Comprehensive index calculation method for automatic driving scene library

Abstract

The invention discloses a comprehensive index calculation method for an automatic driving scene library, and belongs to the technical field of automatic driving. In an automatic driving scene, accurate calculation of a comprehensive index to realize scene classification is of great significance, but an existing method has the problems of inaccurate data processing, inaccurate index calculation and the like. According to the method, dynamic lane width data, local coordinate system parameters and longitudinal and transverse collision time are calculated by obtaining three-dimensional positioning data, wheel speed pulse data and image data through the steps of fusion, complementation, processing and the like, then a dynamic collision risk index is generated, and finally a scene classification label is output. The method is mainly used for construction and analysis of an automatic driving scene library and provides powerful support for automatic driving decision making.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving technology. More specifically, the present invention relates to a comprehensive index calculation method for an autonomous driving scenario library. Background Art

[0002] In the development of autonomous driving technology, accurately constructing a scenario library and calculating comprehensive indicators are crucial to achieving safe and efficient autonomous driving.

[0003] At present, data collection of autonomous vehicles mainly relies on a variety of sensors, such as on-board GNSS modules, wheel speed sensors, and front-view cameras. However, data from different sensors are subject to errors and instability. For example, in certain environments (such as urban canyons and tunnels with tall buildings), the on-board GNSS module may have inaccurate positioning, resulting in deviations in the three-dimensional positioning data; the wheel speed pulse data obtained by the wheel speed sensor may be affected by factors such as tire wear and road conditions, resulting in errors; the image data obtained by the front-view camera may be affected by conditions such as lighting and weather, resulting in reduced image quality and affecting subsequent analysis and processing.

[0004] In terms of data fusion, existing methods often fail to fully utilize the advantages of different sensor data, resulting in missing or inaccurate fused trajectory data. When there are several consecutive frames of empty value areas in the fused trajectory data, how to accurately mark and complete these missing data is an urgent problem to be solved. Traditional data completion methods may not take into account the temporal characteristics of the data, resulting in a large deviation between the completed trajectory data and the actual situation.

[0005] For lane width calculation, most existing methods use fixed values ​​or simple average values, which cannot be adjusted dynamically according to actual driving conditions. In actual driving, lane width may change due to factors such as road construction and vehicle merging. If the dynamic lane width cannot be calculated in a timely and accurate manner, it may cause the autonomous vehicle to cross the line, deviate from the lane, and other dangerous situations during driving.

[0006] In terms of collision risk assessment, existing index calculation methods often only consider a single factor, such as longitudinal distance or lateral distance, while ignoring other important factors such as road curvature. Changes in road curvature will affect the vehicle's driving trajectory and speed, which will have a significant impact on the collision risk. Therefore, how to comprehensively consider multiple factors and accurately calculate dynamic collision risk indicators is the key to improving the safety of autonomous driving.

[0007] In addition, in terms of scene classification, existing methods lack comprehensive consideration of multiple indicators, resulting in inaccurate scene classification. The characteristics of car-following scenes, lane change scenes, and intersection scenes are complex and diverse, and it is difficult to accurately distinguish these scenes by relying on a single indicator. Therefore, a method that can comprehensively consider multiple factors is needed to achieve accurate classification of autonomous driving scenes. Summary of the invention

[0008] An object of the present invention is to solve at least the above problems and to provide at least the advantages which will be described hereinafter.

[0009] The present invention aims to improve the accuracy of comprehensive index calculation and the reliability of scene classification through a series of steps, including data acquisition, fusion, completion, processing, index calculation and scene classification.

[0010] In order to achieve these purposes and other advantages according to the present invention, a comprehensive index calculation method for an autonomous driving scenario library is provided, characterized in that it includes the following steps: S1, obtain three-dimensional positioning data through the vehicle-mounted GNSS module, obtain wheel speed pulse data through the wheel speed sensor, and obtain image data through the front-view camera; S2, inputting the three-dimensional positioning data and wheel speed pulse data of step S1 into the extended Kalman filter to generate fused trajectory data; S3, detecting a plurality of consecutive frames of null value regions in the fused trajectory data of step S2, and generating missing trajectory mark data; S4, inputting the fused trajectory data of step S2 and the missing trajectory mark data of step S3 into the cascade structure of the temporal convolutional network and the micro-gated recurrent unit network, and outputting the completed trajectory data; S5, performing ROI region cropping and grayscale processing on the image data of the front-view camera in step S1 to generate pre-processed image data; S6: The pre-processed image data of step S5 is integrated with the historical trajectory database and map data to calculate the dynamic lane width data. The output formula is W real =0.6W cam +0.4W hist , where W real Represents the true value of lane width, W cam Indicates the value measured by the camera, W hist Indicates the lane width history value; S7, based on the completed trajectory data of step S4 and the dynamic lane width data of step S6, generate local coordinate system parameters through Frenet frame calculation; S8, projecting the completed trajectory data of step S4 to the local coordinate system of step S7, and decomposing it into the longitudinal velocity component v x With the lateral velocity component vy ; S9, according to step S8 v x 、v y The relative distances Δx and Δy from the target object are used to calculate the longitudinal collision time TTC. x =Δx / v x The time to lateral collision TTC y =Δy / v y ; S10, TTC of fusion step S9 x 、TTC y And the road curvature κ extracted from the map, through the formula Generate a dynamic collision risk index, where α(κ)+β(κ)=1 and α(κ)=0.95-0.5κ, where α(κ) and β(κ) represent weight coefficients related to the road curvature κ; S11. The variance of the DRC index and the lateral velocity component of step S10 is judged based on a preset threshold value, and classification labels of the following scene, the lane change scene and the intersection scene are output.

[0011] Preferably, the method for constructing the temporal convolutional network includes: S401, configure a three-level cascaded dilated convolution module, set the dilated ratios of the first-level dilated convolution module, the second-level dilated convolution module, and the third-level dilated convolution module to 1, 2, and 4, respectively, preset 64 convolution kernels of size 3×3 for each dilated convolution module, set the convolution sliding step to 1, and use a symmetric filling method in the time dimension to maintain the size of the feature map; S402, inputting the feature map output by the front neural network into the first-level dilated convolution module to generate a primary feature map, inputting the primary feature map into the second-level dilated convolution module to generate an intermediate feature map, and inputting the intermediate feature map into the third-level dilated convolution module to generate a high-level feature map; S403, grouping the 64 convolution kernels of the first-level dilated convolution module, the second-level dilated convolution module, and the third-level dilated convolution module, and splitting the 64 convolution kernels of each level of dilated convolution module into 4 groups, each group containing 16 convolution kernels, and sharing parameters within the group; S404, performing structured pruning on the grouped 64 convolution kernels, retaining the first 50% of the highly activated convolution kernels in each level of the module through L1 norm analysis, and setting the remaining 50% of the convolution kernels to zero; S405, performing channel compression on the high-level feature map output by the pruned third-level hole convolution module through 1×1 convolution to generate a compressed high-level feature map; S406, performing feature fusion on the compressed high-level feature map and the input feature map of the micro-gated recurrent unit network through a tensor addition operation to generate a residual connection feature map; S407, applying a linear unit activation function with leakage correction to the output feature map of each dilated convolution module, wherein the leakage factor of the linear unit with leakage correction is set to 0.01; S408, outputting the residual connection feature map to the micro-gated recurrent unit network, and outputting the completed trajectory data.

[0012] Preferably, the configuration method of the micro-gated recurrent unit network includes: S409, constructing a micro gated recurrent unit network, wherein the hidden layer adopts a 16-node topology structure, and the gated unit weight initialization follows the Xavier normal distribution; S410, performing 8-bit integer quantization on the 32-bit floating-point weight Wfloat32 of the network, converting it into 8-bit integer weight Wint8, and recalculating quantization parameters μ and σ every 1000 training iterations, where μ and σ are the mean and standard deviation of Wfloat32 respectively; S411, dequantize Wint8 through inverse transformation to obtain the dequantized floating point weight W dequant , ; S412. Use the straight-through estimator (STE) to approximate the gradient in the back-propagation stage, keeping the gradient of the rounding operation to 1.

[0013] Preferably, the method for calculating α(κ) and β(κ) based on the road curvature includes: S1001, obtaining road curvature κ; S1002, determine the range of κ: like , according to the formula Calculate α(κ), and then according to Calculate β(κ); like , α(κ)=0.4±0.02 and β(κ)=0.6 are forced to be set, and the correction coefficient 0.02 is dynamically adjusted through the Kalman filter residual variance.

