Hierarchical decoupling method for autonomous driving city scenario

Abstract

The application belongs to the technical field of automatic driving, and discloses a hierarchical decoupling method for an automatic driving urban scene, which comprises the following steps: step one, unifying static scene elements in a test scene to a traffic infrastructure coordinate system, and obtaining a grid map by using a probability occupation method; step two, unifying vehicle-mounted multi-sensor data to a vehicle body coordinate system, and then establishing a time sequence diagram by using a Kalman filter prediction; step three, based on the time sequence diagram, constructing an attention model by using a convolutional neural network and extracting traffic flow characteristics; step four, decoupling the nonlinear system by using a decoupling pre-compensator, and then outputting element characteristics after decoupling; and step five, clustering and extracting evaluation elements based on element characteristics obtained by decoupling of the modified decoupling pre-compensator. The application combines static scene elements, sensor data and a multivariable control system at the vehicle individual level, and improves the confidence in the test.

Description

Technical Field

[0001] This invention belongs to the field of autonomous driving technology, specifically relating to a layered decoupling method for autonomous driving in urban scenarios. Background Technology

[0002] In urban scenarios, the vehicle driving environment is highly complex and random. Static and dynamic information within the scenario are coupled with various levels of the vehicle itself. Current testing methods for scenario information acquisition and reconstruction rarely consider the impact of hierarchical coupling. High coupling between levels affects the scenario's complexity, coverage, and generalization ability, making it difficult to reflect the real, highly complex, and random traffic environment; consequently, test confidence is insufficient. While existing technologies offer solutions based on the vehicle's physical structure, logical relationships, and application functions for hierarchical decoupling, these solutions only consider the vehicle's own influence. The analysis, classification, and extraction of evaluation elements cannot adapt to the actual situation in the complex urban environment. Therefore, the coupling problem between different levels of the scenario still exists in current testing. A comprehensive consideration of environmental factors is needed for urban scenarios to improve test confidence. Summary of the Invention

[0003] The purpose of this invention is to provide a layered decoupling method for urban autonomous driving scenarios, which addresses the technical problem in existing technologies where the high complexity and randomness of urban environments make it difficult to adapt to complex environments by considering only the vehicle's own factors. This results in the determination of evaluation elements still being affected by the coupling problem between different levels of the scenario, leading to insufficient confidence in the test.

[0004] The layered decoupling method for urban autonomous driving scenarios includes the following steps.

[0005] Step 1: Unify the static scene elements in the test scenario to the coordinate system of transportation infrastructure, then divide the test environment and use the probabilistic occupancy method to obtain a grid map.

[0006] Step 2: Unify the data from the vehicle's multiple sensors into the vehicle coordinate system, and then use a Kalman filter to predict and build a time series diagram.

[0007] Step 3: Based on the corrected and updated time series graph, use a convolutional neural network to build an attention model and extract traffic flow features related to autonomous driving actions.

[0008] Step 4: The vehicle chassis adopts an integrated multivariable control nonlinear system. The nonlinear system is decoupled by a decoupling pre-compensator, and then the decoupled feature elements are output.

[0009] Step 5: Based on the modified decoupling pre-compensator, decouple the obtained feature characteristics to complete the clustering of evaluation features, thereby extracting evaluation features related to traffic flow characteristics.

[0010] Preferably, in step one, the probabilistic occupancy method for generating a grid map includes: collecting driving area information through laser ranging, gridding the driving area, using probability values ​​to represent the probability of each grid cell being occupied, setting a threshold for the probability of each grid cell being occupied, determining that the grid cell is occupied when the probability is not less than the threshold, and determining that the grid cell is not occupied when the probability is less than the threshold, thereby obtaining a grid map by determining whether each grid cell is occupied.

[0011] Preferably, in step two, when establishing the time series diagram, the predicted data of the time series diagram is compared with the actual data in the raster map, and the observation noise and process noise are manually adjusted multiple times to correct and update the Kalman filter so that the time series diagram reaches the required accuracy.

