to the student's terminal;

[0076] The process decision module communicates with the document simulation module and is used to output deployment instruction data based on electronic form data.

[0077] The behavior acquisition module is used to collect real-time operation trajectory data generated by the student's terminal within the virtual scene; and

[0078] The teaching management backend communicates with the behavior collection module to receive and store operation trajectory data.

[0079] In the above technical solution, the virtual reality scene generation module includes:

[0080] Point cloud data interface, used to receive external laser point cloud data;

[0081] A 3D reconstruction unit, connected to a point cloud data interface, is used to convert laser point cloud data into triangular mesh data; and

[0082] The rendering unit, connected to the 3D reconstruction unit, is used to render triangular mesh data into a frame sequence in real time and output it to the student terminal.

[0083] Furthermore, in the above technical solution, the document simulation module includes:

[0084] A form database for storing at least one electronic form template;

[0085] The input interface is used to receive field data entered by students' terminals; and

[0086] The verification unit, connected to the input interface, is used to compare field data with preset rule data and output verification result data.

[0087] Furthermore, in the above technical solution, the behavior acquisition module includes:

[0088] Handle gesture sensor, used to output the spatial coordinate data of the handle;

[0089] Eye-tracking sensors are used to output data on the direction of eye movement; and

[0090] The timing integration unit, connected to the handle posture sensor and eye-tracking sensor, is used to align spatial coordinate data and gaze direction data according to timestamps and generate operation trajectory data.

[0091] like Figure 1 As shown, the present invention provides a flowchart of a virtual reality-based customs declaration process teaching method, including the aforementioned virtual reality-based customs declaration process teaching system, wherein the method includes the following steps:

[0092] S10: Construct a 3D virtual scene and distribute it to student terminals;

[0093] S20: Send role identification data to the student's terminal;

[0094] S30: Output electronic form data to the student's terminal;

[0095] S40: Receives field data returned by the student terminal and generates deployment instruction data according to preset rule data;

[0096] S50: Collects operation trajectory data generated by student terminals in the virtual scene; and stores the operation trajectory data in the teaching management backend.

[0097] The specific implementation methods of the above steps are described in detail below.

[0098] Running in a local area network or 5G private network environment, the teaching management backend acts as the master time node, uniformly distributing time signals to student terminals, teacher terminals, and various modules to ensure that the five data streams of scene, role, document, deployment, and behavior advance on the same time base. This method includes the following steps:

[0099] The teaching management backend automatically matches the current customs declaration case package according to the course schedule and broadcasts the metadata frame containing the 3D virtual scene index code, number of roles, document template version number and risk rule timestamp to the virtual reality scene generation module, role allocation module, document simulation module and process decision module to complete the pre-class synchronization.

[0100] After receiving the metadata frame, the virtual reality scene generation module reads the triangular mesh data packet corresponding to the index code from the local cache, generates a 90fps frame sequence through the GPU-side rendering pipeline, and continuously pushes it to the student terminal in UDP multicast mode. The student terminal presents a 1:1 scale virtual scene of the port inspection area in the head-mounted display.

[0101] After receiving the metadata frame, the role allocation module cyclically sends role identification data to the logged-in student terminals based on the role quantity register value. This data is an 8-byte UID, with the high 4 bytes identifying the role category and the low 4 bytes identifying the terminal serial number, ensuring that roles are not duplicated or missing within the same virtual scene.

[0102] After receiving the metadata frame, the document simulation module loads the corresponding version of the electronic form template data into the DDR and pushes the blank customs declaration interface to the student terminal through the WebSocket channel. The student completes the field entry in the VR overlay panel, and the field data is transmitted back to the verification unit of the document simulation module in real time.

[0103] The process decision module continuously monitors the verification result data output by the verification unit. Once it receives the "pass" flag, it immediately compares the field data with the local rule database and generates deployment instruction data based on the comparison result. This data is 16 bytes long, with the high 8 bytes being the instruction code and the low 8 bytes being the parameters. The instruction code is broadcast to the student terminal via the RS-485 differential bus, triggering the virtual scene state machine to jump to the "inspection" or "release" branch.

[0104] The behavior acquisition module uses an FPGA as the timing master controller. It acquires the spatial coordinate data of the handle posture sensor at a default 1kHz interval and the gaze direction data of the eye-tracking sensor at a default 90Hz interval. Inside the FPGA, a 64-bit timestamp is used to hard align the two different frequency data to generate 32 bytes of operation trajectory data, which is then uploaded to the teaching management backend in real time via a TLS-TCP link.

