Adaptive interpolation restoration device and method for automatic driving continuous space-time data transmission

Abstract

The application discloses an automatic driving continuous space-time data transmission adaptive interpolation restoration device and method, belongs to the automatic driving data transmission and processing technical field, and solves the problems of high bandwidth occupation, insufficient restoration precision, poor scene adaptability and the like of an existing transmission scheme. The device contains six modules of data acquisition, vehicle end edge calculation, adaptive interpolation restoration, error checking and the like. The method is that multiple types of space-time data are collected at the vehicle end, and an identifier is added, dynamic characteristic parameters are extracted to divide low, medium and high dynamic driving scenes, adaptive configuration of start and end point intervals is carried out, and customized data packaging is carried out, data is restored according to a scene matching interpolation algorithm at the cloud end, error checking is carried out through original space-time data cached at the vehicle end, and if a threshold value is exceeded, supplementary transmission is triggered and the start and end point intervals are reduced. The application greatly reduces data transmission quantity and bandwidth occupation, improves restoration precision and data reliability, reduces vehicle end energy consumption, and adapts to real-time data interaction requirements of automatic driving in all working conditions.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving data transmission and processing technology, and in particular to an adaptive interpolation restoration device and method for continuous spatiotemporal data transmission in autonomous driving. Background Technology

[0002] In the autonomous driving technology system, continuous spatiotemporal data, such as LiDAR 3D point clouds, GPS positioning trajectories, and IMU attitude data, are the core data support for achieving environmental perception and path planning. This type of data is characterized by "high sampling rate, high dimensionality, and strong temporal continuity." Currently, the industry's transmission solutions for this type of data are mainly divided into the following two categories:

[0003] (1) Full Transmission Scheme: Continuous spatiotemporal data collected by vehicle-side sensors is transmitted directly to the cloud or roadside unit via the 5G / V2X communication module without simplification. During transmission, only a general compression algorithm is used to perform lossless compression on the data. After receiving the data, the cloud directly decodes and restores the original spatiotemporal data. This scheme is currently the mainstream basic scheme for autonomous driving data transmission and is widely used in scenarios such as high-precision map updates and multi-vehicle collaborative perception.

[0004] (2) Keyframe Filtering and Transmission Scheme: Based on the temporal variation of the data, keyframe data with variations exceeding a threshold are filtered for transmission, while redundant frames with smaller variations are discarded. For example, when a vehicle is traveling at a constant speed in a straight line, the continuous frame data collected by the lidar has small differences, and only one keyframe is transmitted every 5-10 frames; when the vehicle is turning or encountering obstacles, the keyframe filtering frequency is increased. After receiving the keyframes, the cloud uses simple linear interpolation to fill in the missing intermediate frame data. This scheme is commonly used in low-to-medium precision vehicle-road cooperative data interaction scenarios.

[0005] The two existing transmission schemes have significant technical shortcomings in highly dynamic autonomous driving scenarios, and they do not combine edge computing architecture to achieve coordinated optimization of data processing and transmission, as detailed below:

[0006] (1) Deficiencies of the full-volume transmission scheme: ① High bandwidth consumption: The high sampling rate of continuous spatiotemporal data results in a huge amount of data. Even after general compression, the transmission bandwidth requirement can still reach tens of Mbps. In the scenario of multi-vehicle concurrent transmission in vehicle-road cooperative systems, it is easy to cause network congestion and transmission delay, which cannot meet the needs of real-time decision-making in autonomous driving; ② High energy consumption: Long-term high-load transmission of vehicle-side communication modules will significantly increase vehicle energy consumption and affect range; ③ Waste of redundant data: There is a large amount of temporal redundancy in continuous spatiotemporal data. Full-volume transmission will cause ineffective occupation of communication resources. The core reason for the defects is that the temporal continuity characteristics of continuous spatiotemporal data have not been specifically simplified, the local processing capabilities of edge computing have not been utilized to reduce invalid data transmission, and relying solely on general compression algorithms cannot fundamentally reduce the amount of data.

[0007] (2) Deficiencies of the key frame screening and transmission scheme: ① Insufficient restoration accuracy: The existing scheme uses a fixed threshold to screen key frames. In high dynamic driving scenarios, such as when the vehicle makes a sharp turn, the data changes drastically within the key frame interval. Simple linear interpolation cannot accurately restore the dynamic change trend of the data, resulting in deviations in the data restored in the cloud, which affects the accuracy of perception and decision-making; ② Poor adaptability: The fixed key frame screening threshold cannot be adapted to different dynamic driving scenarios. For example, the optimal key frame interval differs greatly between high-speed uniform speed and urban congestion scenarios. In such scenarios, there is a problem that the lack of key frames in high dynamic driving scenarios leads to restoration distortion, while the excessive number of key frames in low dynamic driving scenarios leads to bandwidth waste; ③ No error verification mechanism: The data restored by interpolation has not been verified for accuracy. If packet loss or noise interference occurs in the key frame data during transmission, it will directly lead to the failure of the restored data, and it will not be detected in time; ④ Failure to combine the advantages of edge computing: Real-time judgment and data preprocessing of dynamic driving scenarios are not implemented in the vehicle-side edge computing module (2), resulting in a lag in the response of the screening logic. The defects are caused by the fact that neither the keyframe filtering logic nor the interpolation algorithm is adaptively optimized in combination with the dynamic characteristics of the data, and there is a lack of end-to-end accuracy guarantee mechanism for transmission and restoration, and the low-latency processing capability of edge computing is not fully utilized. Summary of the Invention

[0008] The purpose of this invention is to provide an adaptive interpolation restoration device and method for continuous spatiotemporal data transmission in autonomous driving, so as to overcome or at least mitigate at least one of the above-mentioned defects of the prior art.