[0014] Preferably, the historical lane width in the dynamic lane width data in step S6 is calculated using an exponentially decaying moving average algorithm, with the decay factor set to 0.85; the initial historical lane width value is taken from the nominal lane width value of the corresponding road section in the map; when the camera confidence for several consecutive frames is lower than a preset threshold, the historical lane width is forcibly reset to the map reference value, and the lane line re-detection process based on color space segmentation is activated.

[0015] Preferably, the cubic spline interpolation control point interval of the road centerline in the Frenet frame calculation in step S7 is dynamically adjusted according to the real-time curvature: when the real-time curvature is greater than 0.1 per meter, the control point interval is shortened to 2 meters; when the real-time curvature is less than or equal to 0.1 per meter, the control point interval is extended to 10 meters; the curvature calculation adopts the central difference method to obtain the second-order derivative of the road centerline parameter equation.

[0016] Preferably, the parameter configuration of the extended Kalman filter satisfies: The state vector is defined as a two-dimensional vector X, which contains the longitudinal position s and the longitudinal velocity v; The process noise covariance matrix Q is configured as a diagonal matrix diag(0.1,0.05), reflecting the process noise variance of the longitudinal position and velocity; The observation noise covariance matrix R is configured as a diagonal matrix diag(0.2,0.1), which characterizes the measurement noise intensity of the longitudinal position and velocity; The state transfer matrix F adopts the uniform kinematics model: F = [[1, Δt], [0, 1]], where Δt = 0.1 seconds is the discretization time step; The observation matrix H is configured as a 2×2 unit matrix I to achieve full-dimensional observation of the state vector.

[0017] Preferably, the ROI area and processing flow of the image preprocessing module include: S501, in the pixel coordinate system of the front-view camera, a rectangular area from the upper left corner coordinate to the lower right corner coordinate of the road detection range of 100 meters in front of the vehicle is selected as the ROI area; S502, obtaining an input image, converting it from RGB format to YUV420 format, wherein the chrominance component adopts a 4:2:0 subsampling mode to generate a YUV420 format image with a 30% reduction in chrominance data volume; S503, performing a CLAHE algorithm of grid division on the Y component of the YUV420 format image in S502, setting an 8×8 pixel grid unit, a contrast limit threshold of 2.0, and using bilinear interpolation to eliminate the grid boundary effect, thereby generating a processed Y component; S504: Perform bilateral filtering on the U / V components of the YUV420 format image in S502, and set the spatial domain standard deviation σ s = 1.5 pixels, color domain standard deviation σ c =15 gray level, filter window diameter d=5 pixels, generating filtered U / V components; S505, combining the processed Y component generated in S503 with the filtered U / V component generated in S504, and outputting the pre-processed image data in YUV420 format; Among them, the Y component represents the brightness component, the U component represents the difference between the blue chromaticity information and the brightness information in the image, and the V component represents the difference between the red chromaticity information and the brightness information in the image.

[0018] The present invention has at least the following beneficial effects: First, by completing the missing trajectory through the adversarial network model, the autonomous driving system can fully grasp the vehicle's motion status and avoid decision-making errors. Multi-dimensional collision risk assessment takes into account factors such as road curvature, and can provide timely warning and braking under complex road conditions. For example, on high-curvature sections, the lateral risk recognition rate is greatly improved, and the emergency braking trigger delay is significantly shortened, effectively reducing collision accidents and ensuring personnel safety.

[0019] Second, based on the dynamic time warping algorithm, driving modes are identified and scene features are accurately extracted and classified, so that the autonomous driving system can quickly identify scenes such as lane changing and following, understand driving intentions, optimize driving paths and speed planning, improve driving smoothness and comfort, and reduce unnecessary acceleration, deceleration and lane changing operations.

[0020] Third, the trajectory data processing mechanism has been improved, speed data noise has been reduced, and data accuracy and reliability have been improved. High-quality data provides a solid foundation for algorithm training and model optimization of the autonomous driving system, enhances the system's adaptability and stability to different scenarios, and reduces the interference of abnormal data on the system.

[0021] Fourth, the scenario indicator credibility assessment model can effectively identify low-confidence data and abnormal conditions, and prevent the system from using unreliable data for decision-making. At the same time, accurate risk scenario marking and confidence interval output allow the system to prepare for response in advance, reduce the probability of misjudgment and wrong decision-making, and improve the stability and reliability of system operation.

[0022] Fifth, it solves key technical problems in the industry, provides a more scientific indicator calculation method for the construction of an autonomous driving scenario library, promotes the research and innovation of autonomous driving technology, accelerates the transformation of autonomous driving technology from theoretical research to practical application, and promotes the development of the entire autonomous driving industry.

[0023] Other advantages, objectives and features of the present invention will be embodied in part through the following description, and in part will be understood by those skilled in the art through study and practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 This is a flow chart of the method for calculating the comprehensive index of the autonomous driving scenario library. DETAILED DESCRIPTION

[0025] The present invention is further described in detail below in conjunction with the accompanying drawings so that those skilled in the art can implement the invention with reference to the description.

[0026] It should be noted that the experimental methods described in the following embodiments are conventional methods unless otherwise specified, and the reagents and materials are commercially available unless otherwise specified; in the description of the present invention, the orientation or positional relationship indicated by the terms is based on the orientation or positional relationship shown in the accompanying drawings, which is only for the convenience of describing the present invention and simplifying the description, and does not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention.

[0027] like Figure 1 As shown, the present invention provides a comprehensive index calculation method for an autonomous driving scenario library, comprising the following steps: S1, obtain three-dimensional positioning data through the vehicle-mounted GNSS module, obtain wheel speed pulse data through the wheel speed sensor, and obtain image data through the front-view camera; S2, inputting the three-dimensional positioning data and wheel speed pulse data of step S1 into the extended Kalman filter to generate fused trajectory data; S3, detecting a plurality of consecutive frames of null value regions in the fused trajectory data of step S2, and generating missing trajectory mark data; S4, inputting the fused trajectory data of step S2 and the missing trajectory mark data of step S3 into the cascade structure of the temporal convolutional network and the micro-gated recurrent unit network, and outputting the completed trajectory data; S5, performing ROI region cropping and grayscale processing on the image data of the front-view camera in step S1 to generate pre-processed image data; S6: The pre-processed image data of step S5 is integrated with the historical trajectory database and map data to calculate the dynamic lane width data. The output formula is W real =0.6W cam +0.4W hist , where W real Represents the true value of lane width, W cam Indicates the value measured by the camera, W hist Indicates the lane width history value; S7, based on the completed trajectory data of step S4 and the dynamic lane width data of step S6, generate local coordinate system parameters through Frenet frame calculation; S8, projecting the completed trajectory data of step S4 to the local coordinate system of step S7, and decomposing it into the longitudinal velocity component v x With the lateral velocity component v y ; S9, according to step S8 v x 、v y The relative distances Δx and Δy from the target object are used to calculate the longitudinal collision time TTC.x =Δx / v x The time to lateral collision TTC y =Δy / v y ; S10, TTC of fusion step S9 x 、TTC y And the road curvature κ extracted from the map, through the formula Generate a dynamic collision risk index, where α(κ)+β(κ)=1 and α(κ)=0.95-0.5κ, where α(κ) and β(κ) represent weight coefficients related to the road curvature κ; S11. The variance of the DRC index and the lateral velocity component of step S10 is judged based on a preset threshold value, and classification labels of the following scene, the lane change scene and the intersection scene are output.

[0028] In the above technical solution, by integrating multiple sensor data, performing data fusion, completion and multi-step calculation, the dynamic collision risk index is finally obtained and the scene classification is realized. The comprehensive use of multi-sensor data improves the comprehensiveness and accuracy of the data, and can more realistically reflect the environment in which the autonomous driving vehicle is located. By extending the Kalman filter to fuse positioning and wheel speed data, the influence of single sensor errors is effectively reduced. The completed trajectory data provides a complete data basis for subsequent analysis. The calculation of dynamic lane width takes into account real-time images and historical data, which is more in line with actual road conditions. The dynamic collision risk index generated by comprehensively considering the longitudinal and lateral collision time and road curvature can more accurately assess the collision risk. The scene classification finally achieved helps the autonomous driving system make more reasonable decisions according to different scenarios and improve the safety and reliability of autonomous driving.

[0029] Specifically, in actual autonomous driving scenarios, the vehicle first needs to use different on-board sensors to obtain relevant data. The on-board GNSS module obtains the vehicle's three-dimensional positioning data by receiving satellite signals, which contains the vehicle's position information in space. The wheel speed sensor obtains wheel speed pulse data by detecting the rotation of the wheels, and then calculates the vehicle's driving speed. The front-view camera captures the road image in front of the vehicle in real time to obtain image data.