[0012] Preferably, the method for establishing a time series graph includes: first, collecting sensor data from vehicle-mounted sensors such as cameras, radar, and lidar, and performing preprocessing such as denoising and normalization; then, using a transformation matrix to unify the sensor data into the vehicle coordinate system based on the positional relationship between sensor coordinates and vehicle coordinates; constructing a prediction model and a measurement model for the time series graph, calculating the prior estimate covariance, updating the current time estimate based on the measured value and the prior estimate, and then calculating and updating the posterior estimate covariance; adjusting the process noise based on the accuracy of the motion model, and adjusting the observation noise based on sensor performance; setting the initial estimate to 0 and the initial covariance to 1; comparing the predicted data of the time series graph with the actual data in the raster map, and manually adjusting the observation noise and process noise multiple times to achieve the required accuracy in the prediction results; then repeating the prediction and update steps to obtain new state estimates and covariance matrices, thereby drawing time series graphs for various urban scenarios and situations.

[0013] Preferably, in step three, the image is input into the first convolutional layer of the attention model to form a feature map. This convolutional layer contains a set of learnable convolutional kernels for detecting basic features such as edges and corners in the image. A max pooling layer is inserted after the convolutional layer to preserve the features of the feature map and simplify the output of the convolutional layer. After passing through a second convolutional pooling layer, the resulting pixel map is flattened and input into a fully connected layer to output traffic flow features. When the features do not meet the accuracy requirements, backpropagation is performed, and the optimizer updates the weights and biases of each channel of the weighted convolutional kernel. The attention model then repeatedly processes and outputs the extraction results until the extraction results meet the accuracy requirements.

[0014] Preferably, in step four, the same feature data is collected from the time series diagram and compared with the decoupled feature data. When the root mean square error is greater than the error threshold, the decoupling pre-compensator is corrected.

[0015] Preferably, in step four, the relationship between input and output variables is analyzed based on the physical structure, logical relationship, and application function of the autonomous vehicle. Based on the analysis results, the Interactor algorithm is used to determine the invertibility of the corresponding multivariable control nonlinear system and to determine the inverse system model. The original system is compensated into a pseudo-linear system with linear transitivity using the inverse system model, and the inverse system model is a decoupling pre-compensator.

[0016] Preferably, in step five, clustering of evaluation elements is completed based on traffic flow characteristics. If the cluster boundaries are blurred and cannot be extracted, the edges in the data map are removed to optimize the data map. Clustering is then performed based on the optimized data map until the evaluation elements related to traffic flow characteristics are successfully extracted.

[0017] Preferably, in step five, for the previously obtained feature elements, each feature element is unified to a preset interval to obtain a normalized value for each evaluation element; then, a data spectrum is determined based on the normalized value, where each data point in the data spectrum is an evaluation element corresponding to the normalized value; for each data point, K neighboring points are found; the data point is connected to its surrounding neighboring points to form a local point cluster structure; the weights between the data points are calculated; the dimensionality-reduced feature vector and feature value are obtained based on the weights; and a low-dimensional matrix is ​​obtained based on the feature value and feature vector to complete the clustering of feature elements.

[0018] This invention has the following advantages: It utilizes a grid map established at the road infrastructure level to correct and update a time-series map established at the traffic flow level. The time-series map is built based on sensor data from vehicle-mounted sensors, and further uses a convolutional neural network to extract traffic flow features related to autonomous driving actions. At the vehicle-level, a decoupling pre-compensator is used to hierarchically decouple multiple variables of the nonlinear system. The decoupled features correspond to traffic flow characteristics, and the accuracy of the decoupling model is verified by collecting the same feature data from the time-series map, thereby correcting the model to improve decoupling accuracy. Subsequently, a data map is built based on feature data, and clustering of evaluation elements is completed based on traffic flow characteristics. In this way, this invention combines static scene elements at the road infrastructure level, sensor data at the traffic flow level, and a multivariable controlled nonlinear system at the vehicle-level, allowing complex environmental factors in urban scenarios to be used to optimize the decoupling pre-compensator and guide the clustering of evaluation elements. This overcomes the problem that the determination of evaluation elements is still affected by the coupling between different levels of the scene, thus improving the confidence level in autonomous driving technology testing. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the basic steps of the hierarchical decoupling method for autonomous driving in urban scenarios according to the present invention.