[0009] To achieve the above objectives, the present invention provides an adaptive interpolation and restoration device for continuous spatiotemporal data transmission in autonomous driving, which includes a data acquisition module, a vehicle-side edge computing module, a vehicle-side data transmission module, a cloud-side data transmission module, an adaptive interpolation and restoration module, and an error verification module.

[0010] The data acquisition module is used to collect multiple types of continuous spatiotemporal data from autonomous vehicles, and after adding a millisecond-level unique timestamp and sensor identifier to each frame of data, it is transmitted to the vehicle-side edge computing module as raw spatiotemporal data.

[0011] The vehicle-side edge computing module is deployed on the vehicle side to extract dynamic feature parameters of several consecutive frames of raw spatiotemporal data. The dynamic feature parameters are compared with preset thresholds to identify the dynamic driving scene level. The start and end time intervals are adaptively configured according to the scene level and data pairs are filtered. For high-dynamic driving scenes, the feature intermediate point with the largest data change is added between the start and end points. The start and end data pairs are customized and encapsulated before the data header information is compressed.

[0012] The vehicle-side data transmission module is used to add a checksum to the encapsulated data pairs from the vehicle-side edge computing module and then transmit them to the cloud data transmission module.

[0013] The cloud data transmission module is used to receive encapsulated data pairs and realize high-speed data forwarding between the vehicle and the cloud;

[0014] The adaptive interpolation restoration module is used to restore continuous spatiotemporal data by selecting a matching interpolation algorithm based on the scene level identifier after performing preliminary verification on the encapsulated data pairs.

[0015] The error verification module is used to call the original spatiotemporal data cached by the vehicle edge computing module and compare it with the restored data of the adaptive interpolation restoration module. If the error exceeds the preset threshold, a retransmission request is sent to the vehicle edge computing module. The vehicle then retransmits the start and end data pairs for that time period and reduces the interval between the start and end points. If the error does not exceed the threshold, the restored data is output to the autonomous driving decision system.

[0016] Furthermore, the vehicle-side edge computing module includes:

[0017] The dynamic driving scene recognition submodule is used to extract dynamic feature parameters from several consecutive frames of raw spatiotemporal data. The dynamic feature parameters include position change, attitude change rate and LiDAR point cloud dispersion. The dynamic feature parameters are compared with preset thresholds to classify three driving scenes: low dynamic, medium dynamic and high dynamic.

[0018] The start and end point filtering submodule is used to configure the frame interval between the start and end points according to the scene level: for low-dynamic driving scenes and medium-dynamic driving scenes, the first frame of the continuous frame is extracted as the start point and the n1th frame as the end point. For high-dynamic driving scenes, the interval between the start and end points is preset to n2 frames, and a feature intermediate point is added between the start and end points. The feature intermediate point is the frame with the largest data change determined by real-time calculation of the data change of adjacent frames. The start and end point data pairs are customized and encapsulated before the data header information is compressed.

[0019] Furthermore, the start and end point filtering submodule is customized into differentiated data encapsulation, specifically as follows:

[0020] The encapsulated content of the start-end data pairs for low, medium, and high dynamic driving scenarios includes the full start-end data, the full end-end data, the dynamic level identifier, the timestamp range, and the sensor identifier.

[0021] The encapsulated content of the start and end point data pairs in high-dynamic driving scenarios also includes obstacle contour coordinates of the preset area of ​​the feature intermediate point and the GPS key steering position.

[0022] Furthermore, in the adaptive interpolation module, the interpolation algorithm that matches the low, medium, and high dynamic driving scenario level identifiers one-to-one is as follows:

[0023] In low-dynamic driving scenarios, a linear interpolation algorithm is used to uniformly interpolate and complete the intermediate frame data based on the numerical change trends of the start and end points.

[0024] In the dynamic driving scenario, a quadratic polynomial interpolation algorithm is used. The least squares method is used to solve the fitting coefficients, and the fitting quadratic polynomial curve is used to restore the gradual trend of data.

[0025] In high-dynamic driving scenarios, a combination algorithm of adaptive spline interpolation and contour constraint interpolation is adopted. The adaptive spline interpolation algorithm combines the three-dimensional data of the starting point, feature intermediate point and the ending point to fit a piecewise spline curve, and each curve segment is continuously constrained by the first derivative of three adjacent points. The contour constraint interpolation is based on prior knowledge of the actual size of the obstacle to constrain the coordinate range of the interpolation points, ensuring that the error between the restored obstacle contour and the actual size is less than the preset error.

[0026] Furthermore, the error verification module includes:

[0027] The data preparation unit verifies the data and calls the original spatiotemporal data within the most recent preset time period cached locally by the vehicle-side edge computing module.

[0028] The error calculation unit is used to extract any two non-starting and non-ending values ​​from the restored data of the adaptive interpolation restoration module in the cloud, request the original spatiotemporal data with the corresponding timestamp from the vehicle edge computing module, and calculate the absolute errors of position error and attitude angle error.

[0029] The accuracy judgment and retransmission mechanism unit is used to preset an error threshold. If the error of the restored data is less than the threshold, it is judged as qualified and output to the autonomous driving decision system; if the error is greater than the threshold, it is judged as unqualified and a retransmission request is sent to the vehicle edge computing module. The request contains the timestamp range of the unqualified data, which triggers the vehicle to retransmit the start and end data pairs of the time period. During the retransmission, the interval between the start and end points is reduced to ensure restoration accuracy.