[0030] The acquired 3D positioning data and wheel speed pulse data are input into the extended Kalman filter. The extended Kalman filter is a commonly used data fusion algorithm that estimates and updates the vehicle state based on the vehicle's motion model and sensor measurements to generate fused trajectory data. In this process, the extended Kalman filter will continuously adjust the estimation of the vehicle's position and speed based on the new measurements to reduce errors.

[0031] Next, the fused trajectory data is checked to detect the empty value regions in several consecutive frames. These empty value regions may be caused by sensor failure, signal loss, etc. Once the empty value regions are detected, missing trajectory marker data will be generated to mark the location of these missing data.

[0032] The fused trajectory data and the missing trajectory marker data are input into the cascade structure of the temporal convolutional network and the micro-gated recurrent unit network. The temporal convolutional network can extract the temporal features of the data, while the micro-gated recurrent unit network can process the long-term dependencies in the sequence data. Through the collaborative work of these two networks, the missing trajectory data can be completed and the completed trajectory data can be output.

[0033] The image data obtained by the front-view camera needs to be preprocessed. First, the ROI area is cropped to select a specific range of road area in front of the vehicle as the region of interest, which can reduce the interference of irrelevant information. Then grayscale processing is performed to convert the color image into a grayscale image for subsequent analysis and processing. After these processes, preprocessed image data is generated.

[0034] The pre-processed image data is fused with the historical trajectory database and map data. The historical trajectory database records the relevant information of the vehicle in the past driving process, and the map data contains the basic information of the road, such as lane width, road curvature, etc. By fusing these data, the dynamic lane width data can be calculated using the formula W real =0.6W cam +0.4W hist To comprehensively consider the lane width detected by the camera and the historical lane width.

[0035] Based on the completed trajectory data and dynamic lane width data, the local coordinate system parameters are calculated using the Frenet frame. The Frenet frame is a coordinate system based on the centerline of the road. It can convert the trajectory and position of the vehicle into a local coordinate system for subsequent analysis and processing.

[0036] The completed trajectory data is projected into the local coordinate system. According to the definition of the coordinate system, the trajectory data can be decomposed into the longitudinal velocity component v x and the lateral velocity component v y These two velocity components can describe the longitudinal and lateral motion of the vehicle respectively.

[0037] According to the longitudinal velocity component v x , lateral velocity component v y And the relative distances Δx and Δy between the target object and the vehicle, respectively calculate the longitudinal collision time TTC x =Δx / v xThe time to lateral collision TTC y =Δy / v y These two times can be used to evaluate the possibility of collision between the vehicle and the target object in the longitudinal and lateral directions.

[0038] Finally, the longitudinal collision time, the lateral collision time and the road curvature κ extracted from the map are integrated to obtain the following formula: Generate a dynamic collision risk index. α(κ) and β(κ) are coefficients determined according to the road curvature κ, and α(κ)+β(κ)=1. The variance of the DRC index and the lateral velocity component is judged based on a preset threshold, and the classification labels of the following scenario, lane change scenario, and intersection scenario are output according to the judgment result. In this way, the autonomous driving system can make corresponding decisions according to different scenarios.

[0039] In another technical solution, the method for constructing the temporal convolutional network includes: S401, configure a three-level cascaded dilated convolution module, set the dilated ratios of the first-level dilated convolution module, the second-level dilated convolution module, and the third-level dilated convolution module to 1, 2, and 4, respectively, preset 64 convolution kernels of size 3×3 for each dilated convolution module, set the convolution sliding step to 1, and use a symmetric filling method in the time dimension to maintain the size of the feature map; S402, inputting the feature map output by the front neural network into the first-level dilated convolution module to generate a primary feature map, inputting the primary feature map into the second-level dilated convolution module to generate an intermediate feature map, and inputting the intermediate feature map into the third-level dilated convolution module to generate a high-level feature map; S403, grouping the 64 convolution kernels of the first-level dilated convolution module, the second-level dilated convolution module, and the third-level dilated convolution module, and splitting the 64 convolution kernels of each level of dilated convolution module into 4 groups, each group containing 16 convolution kernels, and sharing parameters within the group; S404, performing structured pruning on the grouped 64 convolution kernels, retaining the first 50% of the highly activated convolution kernels in each level of the module through L1 norm analysis, and setting the remaining 50% of the convolution kernels to zero; S405, performing channel compression on the high-level feature map output by the pruned third-level hole convolution module through 1×1 convolution to generate a compressed high-level feature map; S406, performing feature fusion on the compressed high-level feature map and the input feature map of the micro-gated recurrent unit network through a tensor addition operation to generate a residual connection feature map; S407, applying a linear unit activation function with leakage correction to the output feature map of each dilated convolution module, wherein the leakage factor of the linear unit with leakage correction is set to 0.01; S408, outputting the residual connection feature map to the micro-gated recurrent unit network, and outputting the completed trajectory data.

[0040] In the above technical solution, by setting a three-level cascaded hole convolution module with different hole rates, it is possible to extract time series features at different scales, enrich feature expression, improve the feature capture capability of complex time series data, and more comprehensively mine data information than a single hole rate. Grouping and structured pruning of convolution kernels reduces the number of network parameters, reduces the amount of calculation and memory usage, improves the model operation efficiency, and retains key convolution kernels to avoid performance degradation caused by parameter reduction to a certain extent, balancing model complexity and performance. The compressed high-level feature map is fused with the input feature map of the micro-gated recurrent unit network, and residual connections are used to enable the network to better learn the differences and connections between features, improve feature utilization efficiency, help to more accurately output complete trajectory data, and enhance the accuracy of the model's processing of trajectory data. The use of a linear unit activation function with leakage correction and setting a suitable leakage factor can effectively solve the problem of neuron death in the negative semi-axis of the traditional ReLU function, so that the network can better transfer gradients during training, accelerate convergence, and improve the model training effect.

[0041] Specifically, the configuration of the dilated convolution module: When constructing the network, you must first create a three-level cascade of dilated convolution modules. The dilated convolution module's dilated rate setting is key. The first level is set to 1, which is similar to ordinary convolution and can capture local, relatively fine features. The second-level dilated rate is set to 2. At this time, the convolution kernel will skip some positions during the convolution process, the receptive field becomes larger, and features in a slightly larger range can be captured. The third-level dilated rate is set to 4 to further expand the receptive field and obtain more global feature information. Each dilated convolution module is configured with 64 convolution kernels of size 3×3. The size of 3×3 has a good balance between computational complexity and feature extraction capabilities. Set the convolution sliding step size to 1, so that the feature map can be scanned carefully without missing any information. The time dimension is symmetrically filled, that is, the same amount of data is filled before and after the time series to ensure that the size of the feature map in the time dimension remains unchanged after the convolution operation, maintain the time continuity of the data, and facilitate subsequent processing.

[0042] Feature map generation process: The feature map output by the front neural network is used as input to the first-level dilated convolution module. In this module, 64 convolution kernels perform convolution operations on the input feature map according to the set parameters to generate a primary feature map. The primary feature map contains the feature information of the input data after the first-level dilated convolution. The primary feature map is then input to the second-level dilated convolution module, where 64 convolution kernels perform convolution operations with a dilation rate of 2 to generate an intermediate feature map. The intermediate feature map combines the information of the primary feature map and the broader features brought by the second-level dilated convolution. The intermediate feature map is then input to the third-level dilated convolution module, and a convolution operation with a dilation rate of 4 is performed to generate a high-level feature map. The high-level feature map integrates the features of the previous levels and has richer and more global information.

[0043] Convolution kernel grouping: For each level of the atrous convolution module, the 64 convolution kernels are divided into 4 groups, each containing 16 convolution kernels. After grouping, the 16 convolution kernels in the group share parameters. The benefit of sharing parameters is that it reduces the total number of parameters that the network needs to learn and reduces the computational complexity. In the actual calculation process, for the convolution kernels in the same group, the same weight parameters are used when performing convolution operations on feature maps, which greatly improves the computational efficiency and also helps improve the generalization ability of the model.

[0044] Structured pruning operation: After the grouping is completed, the 16 convolution kernels in each group (that is, 64 convolution kernels in each level) are subjected to structured pruning. Through L1 norm analysis, the L1 norm of each convolution kernel is calculated. The L1 norm reflects the sum of the absolute values ​​of the convolution kernel parameters. The convolution kernels are sorted according to the size of the L1 norm, and the first 50% of the highly activated convolution kernels in each level of the module are retained, and the remaining 50% of the convolution kernels are set to zero. This can remove convolution kernels that contribute less to model performance, reduce redundant parameters, further reduce the amount of calculation and memory usage, and avoid excessive negative impact on model performance, ensuring that the model can still maintain good feature extraction capabilities while reducing parameters.