[0020] Figure 2 This is a flowchart of the hierarchical decoupling method for urban autonomous driving scenarios according to the present invention. Detailed Implementation

[0021] The following detailed description of the embodiments, with reference to the accompanying drawings, will further illustrate the specific implementation of the present invention, in order to help those skilled in the art to have a more complete, accurate, and thorough understanding of the inventive concept and technical solutions of the present invention.

[0022] like Figures 1-2 As shown, the present invention provides a layered decoupling method for urban autonomous driving scenarios, including the following steps.

[0023] Step 1: Unify the static scene elements in the test scenario to the coordinate system of transportation infrastructure, then divide the test environment and use the probabilistic occupancy method to obtain a grid map.

[0024] Static scene elements include static obstacles, lane lines, and traffic signs, which belong to the road infrastructure level. The probabilistic occupancy methods for generating raster maps include the following.

[0025] Information about the driving area is collected using laser ranging, and the driving area is then rasterized and denoted as m. Thus, m = {m i}, i = 1, ..., M, where M is the total number of grid cells; m i This represents the i-th grid cell, with a value of 0 or 1. 0 represents an empty area, i.e., a drivable area; 1 represents a placeholder area, i.e., an occupied area.

[0026] Let S be the data obtained from a single measurement by the laser rangefinder, where S = {s1, ..., s2} N Let N be the total number of laser data points, and each point in S contain distance and angle information. Calculate the probability that a grid cell is occupied. Assuming that grid cells are independent of each other, the formula for the probability of a grid cell being occupied is as follows:

[0027]

[0028] In the above formula, P(m) i |s1,…,s N ) represents the probability of a raster map being occupied, and the probability of each raster cell being occupied is estimated using the data S obtained from laser measurement as a condition. i |s1,...,s N The probability of each grid cell being occupied satisfies the following formula:

[0029]

[0030] That is, the probability of being occupied / the probability of being idle. According to Bayes' rule, the recursive form of the above formula is:

[0031]

[0032] Among them, l i,N The update formula is as follows:

[0033]

[0034] The threshold for the probability of each grid cell being occupied is set to 0.6, meaning that when the probability is greater than 0.6, m... i The value is 1; when the probability is less than 0.6, m i The value is 0. The raster map is calculated using the probability placeholder method described above.

[0035] Step 2: Unify the data from the vehicle's multiple sensors into the vehicle coordinate system, and then use a Kalman filter to predict and build a time series map. When building the time series map, compare the predicted data of the time series map with the actual data in the grid map, and correct and update the Kalman filter to make the time series map reach the required accuracy.

[0036] The method for establishing a time sequence diagram includes the following steps: First, sensor data from vehicle-mounted sensors such as cameras, radar, and lidar are collected and preprocessed, including noise reduction and normalization. Then, a transformation matrix is ​​used to unify the sensor data into the vehicle coordinate system based on the positional relationship between the sensor coordinates and the vehicle coordinates.

[0037] Construct prediction and measurement models for the time series plot, calculate the prior estimate covariance, update the current time-time estimate based on the measured and prior estimates, and then calculate and update the posterior estimate covariance. Adjust process noise based on the accuracy of the motion model and observation noise based on sensor performance. The initial estimate P0 is set to 0, and the initial covariance is... Set the value to 1. Compare the predicted data from the time series plot with the actual data in the raster map. After multiple manual adjustments to the observation noise and process noise, the prediction results are made to meet the required accuracy. Then, repeat the prediction and update steps, that is, repeat the prediction and update for new observation data to obtain new state estimates and covariance matrices, thereby drawing time series plots for various urban scenarios and situations.

[0038] The above process takes the observation of the environment on a straight road as an example. When the vehicle in front is moving in a straight line at a constant speed, the position p of the vehicle in front is extracted. i and speed V i .