[0030] This invention also provides an adaptive interpolation restoration method for continuous spatiotemporal data transmission in autonomous driving, comprising:

[0031] Step 1: On the vehicle side, collect continuous spatiotemporal data of autonomous vehicles, add a unique timestamp and sensor identifier to each frame of data as raw spatiotemporal data, extract dynamic feature parameters from several consecutive frames of raw spatiotemporal data, compare the dynamic feature parameters with preset thresholds to identify the dynamic driving scenario level, adaptively configure the start and end time intervals according to the scenario level and filter data pairs. For high dynamic driving scenarios, supplement the feature intermediate point with the largest data change between the start and end points. After custom encapsulating the start and end point data pairs, compress the data header information, add a check code, and then transmit it to the cloud.

[0032] Step 2: In the cloud, after completing the initial verification of the encapsulated data pairs, select the matching interpolation algorithm according to the scene level identifier to restore the continuous spatiotemporal data;

[0033] Step 3: Call the original spatiotemporal data within the most recent preset time period cached in Step 1 and compare it with the restored data in Step 2. If the error exceeds the preset threshold, a retransmission request is sent to the vehicle. The vehicle retransmits the start and end data pairs of the time period and reduces the interval between the start and end points. If the error does not exceed the threshold, the restored data is output to the autonomous driving decision system.

[0034] Furthermore, in step 1, the method for identifying the dynamic driving scenario level specifically includes:

[0035] Dynamic feature parameters are extracted from several consecutive frames of raw spatiotemporal data. The dynamic feature parameters include position change, attitude change rate and LiDAR point cloud dispersion. The dynamic feature parameters are compared with preset thresholds to classify three driving scenarios: low dynamic, medium dynamic and high dynamic.

[0036] The frame interval between the start and end points is configured according to the scene level: for low-dynamic driving scenes and medium-dynamic driving scenes, the first frame of the continuous frame is extracted as the start point and the n1th frame as the end point. For high-dynamic driving scenes, the interval between the start and end points is preset to n2 frames, and a feature intermediate point is added between the start and end points. The feature intermediate point is the frame with the largest data change determined by real-time calculation of the data change of adjacent frames. The start and end point data pairs are customized and encapsulated before the data header information is compressed.

[0037] Furthermore, in step 1, the customized packaging includes:

[0038] The encapsulated content of the start-end data pairs for low, medium, and high dynamic driving scenarios includes the full start-end data, the full end-end data, the dynamic level identifier, the timestamp range, and the sensor identifier.

[0039] The encapsulated content of the start and end point data pairs in high-dynamic driving scenarios also includes obstacle contour coordinates of the preset area of ​​the feature intermediate point and the GPS key steering position.

[0040] Furthermore, in step 2, the method for selecting a matching interpolation algorithm to restore continuous spatiotemporal data based on the scene level identifier includes:

[0041] In low-dynamic driving scenarios, a linear interpolation algorithm is used to uniformly interpolate and complete the intermediate frame data based on the numerical change trends of the start and end points.

[0042] In the dynamic driving scenario, a quadratic polynomial interpolation algorithm is used. The least squares method is used to solve the fitting coefficients, and the fitting quadratic polynomial curve is used to restore the gradual trend of data.

[0043] In high-dynamic driving scenarios, a combination algorithm of adaptive spline interpolation and contour constraint interpolation is adopted. The adaptive spline interpolation algorithm combines the three-dimensional data of the starting point, feature intermediate point and the ending point to fit a piecewise spline curve, and each curve segment is continuously constrained by the first derivative of three adjacent points. The contour constraint interpolation is based on prior knowledge of the actual size of the obstacle to constrain the coordinate range of the interpolation points, ensuring that the error between the restored obstacle contour and the actual size is less than the preset error.

[0044] Furthermore, step 3 specifically includes:

[0045] Step 31: Extract any two non-starting and non-ending values ​​from the restored data in Step 2 in the cloud, request the original spatiotemporal data with the corresponding timestamps from the vehicle, and calculate the absolute errors of position error and attitude angle error.

[0046] Step 32: Preset an error threshold. If the error of the restored data is less than the threshold, it is judged as qualified and output to the autonomous driving decision system. If the error is greater than the threshold, it is judged as unqualified and a retransmission request is sent to the vehicle. The request contains the timestamp range of the unqualified data, triggering the vehicle to retransmit the start and end data pairs of the time period. During the retransmission, the interval between the start and end points is reduced to ensure restoration accuracy.

[0047] The present invention has the following advantages due to the adoption of the above technical solutions:

[0048] This invention achieves a comprehensive improvement in data transmission efficiency, restoration accuracy, scene adaptability, and data reliability through core technologies such as dynamic driving scene recognition, differentiated data filtering and encapsulation, scene matching interpolation algorithms, and end-to-end error verification and retransmission. At the same time, it effectively reduces vehicle-side energy consumption and ensures the real-time performance of data processing, perfectly meeting the real-time decision-making needs of autonomous driving environment perception and path planning. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the adaptive interpolation and restoration device for continuous spatiotemporal data transmission in autonomous driving provided in an embodiment of the present invention. Detailed Implementation

[0050] In the accompanying drawings, the same or similar reference numerals are used to denote the same or similar elements or elements having the same or similar functions. The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0051] In the description of this invention, the terms "center," "longitudinal," "lateral," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting the scope of protection of this invention.