[0045] Channel compression and feature fusion: The high-level feature map output by the pruned third-level atrous convolution module has a large number of channels. In order to further optimize the calculation and feature expression, channel compression is performed through 1×1 convolution. The role of the 1×1 convolution kernel is to adjust the number of channels without changing the spatial size of the feature map. After 1×1 convolution, a compressed high-level feature map is generated, and its number of channels is reduced, which is more convenient for subsequent processing. The compressed high-level feature map is tensor-added with the input feature map of the micro-gated recurrent unit network. Tensor addition is the addition of elements at corresponding positions. In this way, features from different sources are fused to generate a residual connection feature map. Residual connection helps the network better learn the differences between features, prevents problems such as gradient disappearance in deep networks, and improves feature utilization efficiency.

[0046] Application of activation function: Apply the leaky corrected linear unit (Leaky ReLU) activation function to the output feature map of each dilated convolution module. Based on the traditional ReLU function, when the input is negative, the Leaky ReLU no longer sets the output to zero, but multiplies it by a smaller leakage factor (set to 0.01 here). In actual calculations, for the feature map output by the dilated convolution module, each element is calculated according to the rules of the Leaky ReLU function. This can avoid the gradient vanishing problem caused by the complete inactivation of neurons when the input is negative during training, so that the network can more effectively transfer gradients during training, accelerate convergence, and improve model training effects.

[0047] Final output: The generated residual connection feature map is output to the micro-gated recurrent unit network. The micro-gated recurrent unit network further processes the residual connection feature map and finally outputs the completed trajectory data according to the internal structure of the network and the trained parameters. In this process, the micro-gated recurrent unit network uses the feature information extracted and fused by the previous temporal convolutional network, combined with its own processing ability for time series data, to complete the task of completing the trajectory data and output the completed trajectory data that meets the requirements.

[0048] In another technical solution, the configuration method of the micro-gated recurrent unit network includes: S409, construct a micro gated recurrent unit network, whose hidden layer adopts a 16-node topology structure, and the gated unit weight initialization follows the Xavier normal distribution. S410, performing 8-bit integer quantization on the 32-bit floating-point weight Wfloat32 of the network, converting it into 8-bit integer weight Wint8, and recalculating quantization parameters μ and σ every 1000 training iterations, where μ and σ are the mean and standard deviation of Wfloat32 respectively; S411, dequantize Wint8 through inverse transformation to obtain the dequantized floating point weight W dequant , ; S412. Use the straight-through estimator (STE) to approximate the gradient in the back-propagation stage, keeping the gradient of the rounding operation to 1.

[0049] In the above technical solution, a micro-gated recurrent unit network with a 16-node topology is constructed for the hidden layer, and the gated unit weights are initialized according to the Xavier normal distribution, which enables the network to effectively process time series data at a reasonable scale. The 16-node topology strikes a balance between computational complexity and model performance. The Xavier normal distribution weight initialization helps the network converge quickly in the early stages of training, avoids the gradient vanishing or explosion problems caused by improper weight initialization, and improves network training efficiency. 8-bit integer quantization is performed on 32-bit floating-point weights, which significantly reduces the memory space required for network storage and calculation, improves the computing speed, and is particularly suitable for resource-constrained environments. Recalculating the quantization parameters every 1,000 training iterations can adapt to changes in weight distribution during network training, making quantization more accurate, and minimizing the impact on model performance while reducing precision, maintaining model accuracy. Dequantization and reconstruction of the quantized 8-bit integer weights can restore the floating-point representation of the weights when high-precision calculations are required. Dequantization reconstruction ensures the reversibility of the quantization process, allowing the network to flexibly use weights of different precisions at different stages, enjoying the computing and storage advantages brought by quantization, and returning to high-precision computing when necessary to ensure the overall performance of the model. In the back-propagation stage, a straight-through estimator is used to approximate the gradient and keep the rounding operation gradient to 1, which simplifies the gradient calculation process. Since the rounding operation in the quantization process is not differentiable, STE allows the network to smoothly perform back-propagation training in a quantized environment, ensuring the normal progress of training, accelerating model convergence, improving training efficiency, and enabling the network to quickly optimize weights and improve training results.

[0050] Specifically, micro-gated recurrent unit network construction and weight initialization: When constructing a micro-gated recurrent unit network, first determine that the hidden layer adopts a 16-node topology structure. This structure is selected according to the specific application scenario and data characteristics. The 16 nodes can better capture the characteristics and dependencies in the time series data without increasing the amount of calculation too much. In the actual construction process, according to the basic structure of the gated recurrent unit, these 16 nodes are interconnected to form a hidden layer structure capable of processing time series. The weight initialization of the gated unit follows the Xavier normal distribution. The parameters of the Xavier normal distribution are set to a mean of 0, and the standard deviation is calculated based on the number of input and output neurons. For the 16-node structure of the hidden layer, assuming that the number of neurons in the input layer is n, and the number of neurons in the output layer is also 16, the standard deviation is set to , using a random number generator function, generate an initial weight matrix according to the parameters of the Xavier normal distribution, and initialize each weight in the gated unit. At the beginning of training, the initialized weights can make the signal propagate more evenly in the network, which is conducive to faster convergence of the network.

[0051] 8-bit integer quantization and parameter update of 32-bit floating-point weights: Perform 8-bit integer quantization on the network's 32-bit floating-point weights Wfloat32. The basic principle of quantization is to map continuous 32-bit floating-point values ​​to a limited range of 8-bit integer values. First, calculate the dynamic range of the 32-bit floating-point weights, that is, find the maximum and minimum values ​​in the weights. Then, convert each 32-bit floating-point weight value to an 8-bit integer value according to the quantization formula. For example, assuming the weight value range is [min v alue,max v alue], the quantitative formula can be , where round represents rounding operation. The quantization parameters μ and σ are recalculated every 1000 training iterations. At each recalculation, the current 32-bit floating point weight Wfloat32 is traversed to calculate its mean μ and standard deviation σ. The recalculated quantization parameters are used in the next round of quantization process to adapt to the changes in weight distribution during network training and ensure the accuracy of quantization.

[0052] Dequantization reconstruction operation: Dequantize Wint8 through inverse transformation. Dequantization is the inverse process of quantization, and its purpose is to restore 8-bit integer weights to 32-bit floating-point weights. First, dequantization is performed based on the previously calculated and updated quantization parameters μ and σ, as well as the inverse operation of the quantization formula. In actual calculations, for each 8-bit integer weight Wint8, the dequantization formula is used to calculate and obtain the dequantized floating-point weight W. dequant In this way, when high-precision calculations are required, such as in the calculation of certain specific layers or before the final output, the dequantized floating-point weights can be used to ensure the accuracy of the calculation results.

[0053] Gradient approximation in the back-propagation phase: In the back-propagation phase, the straight-through estimator (STE) is used to approximate the gradient, keeping the gradient of the rounding operation at 1. Since the rounding operation in the quantization process is non-differentiable, problems will be encountered when calculating the gradient in back-propagation. STE solves this problem through an approximate method. For the rounding operation, when calculating the gradient, the gradient is directly set to 1. For example, when calculating the gradient about a certain weight, if the weight has undergone a rounding operation during the quantization process, when calculating the gradient in back-propagation, the gradient part related to the rounding operation is regarded as 1, rather than being processed according to the conventional non-differentiable operation. This allows the gradient to pass through the quantization part smoothly, ensuring the normal progress of the back-propagation process, allowing the network to be effectively trained in a quantized environment, continuously optimizing the weights, and improving model performance.

[0054] In another technical solution, a method for calculating α(κ) and β(κ) based on road curvature includes: S1001, obtaining road curvature κ; S1002, determine the range of κ: like , according to the formula Calculate α(κ), and then according to Calculate β(κ); like , α(κ)=0.4±0.02 and β(κ)=0.6 are forced to be set, and the correction coefficient 0.02 is dynamically adjusted through the Kalman filter residual variance.