[0105] The teaching management backend writes the received operation trajectory data to the local SSD, and at the same time calls the capability profiling algorithm to map the trajectory data to four dimensions of compliance, timeliness, communication and emergency response, generating capability profiling data, and pushes it to the teacher's terminal via WebSocket. Teachers can view the radar chart and playback on VR or PC.

[0106] After the course ends, the teaching management backend will package all operation trajectory data of that session into a Merkle tree structure, calculate the root hash value, and write the hash value into the blockchain through a lightweight consortium blockchain gateway to solidify the grades. The employer node can complete the on-chain verification within 3 seconds by scanning the certificate QR code.

[0107] The metadata frame provides a unified time base for each module, ensuring that the five data streams of scene, role, document, deployment, and behavior advance on the same time base, avoiding the time drift caused by traditional software stacking;

[0108] The role identification data, field data, deployment instruction data, and operation trajectory data are all transmitted between modules in fixed-length frames. The transmission length is agreed upon once at the interface to reduce protocol parsing overhead.

[0109] The ability profiling algorithm uses an LSTM time series model. The input is a 42-dimensional behavior sequence, and the output is a 4-dimensional score. The model weights are trained offline in the background, and the inference is run on the CPU side. The single time is less than 200ms, which does not increase the student's waiting time.

[0110] The blockchain is a lightweight consortium blockchain with PoA consensus, with a block interval of 3 seconds. The block header only stores the hash and timestamp, resulting in low on-chain storage overhead, making it suitable for the high-frequency evidence storage needs of universities.

[0111] Among them, the role identification data is used to load the corresponding role's task package and field of view permissions in the virtual scene, so that trainees can only view the document fields and scene areas related to their own role, thus avoiding information overload;

[0112] The field data consists of customs declaration text entered by the trainee in the VR panel, which is encoded in UTF-8 and has a length limit of 1500 bytes to ensure that the document simulation module can complete reading and writing within one DMA transfer during verification.

[0113] Among them, the deployment command data is used to drive the state machine of the virtual scene. The high 4 bits of the command code are "inspection" or "release", the low 4 bits are the risk level, and the parameter bytes are used to specify the inspection method or sampling ratio, so as to achieve second-level feedback consistent with the real customs system.

[0114] Among them, the operation trajectory data is used to replay the student's perspective on the teacher's end. During the replay, the teacher can pause, slow down, and rotate 360 ​​degrees to help discover the student's operation habits and knowledge gaps.

[0115] Among them, the competency profile data is displayed in the form of a radar chart. Teachers can export PDF transcripts with one click. The transcripts include blockchain transaction hashes, which employers can scan to verify the authenticity.

[0116] The blockchain writing process is completed asynchronously in the background, without affecting the start of the next lesson, thus ensuring the continuity of the teaching rhythm.

[0117] like Figure 2As shown, S51: For the operation trajectory data generated by the student terminal in the virtual scene, which contains erroneous operation, vibration or rotation noise, a constant velocity state model of the handle spatial coordinates and eye line direction is established. The theoretical prediction variance and sensor measurement noise variance are used as inputs. The Kalman gain is solved iteratively through the Riccati equation. The theoretical prediction value and the measured value are dynamically weighted and averaged under the meaning of minimum mean square error to obtain the filtered trajectory data.

[0118] S52: The filtered trajectory data is used as the operation trajectory data for subsequent storage and capability profile generation.

[0119] The specific steps are as follows: Noise sources include rapid rotation of the extended handle, momentary vibration of the head-mounted display, or accidental wrist shaking by the student, all of which generate high-frequency pulses in the inertial measurement unit (IMU) and eye-tracking data. Traditional low-pass filtering introduces phase lag, affecting real-time performance. This embodiment implements an extended Kalman filter (EKF) within the FPGA, increasing the state vector to 12 dimensions.

[0120] ;

[0121] in For quaternion pose, Using angular velocity, a joint dynamic model of constant velocity and angular velocity is formed, which can explicitly estimate and subtract vibration and rotation components.

[0122] Forecasting Phase (Time Update):

[0123] Linear part: Position and velocity according to Extrapolation.

[0124] Nonlinear component: Angular velocity drives quaternion updates, using a first-order approximation:

[0125] ;

[0126] in It is an obliquely symmetric matrix for angular velocity.