[0052] like Figure 1 As shown, the adaptive interpolation and restoration device for continuous spatiotemporal data transmission in autonomous driving according to an embodiment of the present invention includes a data acquisition module 1, a vehicle-side edge computing module 2, a vehicle-side data transmission module 3, a cloud-side data transmission module 4, an adaptive interpolation and restoration module 5, and an error verification module 6.

[0053] Data acquisition module 1 is used to collect various types of continuous spatiotemporal data from autonomous vehicles, including but not limited to LiDAR point cloud data, GPS (Global Positioning System) trajectory data, and IMU (Inertial Measurement Unit) attitude data. LiDAR point cloud data typically includes three-dimensional coordinates and reflection intensity; GPS trajectory data includes latitude, longitude, altitude, and timestamps; and IMU attitude data includes pitch, roll, and heading angles. During acquisition, a millisecond-level unique timestamp and sensor identifier are added to each frame of data, outputting raw spatiotemporal data with unique timestamps and sensor identifiers. This processing method ensures the temporal consistency and traceability of the data.

[0054] The vehicle-side edge computing module 2 is deployed on the vehicle side to cache the raw spatiotemporal data output by the data acquisition module 1 locally. The caching method can be segmented by timestamp, and the caching time period can be pre-configured, such as 10 seconds. This way, it is cleaned up in real time after the verification is completed, avoiding preprocessing caused by transmission to the remote end, thus saving vehicle-side storage resources.

[0055] The vehicle-side edge computing module 2 extracts dynamic feature parameters from several consecutive frames of raw spatiotemporal data. It compares these dynamic feature parameters with preset thresholds to identify the dynamic driving scene level. Based on the scene level, it adaptively configures the start-end time interval and filters data pairs. For highly dynamic driving scenes, it supplements the intermediate point with the largest data change between the start and end points. The start-end data pairs are then customized and encapsulated, and the header information is compressed. Thus, the vehicle-side edge computing module 2 outputs encapsulated start-end data pairs, which include a dynamic level identifier and a timestamp range. This preprocessing of tasks with high real-time requirements, such as dynamic driving scene identification and start-end selection, reduces the transmission pressure between the vehicle and the cloud, ensuring timely response in highly dynamic driving scenes.

[0056] In one embodiment, the vehicle-side edge computing module 2 includes an integrated dynamic driving scene recognition submodule and a start-end point filtering submodule.

[0057] The dynamic driving scene recognition submodule aims to identify the dynamic level of the current scene by extracting dynamic feature parameters from the data in real time, providing an accurate basis for subsequent start and end point selection. Leveraging the low latency advantage of edge computing, it also ensures the real-time performance of scene recognition. The specific implementation logic is as follows:

[0058] (1) Extract dynamic feature parameters from three consecutive frames of raw spatiotemporal data. The dynamic feature parameters include position change, attitude change rate, and lidar point cloud dispersion. Compare the dynamic feature parameters with a preset threshold to classify driving scenarios into three categories: low dynamic, medium dynamic, and high dynamic. In this embodiment, the extraction period is synchronized with the sensor sampling period, and the extraction period can be set to no more than 10ms.

[0059] ① The change in position can be calculated, for example, using the Euclidean distance formula provided in equation (1), to calculate the spatial position difference d between adjacent frames:

[0060] (1)

[0061] in, , The spatial coordinates of adjacent frames are represented by the differences in spatial location, which include the difference in GPS latitude and longitude and the offset of the centroid of the LiDAR point cloud.

[0062] ② The rate of change of attitude can be calculated, for example, by the following formula (2) to calculate the rate of change of the IMU attitude angle. :

[0063] (2)

[0064] in, , These are the attitude angles of adjacent frames. The IMU attitude angles include yaw, pitch, and roll. t1 and t2 are the corresponding timestamps of adjacent frames. The symbol ' / ' represents absolute value operation, and ' / ' represents division operation.

[0065] ③ The dispersion of lidar point clouds can be calculated by measuring the spatial distribution variance of lidar point clouds in a single frame. This can reflect whether there are sudden obstacles. The larger the variance, the more dispersed the point cloud distribution is, and the greater the probability of the presence of sudden obstacles.

[0066] (2) Dynamic Level Classification: Based on statistical analysis of no less than 1000 sets of real-vehicle scenario data, three dynamic level thresholds are derived, namely low dynamic, medium dynamic, and high dynamic scenario levels. The dynamic feature parameters are compared with the preset thresholds to output the current driving scenario level. Below is an example of a specific classification standard:

[0067] ① Low-dynamic driving scenario: Vehicle travels in a straight line at a constant speed, with a position change of less than 0.5m / frame, an attitude change rate of less than 1° / s, and a LiDAR point cloud dispersion of less than 0.1; ② Medium-dynamic driving scenario: Vehicle turns slowly, follows other vehicles in urban congestion, with a position change of 0.5-2m / frame, an attitude change rate of 1-5° / s, and a LiDAR point cloud dispersion of 0.1-0.3; ③ High-dynamic driving scenario: Vehicle makes sharp turns, avoids obstacles, or travels quickly in rain or snow, with a position change of more than 2m / frame, an attitude change rate of more than 5° / s, and a LiDAR point cloud dispersion of more than 0.3.

[0068] The start and end point filtering submodule, based on the recognition results of the dynamic driving scene recognition submodule, configures the frame interval between the start and end points according to the scene level. The time interval is the number of frames between two adjacent start and end point data pairs, achieving a dynamic balance between transmission volume and reconstruction accuracy.

[0069] For both low-dynamic and medium-dynamic driving scenarios, the first frame of a continuous sequence is extracted as the starting point and the n1th frame as the ending point.