[0055] In the above technical solution, different calculation methods are used for different κ ranges, which enhances the flexibility and accuracy of the calculation. In the range, α(κ) is calculated by a specific exponential formula, and β(κ) is derived based on this, which can more carefully reflect the changing relationship of related factors within the curvature range. , α(κ) and β(κ) are set compulsorily, and the correction coefficient of α(κ) is dynamically adjusted through the Kalman filter residual variance, which can adapt to the requirements for stability and accuracy of calculation results under complex road conditions, and ensure that the system can have reasonable output under different road curvature conditions. The correction coefficient of 0.02 is dynamically adjusted by the Kalman filter residual variance, and the calculation results can be optimized in real time according to the error feedback during the operation of the system. In actual scenarios with changeable road conditions, the Kalman filter can fuse multi-source data and dynamically adjust the value range of α(κ) according to the residual variance, so that the calculation results are more in line with the actual road conditions and improve the adaptability and reliability of the entire calculation method.

[0056] Specifically, obtaining the road curvature κ: In practical applications, there are many ways to obtain the road curvature κ. For existing high-precision map data, the curvature information of the corresponding road section can be directly read from the map database. When drawing a map, the geometric shape of the road is accurately recorded through measurement and modeling, which includes curvature data. For example, some professional navigation maps will conduct detailed surveys and mapping of roads of different levels and store curvature data in the database. Through the map API interface, the relevant identification of the road section can be input to obtain the accurate curvature κ of the road section. If there is no ready-made map data support, the sensor data on the vehicle can be used for calculation. For example, the inertial measurement unit (IMU) and the global positioning system (GPS) are used to combine data. The IMU can measure information such as the acceleration and angular velocity of the vehicle, and the GPS can provide the location information of the vehicle. By processing the position and attitude data over a period of time, a mathematical algorithm (such as a curve fitting method) is used to estimate the road curvature. Assume that the vehicle is driving on a section of road, and the position coordinates of a series of time points are obtained through GPS. Combined with the attitude information measured by the IMU, these data are fitted into a curve, and then the curvature is calculated based on the mathematical properties of the curve.

[0057] Calculate α(κ) and β(κ) based on the κ range: After obtaining the road curvature κ, first determine its range. , according to the formula Then, according to Calculate β(κ). If , and α(κ)=0.4±0.02 and β(κ)=0.6 are set. Here, the value range of α(κ) is a dynamic range with a lower limit of 0.38 and an upper limit of 0.42. In practical applications, the correction coefficient 0.02 is dynamically adjusted through the Kalman filter residual variance.

[0058] Kalman filter residual variance adjustment correction coefficient: In this scenario, Kalman filter is used to estimate the system state related to road curvature, and the correction coefficient is adjusted according to the residual variance. First, a state space model related to road curvature is established, with road curvature and related parameters as state variables and vehicle sensor measurement data as observation variables. In each iteration, Kalman filter calculates the current state estimate and estimation error covariance based on the state estimate at the previous moment and the current observation data. The residual variance is part of the estimation error covariance. When the residual variance is large, it means that the current estimate is greatly different from the actual value. At this time, the correction coefficient is appropriately increased by 0.02 to make the value range of α(κ) wider to better adapt to changes in road conditions; when the residual variance is small, the correction coefficient is reduced by 0.02 to make the value of α(κ) more accurate. For example, after a period of calculation, it is found that the residual variance gradually increases. This may be due to a sudden change in road conditions, which causes the sensor measurement error to increase. At this time, the correction coefficient is increased from 0.02 to 0.03, and the value range of α(κ) becomes 0.37 to 0.43, so as to dynamically adjust the calculation results.

[0059] In another technical solution, the historical lane width in the dynamic lane width data in step S6 is calculated using an exponentially decaying moving average algorithm, and the decay factor is set to 0.85; the initial historical lane width value is taken from the nominal lane width value of the corresponding road section in the map; when the camera confidence of several consecutive frames is lower than a preset threshold, the historical lane width is forced to be reset to the map baseline value, and the lane line re-detection process based on color space segmentation is activated.

[0060] In the above technical solution, the exponential decay moving average algorithm can effectively integrate historical data with current data, give higher weight to recent data, and better fit the real-time change trend of lane width. The decay factor is set to 0.85, which is more reasonable in balancing the impact of new and old data. It neither over-relies on old data to cause slow response to new changes, nor over-emphasizes new data and ignores historical trends. It improves the accuracy and real-time performance of lane width calculation and provides a reliable lane width reference for related systems. The nominal value of the lane width of the corresponding road section in the map is used as the initial historical lane width value, which provides a reliable starting point for calculation. The map data has been professionally surveyed and mapped, and its nominal value is authoritative and accurate. It can make the calculation based on a reasonable basis during the system startup or data initialization stage, avoid subsequent calculation deviations caused by unreasonable initial values, and ensure the stability of dynamic lane width calculation. By determining that the confidence of several consecutive frames of the camera is lower than the preset threshold, the unreliable situation of the camera collected data can be discovered in time. At this time, forcibly resetting the historical lane width to the map reference value can prevent erroneous data from continuously affecting the calculation results and ensure the reliability of lane width data under abnormal conditions. At the same time, the lane line re-detection process based on color space segmentation is activated to try to obtain more accurate lane line information, recalibrate the lane width, enhance the system's ability to cope with complex environments and data anomalies, and ensure the safety and stability of related applications (such as autonomous driving assistance).

[0061] Specifically, the exponentially decaying moving average algorithm calculates the historical lane width: In the actual calculation process, the exponentially decaying moving average algorithm reflects the change of lane width by continuously updating the historical lane width value. Let the current time be t, and the historical lane width value at the previous time be , the lane width measured at the current moment is C t The calculation formula of the exponential decay moving average algorithm is: , where α is the attenuation factor, which is set to 0.85 here. For example, during the vehicle driving process, after each frame of image acquisition (assuming that the interval between each frame acquisition is a fixed time), a new current lane width measurement value C is obtained. t First, obtain the historical lane width value calculated at the last moment , then C t and Substitute the above formula for calculation. Assume that the historical lane width value at the previous moment is The lane width C is 3 meters. t If the lane width is 3.2 meters, then the current historical lane width value is calculated as 3.17 meters according to the formula. As the vehicle continues to drive, this calculation process is repeated continuously, and the historical lane width value is updated in real time, so that it can quickly respond to changes in lane width.

[0062] Get the initial historical lane width value: To obtain the lane width nominal value of the corresponding road section in the map as the initial historical lane width value, you first need to identify the road section where the vehicle is located. The vehicle's location information can be obtained through the vehicle's positioning system (such as GPS), and then combined with the map matching algorithm to determine the precise road section where the vehicle is currently located. After the road section is determined, use the map data interface to enter the unique identifier of the road section (such as the road section ID) to query the corresponding lane width nominal value from the map database. Different map suppliers may have different data structures and interface methods, but the basic principle is to retrieve relevant information through the road section identifier. After obtaining this value, use it as the initial historical lane width value for the first calculation of the subsequent exponential decay moving average algorithm.

[0063] Camera confidence determination and subsequent operations: Camera confidence is an indicator to measure the reliability of camera data collection. In the actual system, the camera will analyze each frame of the collected image and generate a confidence value. The system will set a preset threshold, such as 0.6. When the camera confidence is lower than this threshold for several consecutive frames (assuming it is set to 5 frames), the camera data is determined to be unreliable. Once it is determined to be unreliable, the historical lane width is first forced to be reset to the benchmark value previously obtained from the map. This step is controlled by the program to directly update the current historical lane width value to the map benchmark value. For example, if the map benchmark value is 3 meters, the current historical lane width value is directly set to 3 meters regardless of its current value. At the same time, the lane line re-detection process based on color space segmentation is activated. In the color space segmentation algorithm, the image is first converted from the RGB color space to a color space that is more conducive to lane line segmentation (such as the HSV space). Then, based on the color characteristics of the lane lines (such as the specific color range of white or yellow lane lines in the HSV space), possible lane line areas are extracted through methods such as threshold segmentation. After morphological processing (such as corrosion and dilation operations) and contour detection, the lane line position is accurately determined, and the lane width is recalculated to provide accurate basic data for subsequent calculations.

[0064] In another technical solution, the cubic spline interpolation control point interval of the road centerline in the Frenet frame calculation in step S7 is dynamically adjusted according to the real-time curvature: when the real-time curvature is greater than 0.1 per meter, the control point interval is shortened to 2 meters; when the real-time curvature is less than or equal to 0.1 per meter, the control point interval is extended to 10 meters; the curvature calculation uses the central difference method to obtain the second-order derivative of the road centerline parameter equation.