[0039] The prediction model is then:

[0040]

[0041] The formula for calculating the prior estimate of covariance is:

[0042]

[0043] in This is the prior estimate of the current position. This is the prior estimate of the current speed. This is the estimated position value from the previous moment. Here is the estimated position value at the previous moment, Δt is the time increment, and a is the position estimate at the previous moment. i This represents the current acceleration.

[0044] The covariance of the current position is estimated prior. The covariance of the current velocity prior estimate The estimated position covariance at the previous time step. The velocity covariance estimate at the previous moment, where Q is the process noise covariance.

[0045] The measurement model is:

[0046]

[0047] Among them, z p The measured value at the current location, z v This is the current speed measurement.

[0048] The estimated value at the current time is updated based on the measured value and prior estimate obtained from the above calculation formula. The update method includes: first calculating the Kalman gain:

[0049]

[0050] The covariance of the previous position prior estimate The prior position covariance at the previous time step is given by: Q, the process noise covariance, and R, the observation noise covariance. The formula for correcting the estimate is:

[0051]

[0052] in, This is the estimated position at the current moment. Let H be the estimated position at the current time, and H be the state observation matrix. The formula for updating the posterior covariance estimate is:

[0053]

[0054] The above method adjusts process noise based on the accuracy of the motion model and observation noise based on sensor performance. The initial estimate P0 is set to 0, and the initial covariance is... Set the value to 1. Compare the predicted data with the actual data from the raster map, and manually adjust the observation noise and process noise multiple times to achieve the required accuracy.

[0055] Step 3: Based on the corrected and updated time series graph, use a convolutional neural network to build an attention model and extract traffic flow features related to autonomous driving actions.

[0056] Based on the aforementioned time series diagram, an attention model is designed using a convolutional neural network (CNN). The collected raw data is mapped into a high-dimensional feature space through convolution and pooling. SENet (channel attention) is then used to automatically identify and focus on the most important features in the traffic flow, such as distance to vehicles ahead, speed differences, and potential pedestrian conflict areas, thus forming traffic flow features.

[0057] In the attention model, the image passes through a first convolutional layer containing a set of learnable convolutional kernels. These kernels are optimized during training to detect basic features such as edges and corners in the image, forming a feature map. After the convolutional layer, a max-pooling layer is inserted, preserving the features of the feature map and simplifying the output of the convolutional layer. This is followed by a second convolutional-pooling layer. The resulting pixel map is flattened and fed into a fully connected layer, which outputs the desired traffic flow features. When the features do not meet the accuracy requirements, backpropagation is performed, using an optimizer (such as gradient descent) to update the weights and biases of each channel of the convolutional kernels. The attention model then repeatedly processes the extracted results until the required accuracy is achieved.

[0058] For example, in the process of identifying the vehicle in front, multiple convolutional kernels in the convolutional layer first extract pixel features from the temporal image to form multiple feature maps. Each convolutional kernel has multiple channels and bias terms. The pooling layer continues to extract features from the feature maps, reducing the feature dimension while preserving the features. The second convolutional pooling layer repeats the above steps to further reduce the feature map dimension. Finally, the low-dimensional feature map is flattened and output through a fully connected layer to obtain the outline of the vehicle in front and the attention score, which do not meet the accuracy standard. The weights and bias terms of the convolutional kernels are updated through backpropagation to increase the attention to important parts and improve accuracy. This guides the extraction of the outline of the vehicle in front to meet the required accuracy. Similarly, other traffic flow features can be extracted.

[0059] Creating time-series diagrams and extracting traffic flow features belong to the traffic flow level.

[0060] Step 4: The vehicle chassis adopts an integrated multivariable control nonlinear system. The nonlinear system is decoupled by a decoupling pre-compensator, and the decoupled feature is output. The same feature is collected from the time series diagram and compared with the decoupled feature. The decoupling pre-compensator is corrected based on the comparison result.