[0127] Process noise covariance Diagonal setting: Linear velocity noise angular velocity noise This reflects that the vibration energy mainly falls in the angular velocity channel.

[0128] Observation phase (measurement update): Observation vector It is composed of IMU quaternions and eye-tracking back-projection coordinates.

[0129] The observation equation is nonlinear: Therefore, a first-order Jacobian matrix H (a 12×6 sparse matrix) is instantiated within the FPGA, and parallel computation is performed using the DSP48E2 hard core. .

[0130] Measurement noise R real-time adaptive: angular velocity channel variance exceeds threshold within a short period of time. When this occurs, it is determined to be a vibration / rotation event. Dynamically expanded to Suppress mutations; after the event ends The exponential decay back to the original value ensures that the filter both smooths the pulse and converges quickly.

[0131] Dynamic model construction: Selecting state vectors:

[0132] ,

[0133] in, It is a three-dimensional coordinate system (handle or line-of-sight projection coordinate system). To correspond to the speed; a constant speed model is adopted:

[0134] ,

[0135] We take 1 / 90s (eye-tracking sampling period). This model concentrates the uncertainty into process noise. This is done through human intervention, rather than through manual correction, which aligns with the Kalman 'model confidence adaptability' principle.

[0136] Variance description and Riccati equation:

[0137] Predicting covariance extrapolation: ;

[0138] Update process: ;

[0139] in For process noise covariance, To measure the noise covariance, both are obtained through static calibration. Iteration of the Riccati equation makes... It converges to a steady state within 10 sampling periods, ensuring filter stability and asymptotic optimality.

[0140] The essence of optimal weights in Kalman gain:

[0141] ;

[0142] When the theoretical prediction variance ( When ) is much larger than R, The value tends towards 1, representing a system trust measurement; conversely, it tends towards 1. The result tends towards 0, indicating a trust model prediction. This dynamic weighted averaging mechanism is completed within the FPGA using only a single matrix multiplication and addition, with a latency of <1µs.

[0143] Using the Xilinx DSP48E2 hard core for single-precision floating-point multiplication and addition, 1 kHz IMU data and 90 Hz eye-tracking data are time-division multiplexed on the same IP core at a 125 MHz clock speed, with resource usage of <6%. After filtering, the peak-to-peak noise of the 3D coordinates is reduced from ±2.1 mm to ±0.3 mm, and the jitter radius of the eye-tracking hotspot is reduced from 1.8° to 0.4°, without the need for secondary smoothing in the background.

[0144] The improved trajectory accuracy is directly reflected in the "compliance" dimension of the competency profile: the standard deviation of two lessons for the same student decreases by 37%, and teachers can accurately determine whether a student has skipped key sections without manual correction during playback, thus truly solving the common pain point of "inaccurate operation trajectory leading to distorted evaluation".

[0145] Through the above steps, this invention achieves full-process hard-wired synchronization of "scenario-role-document-deployment-behavior-evidence" at the method level, enabling teachers to complete pre-class preparation within 60 seconds, and allowing students to receive rule feedback synchronized with the real customs system within a single class, while ensuring that grades cannot be tampered with. This provides an efficient, reliable, and scalable virtual reality teaching implementation method for port logistics talent training.

[0146] Furthermore, in the above technical solution, the steps for constructing a three-dimensional virtual scene include:

[0147] Receive external laser point cloud data;

[0148] Convert laser point cloud data into triangular mesh data; and

[0149] Triangular mesh data is rendered in real time to generate a frame sequence and distributed to student terminals.

[0150] Furthermore, in the above technical solution, the step of generating deployment command data includes:

[0151] Compare the field data with the rule database;

[0152] If the comparison result meets the first condition, then the first deployment command data is output; and

[0153] If the comparison result meets the second condition, then the second deployment command data will be output.

[0154] Furthermore, in the above technical solution, the steps for collecting operation trajectory data include:

[0155] Collect the spatial coordinate data of the handle at the first time interval;

[0156] Data on eye gaze direction was collected at the second time interval; and

[0157] Align spatial coordinate data and line-of-sight data with timestamps to generate operation trajectory data.

[0158] Furthermore, in the above technical solution, the method further includes:

[0159] After storing the operation trajectory data, the system generates competency profile data based on the operation trajectory data and outputs the competency profile data to the teacher's terminal.