[0065] In the above technical solution, dynamically adjusting the cubic spline interpolation control point interval according to the real-time curvature can significantly improve the accuracy of the road centerline fitting. In sections with large curvature, shortening the control point interval to 2 meters can more carefully capture the bending changes of the road, avoid fitting deviations caused by too large intervals, and enable the road centerline to accurately reflect the actual road conditions. In sections with small curvature, the control point interval is extended to 10 meters to reduce the number of control points and reduce the calculation complexity, while ensuring the reasonable fitting of the flat sections, balancing the calculation efficiency and fitting accuracy. The central difference method is used to obtain the second-order derivative of the road centerline parameter equation to calculate the curvature, which has high calculation accuracy. The central difference method uses the information of adjacent data points and can more accurately approximate the second-order derivative of the parameter equation through a reasonable difference formula, thereby obtaining an accurate curvature value. Compared with other simple curvature calculation methods, the central difference method can effectively reduce calculation errors, provide a reliable basis for the dynamic adjustment of the control point interval, and thus improve the accuracy and reliability of the entire Frenet frame calculation. Accurately implement the operation process of dynamically adjusting the control point interval and curvature calculation, providing stable and efficient technical support for related applications. In actual scenarios such as autonomous driving and intelligent traffic monitoring, accurate road centerline information and real-time curvature data are crucial. Through a stable operation process, the system can obtain this information in real time and accurately, providing strong support for vehicle path planning, driving control, and traffic flow analysis, and improving the safety and operation efficiency of the entire transportation system.

[0066] Specifically, real-time curvature acquisition: First, it is necessary to obtain the parametric equation of the road centerline. In practical applications, the data of the road centerline can be obtained in many ways, such as high-precision map data or vehicle sensors (such as lidar, cameras, etc.) collected and processed. When using vehicle sensors (such as lidar, cameras, etc.) to collect, it is assumed that the parametric equations of the road centerline are x=x(t) and y=y(t), where t represents time in seconds, and (x, y) corresponds to the coordinates of different points on the road centerline. The curvature is calculated using the central difference method. In the calculation process, the step size h is reasonably selected to ensure the calculation accuracy and not to make the calculation too large. For example, according to the sampling frequency and accuracy requirements of the actual road data, h=0.1 is selected. The real-time curvature is obtained by continuously calculating the curvature corresponding to different t values.

[0067] Dynamic adjustment of control point interval: Dynamically adjust the cubic spline interpolation control point interval according to the calculated real-time curvature. When the real-time curvature κ>0.1 per meter, set the control point interval to 2 meters. In actual operation, starting from the starting point of the road centerline, a control point is selected every 2 meters. Assuming that the coordinates of the starting point of the road centerline are (x0, y0), the coordinates of the first control point are (x0+2, y1), the value of y1 is calculated by the parametric equation, and the subsequent control points are selected by analogy. When the real-time curvature κ≤0.1 per meter, set the control point interval to 10 meters. Also starting from the starting point, a control point is selected every 10 meters. For example, the starting point coordinates are (x0, y0), the coordinates of the first control point are (x0+10, y1), and y1 is determined by the parametric equation. In the process of selecting control points, it is necessary to ensure that the control points can cover the entire road centerline and accurately reflect the road shape in sections with different curvature changes.

[0068] Cubic spline interpolation and application: After selecting the control points, perform cubic spline interpolation. Cubic spline interpolation requires that the function value, first-order derivative, and second-order derivative are continuous at each control point. By establishing the corresponding set of equations, the coefficients of the cubic spline function are solved. Assume that the control point is (x i ,y i ), i=0,1,...,n, construct a set of equations, and use the function values ​​and derivative conditions of the control points to solve the cubic spline function S(x). In the solution process, matrix operations and other methods can be used to improve the calculation efficiency. After obtaining the road centerline after cubic spline interpolation, it can be applied to related tasks such as Frenet frame calculation. In Frenet frame calculation, the accurate representation of the road centerline is crucial for vehicle positioning, path planning, etc. The precise road centerline obtained by dynamically adjusting the control point interval can provide more accurate basic data for Frenet frame calculation, improving the reliability and practicality of the entire calculation.

[0069] In another technical solution, the parameter configuration of the extended Kalman filter satisfies: The state vector is defined as a two-dimensional vector X, which contains the longitudinal position s and the longitudinal velocity v; The process noise covariance matrix Q is configured as a diagonal matrix diag(0.1,0.05), reflecting the process noise variance of the longitudinal position and velocity; The observation noise covariance matrix R is configured as a diagonal matrix diag(0.2,0.1), which characterizes the measurement noise intensity of the longitudinal position and velocity; The state transfer matrix F adopts the uniform kinematics model: F = [[1, Δt], [0, 1]] where Δt = 0.1 seconds is the discretization time step; The observation matrix H is configured as a 2×2 unit matrix I to achieve full-dimensional observation of the state vector.

[0070] In the above technical solution, the state vector is defined as a two-dimensional vector X containing the longitudinal position s and the longitudinal velocity v, which can concisely and effectively describe the core state information of the system in the longitudinal dimension. These two variables are directly related to many practical applications, such as vehicle driving state monitoring and object motion trajectory tracking. By focusing on these two key elements, the extended Kalman filter can accurately estimate and predict the system state, provide a clear and core basis for subsequent data processing and decision-making, and improve the accuracy and pertinence of the system state description. The process noise covariance matrix Q is set to diag (0.1, 0.05), which reasonably quantifies the process noise variance of the longitudinal position and velocity. A smaller noise variance value indicates that the system process is relatively stable. Such a configuration enables the filter to better adapt to the dynamic characteristics of the system itself, effectively consider the inevitable noise interference in the estimation process, and improve the robustness of the state estimation. The observation noise covariance matrix R is configured as diag (0.2, 0.1), which accurately characterizes the measurement noise intensity. It allows the filter to reasonably weight according to the reliability of the measurement data, and when fusing the observation data, it gives a lower weight to the measurement value with larger noise, thereby improving the accuracy of the estimation result. The state transfer matrix F is based on the uniform kinematics model and adopts the form of [[1,Δt],[0,1]] with Δt=0.1 second, which is consistent with the situation of objects in many practical scenarios moving at approximately uniform speed. This configuration can effectively predict the state of the system at the next moment and provide a reasonable basis for state update for the iterative calculation of the filter. The observation matrix H is set to a 2×2 unit matrix I to achieve full-dimensional observation, ensuring that each element of the state vector can be directly observed, providing comprehensive data support for state estimation, and improving the integrity and reliability of the estimation. Based on the above parameter configuration, the extended Kalman filter can operate stably and efficiently in practical applications. In scenarios such as positioning and speed monitoring of autonomous vehicles, it can accurately integrate system process information and observation data, estimate the system state in real time and accurately, provide strong guarantees for the stable operation and decision-making of the system, and improve the performance and safety of related applications.

[0071] Specifically, state vector definition and initialization: In practical applications, we must first clarify the measurement methods of physical quantities related to the longitudinal position s and longitudinal velocity v in the system. For example, in a vehicle driving scenario, the longitudinal position s can be obtained through the vehicle's GPS device, and the longitudinal velocity v can be measured by the vehicle's speed sensor. When the system starts, the state vector needs to be initialized. According to the initial measurement values, assuming that the initial position of the vehicle when it starts is s0 and the initial velocity is v0, the initial state vector X0=[s0,v0].

[0072] Configuration of process noise covariance matrix Q: Process noise covariance matrix Q=diag(0.1,0.05). The diagonal elements of this matrix correspond to the process noise variances of the longitudinal position s and longitudinal velocity v. In practical understanding, 0.1 represents the process noise variance of the longitudinal position s, which means the degree of interference caused to the position estimation by various uncertain factors (such as road bumps, vehicle vibration, etc.) during the operation of the system. 0.05 represents the process noise variance of the longitudinal velocity v, which reflects the interference in the velocity estimation process.

[0073] Configuration of the observation noise covariance matrix R: The observation noise covariance matrix R=diag(0.2,0.1). 0.2 represents the measurement noise intensity of the longitudinal position s, and 0.1 represents the measurement noise intensity of the longitudinal velocity v. This reflects the accuracy limitation of the measurement equipment itself and the influence of environmental noise on the measurement value. For example, the GPS device has a certain error when measuring the position, and the error size is reflected in the first diagonal element of this matrix; the measurement of the velocity sensor is not completely accurate, and the second diagonal element reflects the noise intensity of the velocity measurement.

[0074] State transfer matrix F configuration: state transfer matrix F = [[1, Δt], [0, 1]], where Δt = 0.1 seconds. This matrix is ​​based on the uniform kinematics model, which means that within each time step Δt, the longitudinal position s will change linearly according to the current velocity v , while the speed v remains unchanged (v {k+1} =v k ). In practical applications, for example, during vehicle driving, assuming that the vehicle is moving at an approximately uniform speed in a short period of time, this matrix can reasonably predict the position and speed of the vehicle at the next moment.