[0061] The vehicle's decision-making layer makes behavioral decisions based on the above characteristics, and the vehicle chassis implements corresponding control based on these behavioral decisions. The vehicle chassis adopts an integrated multivariable control nonlinear system, i.e., a multivariable input-output system. Its output is: y = h(x, u); where x = (x1, x2, ... x... n ) T ∈R n Here, u is the system's state variable, and n is the dimension. u = (u1, u2, ..., un) q ) T ∈R q The system's input variables are y = (y1, y2, ..., y3). q )∈R q This is the system output vector. Therefore, this nonlinear system is an n-order nonlinear dynamic system with q-dimensional input and q-dimensional output. The above nonlinear system can be expressed as: The Interactor algorithm is used to prove the invertibility of multivariable systems. For the vehicle nonlinear system, [δ f T] T System input, [v ω r ] T It is a state variable.

[0062] In the "automatic obstacle avoidance" function test, taking the drive-by-wire chassis in the actuator (physical structure) as an example, the analysis shows that there is a serious coupling between active steering control and active braking control in the "automatic obstacle avoidance" function test. The control structure generates additional front wheel steering angle and yaw moment through the vehicle's active steering system and active braking system for integrated chassis control. Since the tire lateral force is related to both the sideslip angle and the vertical load, and both are affected by the center of gravity sideslip angle and yaw rate, the yaw rate and the center of gravity sideslip angle are coupled through the tire action. When active braking is applied to control the vehicle to drive stably, the yaw moment will also cause the vehicle to yaw. Under the combined action of the yaw rate and the longitudinal velocity, the inertial effect will be manifested in the non-inertial frame, causing the vehicle to move laterally. There is coupling between the yaw and lateral degrees of freedom. Nonlinear decoupling control is required to reasonably configure active braking and active steering to eliminate their interference and coupling.

[0063] Therefore, the vehicle system output variables are the sideslip angle and yaw rate, denoted as [βω]. r ] T For a nonlinear system, find the derivative of y1 with respect to time up to the r-th... i Derivative The variable r clearly contains input variables, which is the output variable. i The Jacobian matrix of the first derivative with respect to the input variable has a rank of 1, expressed as: Then the relative order is α1 = r i .

[0064] Differentiate y2 in the formula until the rj-th derivative is obtained: For a vehicle nonlinear system, there exist non-negative integers 1 and 1 such that the r-th... j The formula for the order derivative is equal to the number of outputs of the system, then the relative order vector of the system is r = [1 1]. T ,and (2 is the order of the system). By the implicit function existence theorem, we can conclude that the original vehicle system has an inverse system, which can be represented as follows: in This represents the functional mapping relationship between the input and output in the inverse system. Z1 = β, Z2 = ω r , If the system is fed into the inverse system, then the inverse system model can be expressed as follows: The inverse system model is connected in series with the vehicle model to compensate the original system into a pseudo-linear system with a linear transfer relationship, thereby achieving decoupling of the original system. This inverse system model is the decoupling pre-compensator.

[0065] Furthermore, for multivariable control systems, by providing a single-input excitation, the relative gain matrices of the front wheel steering angle and yaw moment with respect to the center of gravity sideslip angle are determined. When the relative gain is close to 1, the system can be paired using this variable pairing relationship, i.e., the front wheel steering angle controls the center of gravity sideslip angle, and the yaw rate controls the yaw rate. An inverse system model is then connected in series before the original system to compensate for the original system, transforming the multi-input multi-output system into multiple single-input single-output systems. In the active front wheel steering angle and active yaw moment calculated by the decoupled pre-compensator, the active front wheel steering angle can be directly obtained from the active steering actuator; while the active yaw moment can be calculated by applying braking forces to different wheels. After decoupling the design, at the vehicle-level, this method can extract the output variables, center of gravity sideslip angle and yaw rate, as feature elements for extracting evaluation elements.

[0066] Feature characteristics, such as sideslip angle and yaw rate, can also be obtained from the time series plot. Compare the output variables (slip angle and yaw rate) with those obtained from the time series plot. When the root mean square error exceeds the error threshold of 1%, the inverse system model needs to be modified, such as by improving the tire model or introducing additional dynamic effects.

[0067] Step 5: Based on the modified decoupling pre-compensator, decouple to obtain element features and obtain feature vectors and feature values ​​through normalization and local linear embedding. Then, based on the guidance of traffic flow features, complete the clustering of evaluation elements to extract evaluation elements related to traffic flow features.