[0070] In high-dynamic driving scenarios, the interval between the start and end points is preset to n² frames, which can be either 2 or 3 frames. A feature intermediate point is added between the start and end points. This feature intermediate point is the frame with the largest data change, determined by real-time calculation of the data changes in adjacent frames. The start and end point data pairs are customized and encapsulated before the header information is compressed. The frame with the largest data change is obtained by real-time calculation of the changes in adjacent frames using the vehicle-side edge computing module 2.

[0071] The start and end point filtering submodule completes the filtering of the start and end point data pairs of three consecutive frames of raw spatiotemporal data in the vehicle-side edge computing module 2, replacing the transmission of full data or fixed key frames, and reducing invalid data transmission through local preprocessing by edge computing.

[0072] In one embodiment, the start and end point filtering submodule is customized as a differentiated data encapsulation, specifically as follows:

[0073] The encapsulated content of the start-end data pairs for low, medium, and high dynamic driving scenarios includes the full start-end data, the full end-end data, the dynamic level identifier, the timestamp range, and the sensor identifier.

[0074] The encapsulation of the start-end data pair in high-dynamic driving scenarios also includes obstacle contour coordinates with a preset area for the feature intermediate points and key GPS steering positions. The obstacle contour coordinates are extracted using a point cloud clustering algorithm, and obstacle contours with a preset area after clustering are selected. The preset area can be set to be no less than 0.5㎡. After encapsulation, lightweight Huffman coding is used to compress the data header information, further reducing the transmission volume. The encapsulation process provided in this embodiment is completed in the vehicle-side edge computing module 2, avoiding remote processing delays.

[0075] The vehicle-side data transmission module 3 adds a checksum, such as a CRC32 checksum, to the encapsulated data pairs from the vehicle-side edge computing module 2 before transmitting them to the cloud-side data transmission module 4. In this embodiment, the data transmission module 3 can use the 5G / V2X communication protocol to transmit the encapsulated start-end data pairs from the vehicle-side edge computing module 2 to terminals such as the cloud or roadside units.

[0076] The cloud data transmission module 4 is used to receive encapsulated data pairs and realize high-speed data forwarding between the vehicle and the cloud. It uses a check code to initially determine whether the data is lost or damaged. The check code is transmitted along with the encapsulated data, and the cloud can quickly complete the initial verification.

[0077] The adaptive interpolation restoration module 5 is used to perform preliminary verification on the encapsulated data pairs, and then selects a matching interpolation algorithm based on the scene level identifier to restore continuous spatiotemporal data, achieving accurate matching between the algorithm and the driving scene. Specific implementation logic:

[0078] (1) In low-dynamic driving scenarios, a linear interpolation algorithm is used: based on the numerical change trends of the starting point and the ending point, the intermediate frame data is uniformly interpolated and supplemented. For example, in GPS trajectory data, the starting point coordinates End point coordinates The coordinates of the t-th frame in the middle are , The symbol ' / ' represents multiplication, and ' / ' represents division; the same applies to 'y'. This algorithm has low computational complexity, is suitable for the smooth changes in low-dynamic driving scenarios, and can quickly complete the reconstruction.

[0079] (2) In the dynamic driving scenario, a quadratic polynomial interpolation algorithm is adopted: a quadratic polynomial curve is fitted by the data of the starting point and the ending point. The fitting formula can be: Where a, b, and c are fitting coefficients, solved using the least squares method. This method restores the gradual trend of the data, such as the curvature change of a vehicle's trajectory as it slowly turns, offering a significant improvement in accuracy compared to linear interpolation.

[0080] (3) In high-dynamic driving scenarios, an adaptive spline interpolation algorithm is adopted: combining the three-dimensional data of the starting point, feature intermediate points and the ending point, a piecewise spline curve is fitted. In this embodiment, the first derivative of each curve is continuously constrained by three adjacent points to accurately restore the dramatic change trend of the data, such as the trajectory deviation of the vehicle when it turns sharply and the rapid movement path of the obstacle. At the same time, contour constraint interpolation is added to the lidar point cloud data. Based on the prior knowledge of the actual size of the obstacle, the coordinate range of the interpolation point is constrained to ensure that the error between the restored obstacle contour and the actual size is less than the preset error, such as 5%.

[0081] The adaptive interpolation restoration module 5 outputs the fully restored continuous spatiotemporal data, which is the restored data mentioned elsewhere.

[0082] Error verification module 6 retrieves the raw spatiotemporal data cached by vehicle-side edge computing module 2 and compares it with the restored data from adaptive interpolation module 5. If the error exceeds a preset threshold, it sends a retransmission request to vehicle-side edge computing module 2. The vehicle then retransmits the start and end point data pairs for that time period, reducing the interval between the start and end points. If the error does not exceed the threshold, the restored data is output to the autonomous driving decision-making system. Error verification module 6 performs accuracy verification on the complete interpolated data to ensure data validity.

[0083] In one embodiment, the error verification module 6 is used to implement specific logic including:

[0084] Prepare verification data: Retrieve the original spatiotemporal data within the most recent preset time period (e.g., 10s) from the local cache of the vehicle-side edge computing module 2. After verification, automatically clean up to save storage resources and replace the archived original spatiotemporal data for comparison and verification with the restored data.

[0085] Calculation error: The cloud-based adaptive interpolation restoration module 5 extracts any two values ​​that are not the starting point and the ending point from the restored data. Through, for example, 5G / V2X communication, it requests the corresponding original spatiotemporal data (with the timestamp) from the vehicle-side edge computing module 2 and calculates the absolute error between the two. For example, the position error is calculated using the Euclidean distance formula, and the attitude angle error is calculated using the absolute value of the angle difference.