[0160] Furthermore, in the above technical solution, the method further includes:

[0161] Package the operation trajectory data into data blocks; and write the hash value of the data blocks into the blockchain.

[0162] Specifically, the principle of this invention is as follows: Based on a constant velocity and angular velocity extended state model, the invention incorporates quaternion attitude and three-dimensional velocity into the state vector, forming a twelve-dimensional linear-nonlinear hybrid system. The angular velocity channel is given a larger process noise covariance, making the filter mathematically "skeptical" of sudden increases in angular velocity changes. When the measurement noise covariance expands in real time, the Kalman gain automatically decreases, and the theoretical prediction value dominates, thereby instantly suppressing vibration peaks. The covariance update still strictly follows the Riccati equation, ensuring that the error ellipse converges to a steady state within several sampling periods, achieving asymptotically optimal estimation. In terms of hardware, the FPGA completes matrix multiplication and covariance extrapolation using the same clock, with latency controlled at the microsecond level, ensuring that high-sampled data does not accumulate. The filtered output has the same length as the original packet, allowing direct storage at the backend without protocol modification, forming a closed loop of "front-end denoising, back-end reliability." As a result, the system eliminates the uncertainties caused by vibration, rotation, and misoperation at the edge of the sensor, so that all subsequent teaching analysis and achievement verification are based on a real and smooth operating trajectory, fundamentally solving the industry problem of "inaccurate source" in VR customs declaration training.

[0163] To better understand and implement this invention, the following is an embodiment 1 of a specific application scenario:

[0164] This embodiment uses a 40-seat customs declaration training room as a platform to deploy the teaching system at the hardware level. All modules are fixed in a standard 19-inch rack and interconnected with Wi-Fi 6E APs via the campus gigabit Ethernet. It can be deployed directly in existing classrooms without additional protocol conversion.

[0165] Virtual Reality Scene Generation Module: A 2U high "scene node" is installed at the top of the rack. The front provides an SFP+ optical port as a point cloud data interface, allowing for the insertion of .las files copied from the port scanning vehicle. Internally, it uses a Xilinx Zynq UltraScale+MPSoC, with a quad-core ARM processor running the Poisson reconstruction algorithm to convert the point cloud into a PLY triangular mesh. The FPGA section instantiates an OpenCL kernel, responsible for normal vector estimation and mesh simplification, outputting a single-precision vertex buffer. An NVIDIA RTX A4000 GPU card is inserted at the rear of the node, driving the Vulkan rendering pipeline to render the triangular mesh data into a 90fps H.264 frame sequence in real time. The GPU-side NVENC output is multicast to the classroom AP via RJ45, with a multicast address of 239.192.0.1:5004 and a peak bandwidth not exceeding 250 Mbps, sufficient for simultaneous roaming of 32 VR headsets.

[0166] Role Assignment Module: Shares a rack with scene nodes and adopts a 1U ARM server form factor. The SoC has a built-in 256kBEEPROM, which pre-stores five types of role UID tables; after receiving the "Start Teaching" TCP packet on the network port, the MCU assigns UIDs in a loop according to the terminal MAC address, and unicasts them through the same AP with an 802.11 QoS AC_VO queue. Each packet is 64B, and the air latency is consistently below 3ms, ensuring that students enter their roles simultaneously.

[0167] Document Simulation Module: Inserted into the middle of the rack as a half-width 1U "form node". The node base plate integrates 32MB QSPIFlash to store the latest customs declaration xsd template; the front panel provides a USB-C input interface for receiving UTF-8 field data entered by the trainee's controller / keyboard; the Cortex-M7 hard core has a built-in regular expression engine that compares the fields with the rule segments in the Flash one by one. The comparison results are forwarded to the process decision module via gigabit network port in UDP 1500 B frames. The actual verification latency is less than 20ms, realizing "immediate judgment after filling".

[0168] Process Decision Module: Employing a heterogeneous FPGA+ARM SoC, it is installed in the PCIe 3.0 x 4 slots on the backplane of the cabinet. On the ARM side, the deployment scoring executable file runs. After receiving UDP packets from the document node, it compares them with the risk weight table in the eMMC to generate a 16-byte deployment instruction. On the FPGA side, the instruction is encapsulated into an RS-485 frame, 115200 8E1 format, broadcast at a 100ms cycle. The differential signal directly drives the MCU of the student terminal, triggering the VR scene state machine transition. The hardware path ensures that the end-to-end latency is less than 50ms, consistent with the experience of receiving receipts in a real customs system.