[0068] For the previously obtained feature elements, each feature element is unified to a preset interval to obtain a normalized value for each evaluation element. Then, a data spectrum is determined based on the normalized value, where each data point in the data spectrum is an evaluation element corresponding to the normalized value. For each data point, K neighboring points are found. The data point is connected to its surrounding neighboring points to form a local point cluster structure. The weights between the data points are calculated. The dimensionality-reduced feature vector and feature value are obtained based on the weights. A low-dimensional matrix is ​​obtained based on the feature value and feature vector to complete the clustering of feature elements.

[0069] Specifically, in this embodiment, all feature elements are unified into a roughly similar numerical range to ensure comparability between features of different dimensions. After normalization, all evaluation feature elements are treated as points, and these points are used to construct a graph, with the high-dimensional sample points X... i It is represented by its k nearest neighbors, where the superscript (k) indicates the neighbor number, as shown in the following expression:

[0070]

[0071] Where, x i ∈[1, k], w i ∈[1, k], make That is, w i I k×1 =1; then we have:

[0072] Construct the Lagrange equation:

[0073]

[0074] Differentiating gives in, Will Substituting, we get: Obtain the weight w i .

[0075] Then, a low-dimensional space mapping is performed, resulting in a low-dimensional space point Y. i Can use the k nearest neighbors and w i The expression is:

[0076] The constraints of the above formula include: low-dimensional features, with a mean of 0 and a variance of 1 for each dimension; therefore:

[0077] The positions of the k nearest elements are determined by w. i Fill the elements in the specified positions with 0, and fill the other positions with 0. Thus, the previous expression can be written as:

[0078]

[0079] Then, combining the above equation with the constraints, we construct the Lagrange function:

[0080]

[0081] Differentiating the Lagrange function yields:

[0082] Y(I N×N -W) T (I N×N -W)=λY.

[0083] Since the minimum value is required, the eigenvectors corresponding to the d smallest eigenvalues ​​should be selected. The smallest eigenvalue is often close to 0. Therefore, in practical applications, the eigenvectors Y corresponding to the d smallest eigenvalues ​​other than the smallest eigenvalue are selected as the data after dimensionality reduction. For example, elements in Y that are greater than 0 are grouped into one category, and elements that are less than 0 are grouped into another category. By projecting the elements onto low-level feature points, the clustering of feature elements can be completed.

[0084] In the above clustering process, the clustering of evaluation elements is completed based on the guidance of traffic flow characteristics. If the convolutional neural network extracts that the distance to the vehicle in front (traffic flow characteristic) is too close, the chassis domain extracts the elements related to emergency braking (TTC) under emergency braking conditions as evaluation elements. If the cluster boundaries are blurred and cannot be extracted, the edges in the data graph are removed to optimize the data graph. Clustering is then performed based on the optimized data graph until extraction is successful.

[0085] The present invention has been described above by way of example with reference to the accompanying drawings. Obviously, the specific implementation of the present invention is not limited to the above-described manner. Any non-substantial improvements made using the inventive concept and technical solution of the present invention, or the direct application of the inventive concept and technical solution of the present invention to other occasions without modification, are all within the protection scope of the present invention.