[0086] Accuracy judgment and retransmission mechanism: A preset error threshold is used. If the error of the restored data is less than the threshold, it is judged as qualified and output to the autonomous driving decision-making system. If the error is greater than the threshold, it is judged as unqualified, and a retransmission request is sent to the vehicle-side edge computing module 2. This request contains the timestamp range of the unqualified data, triggering the vehicle to retransmit the start and end point data pairs for that time period. During retransmission, the interval between the start and end points is reduced, for example, by 50%, to ensure restoration accuracy. In this embodiment, the position error threshold can be set to 0.3m, and the attitude angle error is less than 0.5°.

[0087] This invention belongs to the field of autonomous driving data transmission and processing technology, specifically applied to continuous spatiotemporal data interaction scenarios between autonomous vehicles and the cloud, as well as vehicle-road cooperative roadside units. Based on an edge computing architecture, this invention can adapt to the transmission optimization of various types of highly dynamic spatiotemporal data, such as LiDAR point clouds, GPS trajectories, and inertial measurement unit (IMU) time-series data. It is particularly suitable for real-time data interaction needs under highly dynamic conditions such as urban road turning, sudden obstacle avoidance, and driving in rain or snow.

[0088] This invention also provides an adaptive interpolation restoration method for continuous spatiotemporal data transmission in autonomous driving, comprising:

[0089] Step 1: On the vehicle side, collect continuous spatiotemporal data of autonomous vehicles, add a unique timestamp and sensor identifier to each frame of data, extract dynamic feature parameters of several consecutive frames of data, compare the dynamic feature parameters with preset thresholds to identify the dynamic driving scenario level, adaptively configure the start and end time intervals according to the scenario level and filter data pairs. For high dynamic driving scenarios, supplement the feature intermediate point with the largest data change between the start and end points. After custom encapsulating the start and end point data pairs, compress the data header information, add a check code, and then transmit it to the cloud.

[0090] Step 2: In the cloud, after completing the initial verification of the encapsulated data pairs, select the matching interpolation algorithm according to the scene level identifier to restore the continuous spatiotemporal data.

[0091] Step 3: Call the original spatiotemporal data within the most recent preset time period cached in Step 1 and compare it with the restored data in Step 2. If the error exceeds the preset threshold, a retransmission request is sent to the vehicle. The vehicle retransmits the start and end data pairs of the time period and reduces the interval between the start and end points. If the error does not exceed the threshold, the restored data is output to the autonomous driving decision system.

[0092] In one embodiment, step 1, the method for identifying the dynamic driving scenario level specifically includes:

[0093] Dynamic feature parameters are extracted from several consecutive frames of spatiotemporal data. These dynamic feature parameters include position change, attitude change rate, and LiDAR point cloud dispersion. The feature parameters are compared with preset thresholds to classify three driving scenarios: low dynamic, medium dynamic, and high dynamic.

[0094] The frame interval between the start and end points is configured according to the scene level: for low-dynamic driving scenes and medium-dynamic driving scenes, the first frame of the continuous frame is extracted as the start point and the n1th frame as the end point. For high-dynamic driving scenes, the interval between the start and end points is preset to n2 frames, and a feature intermediate point is added between the start and end points. The feature intermediate point is the frame with the largest data change determined by real-time calculation of the data change of adjacent frames. The start and end point data pairs are customized and encapsulated before the data header information is compressed.

[0095] In one embodiment, step 1 includes the following customized encapsulation:

[0096] The encapsulated content of the start-end data pairs for low, medium, and high dynamic driving scenarios includes the full start-end data, the full end-end data, the dynamic level identifier, the timestamp range, and the sensor identifier.

[0097] The encapsulated content of the start and end point data pairs in high-dynamic driving scenarios also includes obstacle contour coordinates of the preset area of ​​the feature intermediate point and the GPS key steering position.

[0098] In one embodiment, step 2, the method for selecting a matching interpolation algorithm to restore continuous spatiotemporal data based on the scene level identifier, includes:

[0099] In low-dynamic driving scenarios, a linear interpolation algorithm is used to uniformly interpolate and complete the intermediate frame data based on the numerical change trends of the start and end points.

[0100] In the dynamic driving scenario, a quadratic polynomial interpolation algorithm is used. The least squares method is used to solve the fitting coefficients, and the fitting quadratic polynomial curve is used to restore the gradual trend of data.

[0101] In high-dynamic driving scenarios, a combination algorithm of adaptive spline interpolation and contour constraint interpolation is adopted. The adaptive spline interpolation algorithm combines the three-dimensional data of the starting point, feature intermediate point and the ending point to fit a piecewise spline curve, and each curve segment is continuously constrained by the first derivative of three adjacent points. The contour constraint interpolation is based on prior knowledge of the actual size of the obstacle to constrain the coordinate range of the interpolation points, ensuring that the error between the restored obstacle contour and the actual size is less than the preset error.

[0102] In one embodiment, step 3 specifically includes:

[0103] Step 31: Retrieve the original spatiotemporal data within the most recent preset time period cached in Step 1. After verification, automatically clean it up and replace it with archived original spatiotemporal data for comparison and verification with the restored data.

[0104] Step 32: Extract any two non-starting and non-ending values ​​from the restored data in Step 2 in the cloud, request the original spatiotemporal data with the corresponding timestamp from the vehicle, and calculate the absolute errors of position error and attitude angle error.