[0169] Behavior Acquisition Module: Each VR headset incorporates a BMI160 IMU and a Tobii EyeChip; a Lattice LFE5U-25 FPGA is soldered to the rear of the headset as a timing integration unit. IMU data enters the FPGA via a 1kHz I²C, and eye-tracking data enters the FPGA via a 90Hz USB isochronous endpoint; the FPGA uses a 125MHz global clock to timestamp each data packet with 64 bits, and... The cycle (64ms) is forcibly merged, outputting a 32B trajectory packet, which is transmitted back to the "acquisition switch" in the cabinet via the gigabit Type-C port at the rear of the head-mounted display, and then uploaded to the teaching management backend via TCP-TLS. There is no software caching throughout the process, ensuring that the error between the controller coordinates and the line of sight is less than 12ms, and that the teacher's viewpoint is completely synchronized with the student's operation during playback.

[0170] The teaching management backend is handled by a 4U industrial PC with an i7-12700 CPU, 32GB RAM, and a 4TB RAID5 SSD. An embedded Nginx-gRPC gateway listens on port 443, simultaneously receiving data from scene nodes and acquisition switches. The SSD stores the raw data paths, MongoDB stores structured fields, and the Vue3 web frontend provides one-click playback, radar chart generation, and blockchain writing capabilities. An RJ45 connector on the backend side panel connects to a Raspberry Pi 4 light node running geth-clique. Each completed lesson writes the Merkel root hash to the consortium blockchain in 3 seconds. Only the hash and timestamp are stored on the blockchain, resulting in less than 1kB of storage per lesson, yet ensuring the grades are immutable. Employers can verify the authenticity of the grades within 3 seconds by scanning a QR code.

[0171] Rack power supply and heat dissipation: It adopts a 3kW 1+1 redundant power supply, which can still run continuously for 7×24h at an ambient temperature of 45°C; four 12038 ball bearing fans are installed in the middle, with front and rear air ducts. Both GPU and FPGA nodes are equipped with aluminum fin heat sinks. The actual measured noise of the whole machine is less than 52dB(A), which does not affect normal teaching.

[0172] Through the above hardware-level implementation, this system completely integrates "scene generation, role allocation, document verification, risk control, behavior collection, and trusted evidence storage" into a standard server rack. Schools only need to provide gigabit network and Wi-Fi to complete the deployment within 30 minutes, realizing "plug and play" and significantly reducing the hardware threshold and maintenance cost of virtual reality customs declaration teaching.

[0173] The following is a specific embodiment 2 of the present invention, in which the specific implementation method of each step is described in detail below. A customs college in East China conducted a week-long intensive training program with 60 students, divided into five groups rotating through the campus. Classes were held for 6 hours daily for five consecutive days. The college's network provided gigabit speeds to the classrooms and full Wi-Fi 6E coverage, but there was no dedicated computer lab. Traditional solutions required hiring external lecturers, printing documents, and traveling to and from the port, which was costly and could not cover the details of risk control. The college decided to adopt the "Virtual Reality-Based Customs Declaration Process Teaching System" and its supporting methods of the present invention, completing all training in an ordinary 80m² lecture hall.

[0174] A 2-meter-wide passageway is reserved at the front of the classroom, with a 42U server rack (hereinafter referred to as the "main rack") placed in the center. All modules are integrated into the main rack using 19-inch plug-in cabinets. The rack is equipped with four swivel casters at the bottom, allowing it to be moved to the next classroom within 30 minutes. The main rack is powered by two 16A AC mains power supplies and a 1+1 redundant UPS, ensuring 15 minutes of backup power during power outages. The internal layout of the rack, from top to bottom, is as follows:

[0175] Virtual Reality Scene Generation Module (2U): Front: SFP+ optical port connects to the campus backbone network, enabling real-time download of the latest point cloud from the port; the side provides 2×USB 3.0 ports for on-site copying. Internal: Zynq UltraScale+MPSoC+RTX A4000, with a single-precision floating-point computing power of 19 TFLOPS, capable of converting a 2GB point cloud into a 5M vertex triangular mesh within 15 minutes, and continuously outputting a 90fps H.264 stream with a peak bitrate of 250Mbps, supporting simultaneous roaming of 60 headsets.