Claims

1. A layered decoupling method for urban autonomous driving scenarios, characterized by: Includes the following steps: Step 1: Unify the static scene elements in the test scenario to the coordinate system of the transportation infrastructure, then divide the test environment and use the probabilistic occupancy method to obtain a grid map. Step 2: Unify the data from the vehicle's multiple sensors into the vehicle coordinate system, and then use a Kalman filter to predict and build a time series diagram; Step 3: Based on the corrected and updated time series graph, construct an attention model using a convolutional neural network and extract traffic flow features related to autonomous driving actions; Step 4: The vehicle chassis adopts an integrated multivariable control nonlinear system. The nonlinear system is decoupled by a decoupling pre-compensator, and then the decoupled feature elements are output. Step 5: Based on the modified decoupling pre-compensator, decouple and obtain the feature characteristics to complete the clustering of evaluation elements, thereby extracting evaluation elements related to traffic flow characteristics; The method for establishing a time series diagram includes: first, collecting sensor data from vehicle-mounted sensors and performing noise reduction and normalization preprocessing; then, using a transformation matrix to unify the sensor data into the vehicle coordinate system based on the positional relationship between sensor coordinates and vehicle coordinates; constructing a prediction model and a measurement model for the time series diagram, calculating the prior estimate covariance, updating the current time-time estimate based on the measured value and the prior estimate, and then calculating and updating the posterior estimate covariance; adjusting the process noise based on the accuracy of the motion model and adjusting the observation noise based on sensor performance; setting the initial estimate to 0 and the initial covariance to 1; comparing the predicted data of the time series diagram with the actual data in the raster map, and manually adjusting the observation noise and process noise multiple times to achieve the required accuracy in the prediction results; then repeating the prediction and update steps to obtain new state estimates and covariance matrices, thereby drawing time series diagrams for various urban scenarios and situations; In step three, the image is input into the first convolutional layer of the attention model to form a feature map. This convolutional layer contains a set of learnable convolutional kernels to detect basic features in the image, including edges and corners. A max pooling layer is inserted after the convolutional layer to preserve the features of the feature map and simplify the output of the convolutional layer. After passing through a second convolutional pooling layer, the resulting pixel map is flattened and input into a fully connected layer to output traffic flow features. When the features do not meet the accuracy requirements, backpropagation is performed, and the optimizer updates the weights and biases of each channel of the weighted convolutional kernel. The attention model then repeats the process to output the extraction results until the extraction results meet the accuracy requirements. In step four, the same feature data is collected from the time series plot and compared with the decoupled feature data. When the root mean square error is greater than the error threshold, the decoupling pre-compensator is corrected.

2. The hierarchical decoupling method for urban autonomous driving scenarios according to claim 1, characterized in that: In step one, the probabilistic occupancy method for generating a grid map includes: collecting driving area information through laser ranging, rasterizing the driving area, using probability values ​​to represent the likelihood of each grid cell being occupied, setting a threshold for the probability of each grid cell being occupied, determining that the grid cell is occupied when the probability is not less than the threshold, and determining that the grid cell is not occupied when the probability is less than the threshold, and obtaining a grid map by determining whether each grid cell is occupied.

3. The hierarchical decoupling method for urban autonomous driving scenarios according to claim 1, characterized in that: In step two, when building the time series diagram, the predicted data of the time series diagram is compared with the actual data in the raster map. The observation noise and process noise are manually adjusted multiple times, and the Kalman filter is corrected and updated to make the time series diagram reach the required accuracy.

4. The hierarchical decoupling method for urban autonomous driving scenarios according to claim 3, characterized in that: In step four, the relationship between input and output variables is analyzed based on the physical structure, logical relationship, and application function of the autonomous vehicle. Based on the analysis results, the Interactor algorithm is used to determine the invertibility of the corresponding multivariable control nonlinear system and to determine the inverse system model. The original system is compensated into a pseudo-linear system with linear transitivity using the inverse system model, which is a decoupling pre-compensator.

5. The hierarchical decoupling method for urban autonomous driving scenarios according to claim 1, characterized in that: In step five, clustering of evaluation elements is completed based on traffic flow characteristics. If the cluster boundaries are blurred and cannot be extracted, the edges in the data graph are removed to optimize the data graph. Clustering is then performed again based on the optimized data graph until the evaluation elements related to traffic flow characteristics are successfully extracted.

6. The hierarchical decoupling method for urban autonomous driving scenarios according to claim 5, characterized in that: In step five, for the element features obtained above, each element feature is unified into a preset range to obtain the normalized value of each evaluation element; then, a data spectrum is determined based on the normalized value, and each data point in the data spectrum is the evaluation element corresponding to the normalized value. For each data point, find K nearest neighbors for that data point; Connect the data points with their neighboring points to form a local point cluster structure; calculate the weights between the data points; obtain the dimensionality-reduced feature vectors and feature values ​​based on the weights; obtain a low-dimensional matrix based on the feature values ​​and feature vectors to complete the clustering of feature elements.

Images