[0075] Observation matrix H configuration: The observation matrix H is set to a 2×2 unit matrix I. The role of the unit matrix is ​​to achieve full-dimensional observation of the state vector, that is, the measurement value can directly reflect each element of the state vector X. In actual measurement, for example, the measurement equipment can directly obtain the position and speed of the vehicle. Then, through this unit matrix, the measurement value can be directly matched with the elements of the state vector, which is convenient for subsequent filtering calculations.

[0076] Extended Kalman filter operation: After completing the above parameter configuration, the extended Kalman filter runs according to its standard process. At each time step, the state is first predicted based on the state transfer matrix F and the process noise covariance matrix Q to obtain the predicted state X {k|k-1} and the predicted covariance P {k|k-1} Then, combining the observation matrix H, the observation noise covariance matrix R and the actual measurement value Z k , through the Kalman gain K kUpdate the state and get the updated state estimate X {k|k} and covariance P {k|k} Throughout the process, the parameters of the above configurations cooperate with each other to ensure that the filter can accurately estimate the system state. k represents a time step identifier, which is used to represent a discrete time point and is a positive integer. .

[0077] The operation process of this technical solution is as follows: first, measure the longitudinal position of the vehicle (for example, the position of the vehicle in the forward direction) and the longitudinal speed (the speed at which the vehicle is moving). The longitudinal position can be determined by the GPS device on the vehicle, and the longitudinal speed can be measured by the vehicle speed sensor. At startup, initialize the state vector. If the position at startup is s0 and the speed is v0, the initial state vector X0 is set to [s0, v0]. Then configure the process noise covariance matrix Q=diag(0.1,0.05). 0.1 indicates the degree of interference of various uncertain factors (such as uneven road surface and vehicle vibration) in the position measurement, and 0.05 indicates the degree of interference of the speed measurement. Then configure the observation noise covariance matrix R=diag(0.2,0.1), where 0.2 indicates that the position measurement device (such as GPS) itself has limited accuracy, and the environmental noise will cause the measured position to have errors; 0.1 indicates that the speed measurement device (speed sensor) is not completely accurate, and the measured speed will also have noise. Then configure the state transfer matrix F= , Δt=0.1 seconds. This matrix is ​​based on a model in which the vehicle travels at an approximately uniform speed in a short period of time, that is, every 0.1 seconds, the position of the vehicle will move forward a certain distance according to the current speed (position change = current speed × 0.1 seconds), while the speed remains unchanged. Then configure the observation matrix H, and set H to the 2×2 unit matrix I, indicating that the measurement equipment can directly measure the position and speed of the vehicle, and the measured values ​​can directly correspond to the position and speed elements in the state vector, so as to facilitate subsequent filtering calculations. After configuring the above parameters, the extended Kalman filter can start working according to its standard process. Each time point is represented by k. First, the state of the vehicle at the current moment is predicted based on the state transfer matrix F and the process noise covariance matrix Q (the predicted state X is obtained). {k|k-1} ) and the uncertainty of the predicted state (obtaining the predicted covariance P {k|k-1} Then, we combine the observation matrix H, the observation noise covariance matrix R, and the actual measured vehicle position and speed values ​​(Z k ), by calculating the Kalman gain K k To update the vehicle's state estimate (get the updated state estimate X {k|k} ) and the uncertainty of the updated state estimate (obtaining the covariance P {k|k} ). These configured parameters work together to ensure that the filter can accurately estimate the state of the vehicle at each moment.

[0078] In another technical solution, the ROI area and processing flow of the image preprocessing module include: S501, in the pixel coordinate system of the front-view camera, a rectangular area from the upper left corner coordinate to the lower right corner coordinate of the road detection range of 100 meters in front of the vehicle is selected as the ROI area; S502, obtaining an input image, converting it from RGB format to YUV420 format, wherein the chrominance component adopts a 4:2:0 subsampling mode to generate a YUV420 format image with a 30% reduction in chrominance data volume; S503, performing a CLAHE algorithm of grid division on the Y component of the YUV420 format image in S502, setting an 8×8 pixel grid unit, a contrast limit threshold of 2.0, and using bilinear interpolation to eliminate the grid boundary effect, thereby generating a processed Y component; S504: Perform bilateral filtering on the U / V components of the YUV420 format image in S502, and set the spatial domain standard deviation σ s = 1.5 pixels, color domain standard deviation σ c =15 gray level, filter window diameter d=5 pixels, generating filtered U / V components; S505, combining the processed Y component generated in S503 with the filtered U / V component generated in S504, and outputting the pre-processed image data in YUV420 format; Among them, the Y component represents the brightness component, the U component represents the difference between the blue chromaticity information and the brightness information in the image, and the V component represents the difference between the red chromaticity information and the brightness information in the image.

[0079] In the above technical solution, a rectangular ROI area covering the 100-meter road detection range in front of the vehicle is selected in the pixel coordinate system of the front-view camera, which can focus on the key area. For vehicle-related applications, such as automatic driving assistance systems, data processing in unnecessary areas can be reduced, processing efficiency can be improved, and the road conditions ahead can be analyzed more targeted, the accuracy of road target detection and recognition can be enhanced, and more effective data support can be provided for subsequent decision-making. Convert the image from RGB format to YUV420 format and adopt 4:2:0 subsampling mode to reduce the amount of chromaticity data by 30%. This reduces the amount of data, storage and transmission costs, and improves the image processing speed without affecting the key information of the image (the brightness component Y is kept intact), which is especially suitable for scenes with high real-time requirements, such as fast image processing during vehicle driving. Y component CLAHE algorithm processing: The Y component of the YUV420 format image is subjected to the CLAHE algorithm of grid division, setting the 8×8 pixel grid unit and the contrast limit threshold to 2.0, and using bilinear interpolation to eliminate the grid boundary effect. The CLAHE algorithm can enhance the local contrast of the image, making the road details clearer, which is conducive to the subsequent recognition of road signs, lane lines, etc. Bilinear interpolation eliminates the grid boundary effect, ensures a smooth transition of the processed image, does not produce obvious block effects, and improves the image quality. Perform bilateral filtering on the U / V components and set the appropriate spatial domain standard deviation σ s =1.5 pixels, color domain standard deviation σ c =15 grayscale levels and filter window diameter d=5 pixels. Bilateral filtering can effectively remove noise and improve image quality while retaining image edge information. For U / V components (chrominance components), it can make color transitions more natural, improve the overall visual effect of the image, and provide more accurate color information for subsequent image analysis. Accurately merge the processed Y component with the filtered U / V component and output the YUV420 format preprocessed image data, ensuring that the image information after different processing steps is complete and accurately integrated. The YUV420 format has good compatibility in many video processing and image analysis scenarios. Outputting preprocessed images in this format can facilitate subsequent docking with other modules and provide high-quality input data for the entire image processing process.

[0080] The present invention provides a comprehensive index calculation system for an autonomous driving scenario library, comprising: A data acquisition module, which is used to obtain three-dimensional positioning data, wheel speed pulse data, and image data; The data fusion and completion module is used to input the three-dimensional positioning data and the wheel speed pulse data into the extended Kalman filter to generate fused trajectory data; detect the continuous null value area of ​​several frames in the fused trajectory data to generate missing trajectory mark data; input the fused trajectory data and the missing trajectory mark data into the cascade structure of the temporal convolution network and the micro-gated recurrent unit network, and output the completed trajectory data; An image preprocessing module is used to perform ROI region cropping and grayscale processing on image data to generate preprocessed image data; Dynamic lane calculation module, which is used to fuse pre-processed image data with historical trajectory database and map data to calculate dynamic lane width data. The output formula is W real =0.6W cam +0.4W hist ; A local coordinate conversion module, which is used to generate local coordinate system parameters through Frenet frame calculation based on the completed trajectory data and dynamic lane width data; The collision time calculation module is used to project the completed trajectory data into the local coordinate system and decompose it into the longitudinal velocity component v x With the lateral velocity component v y According to v x 、v y The relative distances Δx and Δy from the target object are used to calculate the longitudinal collision time TTC. x =Δx / v x The time to lateral collision TTC y =Δy / v y ; Dynamic risk indicator module, which is used to integrate TTC x 、TTC y And the road curvature κ extracted from the map, through the formula Generate a dynamic collision risk index where α(κ)+β(κ)=1 and α(κ)=0.95-0.5κ; The scene classification module is used to judge the variance of the DRC index and the lateral velocity component based on a preset threshold value, and output classification labels for the following scene, lane change scene and intersection scene.