[0105] Step 33: Preset an error threshold. If the error of the restored data is less than the threshold, it is judged as qualified and output to the autonomous driving decision system; if the error is greater than the threshold, it is judged as unqualified and a retransmission request is sent to the vehicle. The request contains the timestamp range of the unqualified data, triggering the vehicle to retransmit the start and end data pairs of the time period. During the retransmission, the interval between the start and end points is reduced to ensure restoration accuracy.

[0106] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Those skilled in the art should understand that modifications can be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; these 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.

Claims

1. An adaptive interpolation and restoration device for continuous spatiotemporal data transmission in autonomous driving, characterized in that, It includes a data acquisition module (1), a vehicle-side edge computing module (2), a vehicle-side data transmission module (3), a cloud-side data transmission module (4), an adaptive interpolation and restoration module (5), and an error verification module (6). The data acquisition module (1) is used to collect multi-type continuous spatiotemporal data of autonomous vehicles, and after adding a millisecond-level unique timestamp and sensor identifier to each frame of data, it is transmitted to the vehicle edge computing module (2) as the original spatiotemporal data. The vehicle-side edge computing module (2) is deployed on the vehicle side to extract dynamic feature parameters of several consecutive frames of original spatiotemporal data, compare the dynamic feature parameters with the preset threshold, identify the dynamic driving scene level, adaptively configure the time interval between the start and end points according to the scene level and filter data pairs. For high dynamic driving scenes, the feature intermediate point with the largest data change is added between the start and end points. The data header information is compressed after the start and end point data pairs are customized and encapsulated. The vehicle-side data transmission module (3) is used to add a check code to the encapsulated data pair of the vehicle-side edge computing module (2) and then transmit it to the cloud data transmission module (4). The cloud data transmission module (4) is used to receive encapsulated data pairs and realize high-speed data forwarding between the vehicle and the cloud; The adaptive interpolation restoration module (5) is used to restore continuous spatiotemporal data by selecting a matching interpolation algorithm according to the scene level identifier after completing the initial verification of the encapsulated data pair; The error verification module (6) is used to call the original spatiotemporal data within the most recent preset time period cached by the vehicle edge computing module (2) and compare it with the restored data of the adaptive interpolation restoration module (5). If the error exceeds the preset threshold, it will send a supplementary transmission request to the vehicle edge computing module (2). The vehicle will retransmit the start and end data pairs of the time period and reduce the interval between the start and end. If the error does not exceed the threshold, it will output the restored data to the autonomous driving decision system.

2. The adaptive interpolation and restoration device for continuous spatiotemporal data transmission in autonomous driving as described in claim 1, characterized in that, The vehicle-side edge computing module (2) includes: The dynamic driving scene recognition submodule (21) is used to extract dynamic feature parameters of several consecutive frames of original spatiotemporal data. The dynamic feature parameters include position change, attitude change rate and laser radar point cloud dispersion. The dynamic feature parameters are compared with preset thresholds to divide the driving scene into three types: low dynamic, medium dynamic and high dynamic. The start and end point filtering submodule (22) is used to configure the frame interval between the start and end points according to the scene level: for low dynamic driving scene and medium dynamic driving scene, the first frame of the continuous frame is extracted as the start point and the n1th frame is the end point. For high dynamic driving scene, the interval between the start and end points is preset to n2 frames, and a feature intermediate point is added between the start and end points. The feature intermediate point is the frame with the largest data change determined by real-time calculation of the data change of adjacent frames. The start and end point data pairs are customized and encapsulated before the data header information is compressed.

3. The adaptive interpolation and restoration device for continuous spatiotemporal data transmission in autonomous driving as described in claim 2, characterized in that, The customized encapsulation of the start and end point filtering submodule (22) is a differentiated data encapsulation, the specific content of which is as follows: The encapsulated content of the start-end data pairs for low, medium, and high dynamic driving scenarios includes the full start-end data, the full end-end data, the dynamic level identifier, the timestamp range, and the sensor identifier. The encapsulated content of the start and end point data pairs in high-dynamic driving scenarios also includes obstacle contour coordinates of the preset area of ​​the feature intermediate point and the GPS key steering position.

4. The adaptive interpolation and restoration device for continuous spatiotemporal data transmission in autonomous driving as described in claim 2, characterized in that, In the adaptive interpolation module (5), the interpolation algorithm that matches the low, medium, and high dynamic driving scenario level identifiers one by one is as follows: In low-dynamic driving scenarios, a linear interpolation algorithm is used to uniformly interpolate and complete the intermediate frame data based on the numerical change trends of the start and end points. In the dynamic driving scenario, a quadratic polynomial interpolation algorithm is used. The least squares method is used to solve the fitting coefficients, and the fitting quadratic polynomial curve is used to restore the gradual trend of data. In high-dynamic driving scenarios, a combination algorithm of adaptive spline interpolation and contour constraint interpolation is adopted. The adaptive spline interpolation algorithm combines the three-dimensional data of the starting point, feature intermediate point and the ending point to fit a piecewise spline curve, and each curve segment is continuously constrained by the first derivative of three adjacent points. The contour constraint interpolation is based on prior knowledge of the actual size of the obstacle to constrain the coordinate range of the interpolation points, ensuring that the error between the restored obstacle contour and the actual size is less than the preset error.