[0176] Role Assignment Module (1U): The SoC uses an STM32MP157 dual-core A7 processor, 1 GB of DDR memory, and 256kB of EEPROM pre-stored five types of role UIDs. Upon receiving a "Lesson Start" TCP packet, the network port automatically and cyclically assigns UIDs and unicasts them via the same Wi-Fi 6E AP using QoS AC_VO, with an over-the-air latency of <3ms, ensuring that 60 students can simultaneously enter their roles without conflict.

[0177] Document Simulation Module (Half-width 1U): The baseboard integrates 32 MB QSPI Flash, storing the latest customs declaration xsd template and 337 regular expression rules; the front-end USB-C port receives input from the handle / keyboard, and the Cortex-M7 hard-core regular expression engine completes the comparison within 20ms. The comparison result is forwarded to the process decision module via gigabit UDP port.

[0178] Process Decision Module (1U): FPGA+ARM heterogeneous SoC, the ARM side runs the deployment scoring executable file, the FPGA side encapsulates the 16-byte deployment instructions into RS-485 frames, broadcasts them in 100ms cycles, and the differential signal directly drives the student terminal MCU, with an end-to-end delay of <50ms.

[0179] Behavior acquisition aggregation node (1U switch form): 48 Gigabit Ethernet ports + 4 SFP+. The trajectory data of all students' head-mounted displays are aggregated via TLS-TCP and sent to the teaching management backend via two 10G uplinks.

[0180] Teaching management backend (4U industrial PC): i7-12700 / 32 GB / 4 TB RAID5 SSD, running Nginx-gRPC gateway, MongoDB, and Vue3 frontend;

[0181] The side panel connects to a Raspberry Pi 4 light node, running the geth-clique PoA chain. After each course is completed, the Merkel root is written to the consortium chain, with a block time of 3 seconds. Only the hash is stored on the chain, and the storage usage is <1kB / time.

[0182] Student Terminal: Each terminal includes a VR headset, controllers, eye-tracking module, Type-C gigabit Ethernet card, and power bank. The headset has an embedded FPGA (LFE5U-25) as a timing integration unit. IMU 1 kHz and eye-tracking 90Hz data are hard-aligned within the FPGA with an error of <12ms. Tracking packets are uploaded to the aggregation switch in real time via Type-C. The terminal weighs <700g, and the power bank provides 3 hours of continuous battery life, sufficient for half a day of continuous teaching.

[0183] On the first day of the course, at 8:00 AM, teachers swipe their cards to log in to the teaching management backend, select the "General Trade Import Case" metadata package, and click "Start Course." The system then executes the following steps:

[0184] Step 1: The backend sends the index code to the scene generation module, and the module completes mesh reconstruction and pushes the stream in 15 minutes;

[0185] Step 2: The role allocation module distributes UIDs in a cyclical manner based on MAC address, and 60 trainees simultaneously obtain roles such as "customs broker", "cargo owner", and "customs officer".

[0186] Step 3: The document simulation module pushes the blank customs declaration form to their respective VR panels, and the trainees complete the pre-entry in the headset, with the field data being transmitted back in real time.

[0187] Step 4: The process decision module returns a control instruction within 20ms. 30% of trainees trigger the "H986 Machine Inspection" branch, and the VR scene automatically plays the machine inspection channel animation.

[0188] Step 5: The behavior acquisition module records the handheld coordinates and eye-tracking hotspots throughout the process, and the background receives a 32B trajectory packet every 64ms;

[0189] Step 6: 3 minutes before the end of get out of class, the background calls the LSTM model to generate a four-dimensional competency profile, and the radar chart on the teacher's web interface is updated in real time.

[0190] Step 7: The teacher clicks "Store Evidence". The system packages the trajectory of 60 people into a Merkel tree, writes the root hash into the consortium blockchain, and automatically generates a QR code. Students can scan the code with their mobile phones to download the electronic transcript with the on-chain hash.

[0191] The five-day, 30-hour training program covered five trade methods, three inspection scenarios, and one system outage emergency drill. Student surveys showed: scenario realism 4.7 / 5, rule synchronization satisfaction 4.8 / 5. Compared to traditional training, the average student saved 8 hours of travel time, and instructors reduced paper document printing by 1200 pages. Employers verified grades via on-chain hashing in an average of 2.8 seconds, with certificate credibility rated 4.9 / 5. The system ran continuously for 30 hours without restarts, with a maximum CPU temperature of 68°C and a network packet loss rate of <0.01%, meeting the requirements for intensive teaching use.