[0081] Although the embodiments of the present invention have been disclosed as above, they are not limited to the applications listed in the specification and the implementation modes, and they can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily implemented. Therefore, without departing from the general concept defined by the claims and the scope of equivalents, the present invention is not limited to the specific details and the illustrations shown and described herein.

Claims

1. A comprehensive index calculation method for an autonomous driving scenario library, characterized in that: The following steps are involved: S1, obtain three-dimensional positioning data through the vehicle-mounted GNSS module, obtain wheel speed pulse data through the wheel speed sensor, and obtain image data through the front-view camera; S2, inputting the three-dimensional positioning data and wheel speed pulse data of step S1 into the extended Kalman filter to generate fused trajectory data; S3, detecting a plurality of consecutive frames of null value regions in the fused trajectory data of step S2, and generating missing trajectory mark data; S4, inputting the fused trajectory data of step S2 and the missing trajectory mark data of step S3 into the cascade structure of the temporal convolutional network and the micro-gated recurrent unit network, and outputting the completed trajectory data; S5, performing ROI region cropping and grayscale processing on the image data of the front-view camera in step S1 to generate pre-processed image data; S6: The pre-processed image data of step S5 is integrated with the historical trajectory database and map data to calculate the dynamic lane width data. The output formula is W real =0.6W cam +0.4W hist , where W real Represents the true value of lane width, W cam Indicates the value measured by the camera, W hist Indicates the lane width history value; S7, based on the completed trajectory data of step S4 and the dynamic lane width data of step S6, generate local coordinate system parameters through Frenet frame calculation; S8, projecting the completed trajectory data of step S4 to the local coordinate system of step S7, and decomposing it into the longitudinal velocity component v x With the lateral velocity component v y ; S9, according to step S8 v x 、v y The relative distances Δx and Δy from the target object are used to calculate the longitudinal collision time TTC. x =Δx / v x The time to lateral collision TTC y =Δy / v y ; S10, TTC of fusion step S9 x 、TTC y And the road curvature κ extracted from the map, through the formula Generate a dynamic collision risk index, where α(κ)+β(κ)=1 and α(κ)=0.95-0.5κ, where α(κ) and β(κ) represent weight coefficients related to the road curvature κ; S11. The variance of the DRC index and the lateral velocity component of step S10 is judged based on a preset threshold value, and classification labels of the following scene, the lane change scene and the intersection scene are output.

2. The comprehensive index calculation method for the automatic driving scenario library according to claim 1, characterized in that: The method for constructing the temporal convolutional network includes: S401, configure a three-level cascaded dilated convolution module, set the dilated ratios of the first-level dilated convolution module, the second-level dilated convolution module, and the third-level dilated convolution module to 1, 2, and 4, respectively, preset 64 convolution kernels of size 3×3 for each dilated convolution module, set the convolution sliding step to 1, and use a symmetric filling method in the time dimension to maintain the size of the feature map; S402, inputting the feature map output by the front neural network into the first-level dilated convolution module to generate a primary feature map, inputting the primary feature map into the second-level dilated convolution module to generate an intermediate feature map, and inputting the intermediate feature map into the third-level dilated convolution module to generate a high-level feature map; S403, grouping the 64 convolution kernels of the first-level dilated convolution module, the second-level dilated convolution module, and the third-level dilated convolution module, and splitting the 64 convolution kernels of each level of dilated convolution module into 4 groups, each group containing 16 convolution kernels, and sharing parameters within the group; S404, performing structured pruning on the grouped 64 convolution kernels, retaining the first 50% of the highly activated convolution kernels in each level of the module through L1 norm analysis, and setting the remaining 50% of the convolution kernels to zero; S405, performing channel compression on the high-level feature map output by the pruned third-level hole convolution module through 1×1 convolution to generate a compressed high-level feature map; S406, performing feature fusion on the compressed high-level feature map and the input feature map of the micro-gated recurrent unit network through a tensor addition operation to generate a residual connection feature map; S407, applying a linear unit activation function with leakage correction to the output feature map of each dilated convolution module, wherein the leakage factor of the linear unit with leakage correction is set to 0.01; S408, outputting the residual connection feature map to the micro-gated recurrent unit network, and outputting the completed trajectory data.

3. The comprehensive index calculation method for the automatic driving scenario library according to claim 2, characterized in that: The configuration method of the micro-gated recurrent unit network includes: S409, constructing a micro gated recurrent unit network, wherein the hidden layer adopts a 16-node topology structure, and the gated unit weight initialization follows the Xavier normal distribution; S410, performing 8-bit integer quantization on the 32-bit floating-point weight Wfloat32 of the network, converting it into 8-bit integer weight Wint8, and recalculating quantization parameters μ and σ every 1000 training iterations, where μ and σ are the mean and standard deviation of Wfloat32 respectively; S411, dequantize Wint8 through inverse transformation to obtain the dequantized floating point weight W dequant , ; S412. Use the straight-through estimator (STE) to approximate the gradient in the back-propagation stage, keeping the gradient of the rounding operation to 1.

4. The comprehensive index calculation method for the automatic driving scenario library according to claim 3, characterized in that: Methods for calculating α(κ) and β(κ) based on road curvature include: S1001, obtaining road curvature κ; S1002, determine the range of κ: like , according to the formula Calculate α(κ), and then according to Calculate β(κ); like , α(κ)=0.4±0.02 and β(κ)=0.6 are forced to be set, and the correction coefficient 0.02 is dynamically adjusted through the Kalman filter residual variance.

5. The comprehensive index calculation method for the automatic driving scenario library according to claim 1, characterized in that: The historical lane width in the dynamic lane width data in step S6 is calculated using an exponentially decaying moving average algorithm, with the decay factor set to 0.85; the initial historical lane width value is taken from the nominal lane width value of the corresponding road section in the map; When the camera confidence of several consecutive frames is lower than the preset threshold, the historical lane width is forced to be reset to the map reference value, and the lane line re-detection process based on color space segmentation is activated.

6. The comprehensive index calculation method for the automatic driving scenario library according to claim 5, characterized in that: The cubic spline interpolation control point interval of the road centerline in the Frenet frame calculation in step S7 is dynamically adjusted according to the real-time road curvature: when the real-time road curvature is greater than 0.1 per meter, the control point interval is shortened to 2 meters; when the real-time road curvature is less than or equal to 0.1 per meter, the control point interval is extended to 10 meters; the road curvature calculation adopts the central difference method to obtain the second-order derivative of the road centerline parameter equation.

7. The comprehensive index calculation method for the automatic driving scenario library according to claim 1, characterized in that: The parameter configuration of the extended Kalman filter satisfies: The state vector is defined as a two-dimensional vector X, which contains the longitudinal position s and the longitudinal velocity v; The process noise covariance matrix Q is configured as a diagonal matrix diag(0.1,0.05), reflecting the process noise variance of the longitudinal position and velocity; The observation noise covariance matrix R is configured as a diagonal matrix diag(0.2,0.1), which characterizes the measurement noise intensity of the longitudinal position and velocity; The state transfer matrix F adopts the uniform kinematics model: F = [[1, Δt], [0, 1]], where Δt = 0.1 seconds is the discretization time step; The observation matrix H is configured as a 2×2 unit matrix I to achieve full-dimensional observation of the state vector.

8. The comprehensive index calculation method for the automatic driving scenario library according to claim 1, characterized in that: The ROI area and processing flow of the image preprocessing module include: S501, in the pixel coordinate system of the front-view camera, a rectangular area from the upper left corner coordinate to the lower right corner coordinate of the road detection range of 100 meters in front of the vehicle is selected as the ROI area; S502, obtaining an input image, converting it from RGB format to YUV420 format, wherein the chrominance component adopts a 4:2:0 subsampling mode to generate a YUV420 format image with a 30% reduction in chrominance data volume; S503, performing a CLAHE algorithm of grid division on the Y component of the YUV420 format image in S502, setting an 8×8 pixel grid unit, a contrast limit threshold of 2.0, and using bilinear interpolation to eliminate the grid boundary effect, thereby generating a processed Y component; S504: Perform bilateral filtering on the U / V components of the YUV420 format image in S502, and set the spatial domain standard deviation σ s = 1.5 pixels, color domain standard deviation σ c =15 gray level, filter window diameter d=5 pixels, generating filtered U / V components; S505, combining the processed Y component generated in S503 with the filtered U / V component generated in S504, and outputting the pre-processed image data in YUV420 format; Among them, the Y component represents the brightness component, the U component represents the difference between the blue chromaticity information and the brightness information in the image, and the V component represents the difference between the red chromaticity information and the brightness information in the image.

Images