5. The adaptive interpolation and restoration device for continuous spatiotemporal data transmission in autonomous driving as described in any one of claims 1-4, characterized in that, The error verification module (6) includes: The verification data preparation unit (61) is used to call the vehicle-side edge computing module (2) to locally cache the original spatiotemporal data within the most recent preset time period; Error calculation unit (62) is used to extract any two non-starting point and ending point values ​​from the restored data of the adaptive interpolation restoration module (5) in the cloud, request the original spatiotemporal data of the corresponding timestamp from the vehicle edge computing module (2), and calculate the absolute error of position error and attitude angle error; The accuracy judgment and retransmission mechanism unit (63) is used to preset the error threshold. If the error of the restored data is less than the threshold, it is judged as qualified and output to the autonomous driving decision system. If the error is greater than the threshold, it is judged as unqualified and a retransmission request is fed back to the vehicle edge computing module (2). The request contains the timestamp range of the unqualified data, which triggers the vehicle to retransmit the start and end data pairs of the time period. When retransmitting, the interval between the start and end points is reduced to ensure the restoration accuracy.

6. An adaptive interpolation restoration method for continuous spatiotemporal data transmission in autonomous driving, characterized in that, include: Step 1: On the vehicle side, collect continuous spatiotemporal data of autonomous vehicles, add a millisecond-level unique timestamp and sensor identifier to each frame of data as raw spatiotemporal data, extract dynamic feature parameters from several consecutive frames of raw spatiotemporal data, compare the dynamic feature parameters with preset thresholds to identify the dynamic driving scenario level, adaptively configure the start and end time intervals according to the scenario level and filter data pairs. For high dynamic driving scenarios, supplement the feature intermediate point with the largest data change between the start and end points, and after custom encapsulating the start and end point data pairs, compress the data header information, add a check code, and then transmit it to the cloud. Step 2: In the cloud, after completing the initial verification of the encapsulated data pairs, select the matching interpolation algorithm according to the scene level identifier to restore the continuous spatiotemporal data; Step 3: Call the original spatiotemporal data within the most recent preset time period cached in Step 1 and compare it with the restored data in Step 2. If the error exceeds the preset threshold, a retransmission request is sent to the vehicle. The vehicle retransmits the start and end data pairs of the time period and reduces the interval between the start and end points. If the error does not exceed the threshold, the restored data is output to the autonomous driving decision system.

7. The adaptive interpolation restoration method for continuous spatiotemporal data transmission in autonomous driving as described in claim 6, characterized in that, In step 1, the method for identifying dynamic driving scenario levels specifically includes: Dynamic feature parameters are extracted from several consecutive frames of raw spatiotemporal data. The dynamic feature parameters include position change, attitude change rate and LiDAR point cloud dispersion. The dynamic feature parameters are compared with preset thresholds to classify three driving scenarios: low dynamic, medium dynamic and high dynamic. The frame interval between the start and end points is configured according to the scene level: for low-dynamic driving scenes and medium-dynamic driving scenes, the first frame of the continuous frame is extracted as the start point and the n1th frame as the end point. For high-dynamic driving scenes, the interval between the start and end points is preset to n2 frames, and a feature intermediate point is added between the start and end points. The feature intermediate point is the frame with the largest data change determined by real-time calculation of the data change of adjacent frames. The start and end point data pairs are customized and encapsulated before the data header information is compressed.

8. The adaptive interpolation restoration method for continuous spatiotemporal data transmission in autonomous driving as described in claim 7, characterized in that, In step 1, the customized packaging includes: The encapsulated content of the start-end data pairs for low, medium, and high dynamic driving scenarios includes the full start-end data, the full end-end data, the dynamic level identifier, the timestamp range, and the sensor identifier. The encapsulated content of the start and end point data pairs in high-dynamic driving scenarios also includes obstacle contour coordinates of the preset area of ​​the feature intermediate point and the GPS key steering position.

9. The adaptive interpolation restoration method for continuous spatiotemporal data transmission in autonomous driving as described in claim 7, characterized in that, In step 2, the method for selecting a matching interpolation algorithm to restore continuous spatiotemporal data based on the scene level identifier includes: In low-dynamic driving scenarios, a linear interpolation algorithm is used to uniformly interpolate and complete the intermediate frame data based on the numerical change trends of the start and end points. In the dynamic driving scenario, a quadratic polynomial interpolation algorithm is used. The least squares method is used to solve the fitting coefficients, and the fitting quadratic polynomial curve is used to restore the gradual trend of data. In high-dynamic driving scenarios, a combination algorithm of adaptive spline interpolation and contour constraint interpolation is adopted. The adaptive spline interpolation algorithm combines the three-dimensional data of the starting point, feature intermediate point and the ending point to fit a piecewise spline curve, and each curve segment is continuously constrained by the first derivative of three adjacent points. The contour constraint interpolation is based on prior knowledge of the actual size of the obstacle to constrain the coordinate range of the interpolation points, ensuring that the error between the restored obstacle contour and the actual size is less than the preset error.

10. The adaptive interpolation restoration method for continuous spatiotemporal data transmission in autonomous driving as described in any one of claims 6-9, characterized in that, Step 3 specifically includes: Step 31: Extract any two non-starting and non-ending values ​​from the restored data in Step 2 in the cloud, request the original spatiotemporal data with the corresponding timestamps from the vehicle, and calculate the absolute errors of position error and attitude angle error. Step 32: Preset an error threshold. If the error of the restored data is less than the threshold, it is judged as qualified and output to the autonomous driving decision system. If the error is greater than the threshold, it is judged as unqualified and a retransmission request is sent to the vehicle. The request contains the timestamp range of the unqualified data, triggering the vehicle to retransmit the start and end data pairs of the time period. During the retransmission, the interval between the start and end points is reduced to ensure restoration accuracy.

Images