[0192] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A virtual reality-based customs declaration process teaching system, characterized in that, include: Virtual reality scene generation module, used to construct and output three-dimensional virtual scene data; The role assignment module is communicatively connected to the virtual reality scene generation module and is used to send role identification data to the student terminals that enter the virtual scene. The document simulation module is used to store and output electronic form data to the student's terminal; The process decision module is communicatively connected to the document simulation module and is used to output deployment instruction data based on the electronic form data. The behavior acquisition module is used to collect the operation trajectory data generated by the student's terminal in the virtual scene in real time; as well as The teaching management backend is connected to the behavior acquisition module and is used to receive and store the operation trajectory data.

2. The virtual reality-based customs declaration process teaching system according to claim 1, characterized in that, The virtual reality scene generation module includes: Point cloud data interface, used to receive external laser point cloud data; A 3D reconstruction unit, connected to the point cloud data interface, is used to convert the laser point cloud data into triangular mesh data; and The rendering unit, connected to the 3D reconstruction unit, is used to render the triangular mesh data into a frame sequence in real time and output it to the student terminal.

3. The virtual reality-based customs declaration process teaching system according to claim 2, characterized in that, The document simulation module includes: A form database for storing at least one electronic form template; The input interface is used to receive field data entered by students' terminals; and The verification unit, connected to the input interface, is used to compare the field data with preset rule data and output verification result data.

4. The virtual reality-based customs declaration process teaching system according to claim 3, characterized in that, The behavior acquisition module includes: Handle gesture sensor, used to output the spatial coordinate data of the handle; Eye-tracking sensors are used to output data on the direction of eye movement; and The timing integration unit, connected to the handle posture sensor and the eye-tracking sensor, is used to align the spatial coordinate data and the gaze direction data according to the timestamp and generate the operation trajectory data.

5. A virtual reality-based teaching method for customs declaration procedures, comprising the virtual reality-based customs declaration procedure teaching system described in any one of claims 1-4, characterized in that, The method includes the following steps: S10: Construct a 3D virtual scene and distribute it to student terminals; S20: Send role identification data to the student's terminal; S30: Output electronic form data to the student's terminal; S40: Receives field data returned by the student terminal and generates deployment instruction data according to preset rule data; S50: Collect the operation trajectory data generated by the student terminal in the virtual scene; and store the operation trajectory data in the teaching management backend.

6. The teaching method for customs declaration process based on virtual reality according to claim 5, characterized in that, The steps for constructing a three-dimensional virtual scene include: Receive external laser point cloud data; Convert the laser point cloud data into triangular mesh data; and The triangular mesh data is rendered in real time to generate a frame sequence and distributed to the student terminals.

7. The teaching method for customs declaration process based on virtual reality according to claim 6, characterized in that, The steps for generating deployment command data include: The field data is compared with the rule database; If the comparison result meets the first condition, then the first deployment command data is output; and If the comparison result meets the second condition, then the second deployment command data will be output.

8. The teaching method for customs declaration process based on virtual reality according to claim 7, characterized in that, The steps for collecting operation trajectory data include: Collect the spatial coordinate data of the handle at the first time interval; Data on eye gaze direction was collected at the second time interval; and Align the spatial coordinate data with the line-of-sight data according to the timestamp to generate operation trajectory data.

9. The teaching method for customs declaration process based on virtual reality according to claim 8, characterized in that, The method, S50, includes: S51: To address the erroneous operation, vibration, or rotation noise contained in the operation trajectory data generated by the student terminal in the virtual scene, a constant velocity state model of the handle spatial coordinates and eye line direction is established. Using the theoretical prediction variance and sensor measurement noise variance as inputs, the Kalman gain is solved iteratively through the Riccati equation, so that the theoretical prediction value and the measured value are dynamically weighted and averaged under the meaning of minimum mean square error to obtain the filtered trajectory data. S52: The filtered trajectory data is used as the operation trajectory data for subsequent storage and capability profile generation.

10. The teaching method for customs declaration process based on virtual reality according to claim 9, characterized in that, The method further includes: After storing the operation trajectory data, capability profile data is generated based on the operation trajectory data; and the capability profile data is output to the teacher terminal. The operation trajectory data is packaged into data blocks; and the hash value of the data blocks is written into the blockchain.

Images