Humanoid intelligent driving track planning method based on RCSM-PL network

Abstract

The invention relates to the technical field of intelligent driving track prediction, in particular to a humanoid intelligent driving track planning method based on an RCSM-PL network. Comprising the following steps of 1, collecting driving control track data, 2, analyzing the driving control track data collected in the step 1, 3, extracting track data characteristics in a radar map processed in the step 2 and an original image, and 4, extracting track data characteristics in the radar map and the original image processed in the step 2. Step 5, according to the fact that future track points, electric quantity and dynamic characteristic parameters output by the compensation network obtained in the step 4 have relatively strong correlation in time, a multi-head attention mechanism is used for learning the space characteristics of the driving track extracted in the step 3 through linear transformation of different dimensions, so that the mapping capability of the network is enhanced; the LSTM network is used as a basic structure to capture and memorize a long-term dependency relationship in a time sequence.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent driving trajectory prediction, and in particular to a humanoid intelligent driving trajectory planning method based on a RCSM-PL network. Background Art

[0002] At present, the research on vehicle energy consumption is mostly focused on motion control, vehicle structure optimization, and environmental impact. There is less research on vehicle trajectory planning. Trajectory planning guides the vehicle's driving. Different driving trajectories will directly affect the vehicle's driving consumption. Therefore, it is very necessary to study the energy consumption of engineering vehicles from the perspective of trajectory planning and decision-making.

[0003] Commonly used trajectory planning methods include graph search, sampling, and artificial potential field methods. Graph search is greatly affected by the contours of the environment, and it is difficult to search for the optimal trajectory in the scene. The sampling method has a large randomness in the search process, and the search trajectory will lead to high energy consumption of the vehicle. When approaching obstacles, the artificial potential field method will have a large swing in the planned trajectory, so its application is limited. Roads in unstructured environments cannot present regular road shapes. The trajectories optimized by the planning methods proposed above will bring more redundant mileage burdens to vehicles. When vehicles are tracking, they will consume more energy. It is also difficult to extract features from the energy consumption aspect to design constraint functions. It is necessary to find a suitable method to plan trajectories in unstructured environments that reduce energy consumption when vehicles are tracking.

[0004] Therefore, a humanoid intelligent driving trajectory planning method based on RCSM-PL network is provided to solve the above problems. The energy consumption problem of the vehicle is studied from the perspective of trajectory planning. By adopting data-driven machine learning methods to learn human control behavior, a better way is planned to plan a trajectory that conforms to human control habits. The machine learning method avoids the process of manually designing constraint functions and can directly learn decision rules from control data. Human operators can adjust different control routes according to complex and changing environments, thereby controlling the vehicle to drive safely and smoothly. Summary of the invention

[0005] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a humanoid intelligent driving trajectory planning method based on a RCSM-PL network to solve the problems raised in the above-mentioned background technology.

[0006] To achieve the above object, the present invention provides the following technical solution: a humanoid intelligent driving trajectory planning method based on RCSM-PL network, comprising the following steps:

[0007] Step 1: Collect driving control trajectory data;

[0008] Step 2: Analyze the driving control trajectory data collected in step 1;

[0009] Step 3: Extract the trajectory data features from the radar image processed in step 2 and the original image;

[0010] Step 4: Use the multi-head attention mechanism to learn the spatial features of the driving trajectory extracted in step 3 through linear transformations of different dimensions;

[0011] Step 5: Based on the future trajectory points, power, and dynamic characteristic parameters output by the compensation network obtained in step 4, which have a strong correlation in time, the LSTM network is used as the basic structure to capture and memorize the long-term dependencies in the time series.

[0012] Preferably, the driving control trajectory data in the human control raw data acquisition module includes driving trajectory power parameters and driving space scene parameters, wherein the driving space scene parameters are radar coordinate data and driving area image pixel data.

[0013] Preferably, the analysis in the human manipulation data analysis module includes the following steps:

[0014] Step 2.1: The human control trajectory and power analysis unit analyzes the power parameters of the driving trajectory, selects 5 driving routes, and then accumulates the power consumption of the left and right wheels of the vehicle during driving under different driving routes. Based on the results of the power consumption accumulation, a cumulative power consumption diagram of the vehicle under the 5 routes is drawn, so as to select Figure 2 The trajectory data corresponding to the ID1 route in the figure is finally processed in groups for deep network training;

[0015] Step 2.2: The driving space scene processing and analysis unit analyzes the driving space scene parameters, including the following steps:

[0016] Step 2.2.1: Simulate human-controlled trajectory information, the coordinate transformation model is: where x l ,y l ,z l They represent the point cloud coordinates in the laser radar coordinate system, λ represents the scale factor, u i, v i represents the point cloud coordinates in the pixel coordinate system, M represents the camera intrinsic parameters, R and T represent the rotation matrix and offset vector between the camera and the lidar, respectively;

[0017] Step 2.2.2: Project the point cloud data collected by the LiDAR in the height direction and process the binary drivable area. The black pixels are the undrivable areas and the white pixels are the drivable areas. Then divide the feasible area into 8 parts. Calculate the percentage of the undrivable area in each part to the total area, and draw a radar map with the proportion of each part as an independent dimension. The calculation model of the proportion of the undrivable area is: where n i Indicates the total number of pixels in different areas, n b It represents the number of black pixels in the corresponding area, and p represents the proportion of black pixels.

[0018] Preferably, the spatiotemporal sequence trajectory planning feature extraction module adopts the residual structure in the resnet50 network, and the residual model is: y=R(x,{W i})+x, where x and y represent the input and output of the residual block, R(·) represents the residual mapping to be learned, and W i represents the parameter to be learned, and R(·)+x represents the sum of the corresponding positions.

[0019] Preferably, in the spatiotemporal sequence trajectory planning feature extraction module, channel attention is introduced in the residual block, useful feature channels are focused on, other channels are ignored, and the spatial information of the input feature map F is aggregated through maximum pooling and average pooling to generate two different types of features: maximum pooling features and average pooling features; then, the two different types of features are transmitted to a weight sharing network composed of a multi-layer perceptron with a hidden layer, and then further maximum pooling and average pooling and a nonlinear activation unit are performed to generate a channel attention map M c (F), the generated M c (F) and the input feature map F are combined through element-wise multiplication to obtain F'. The calculation model is: c (F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F))), Where σ represents the activation function, Represents element-wise multiplication.

[0020] Preferably, in the spatiotemporal sequence trajectory planning feature extraction module, F' is input into the spatial attention, the local area of ​​interest is focused on, the maximum pooling and average pooling operations are performed along the channel dimension to obtain the prominent information area, and the spatial attention map M is obtained after merging and convolution by the convolution layer with a convolution kernel size of (3×3). s (F'), then, through the spatial feature map M sMultiply the elements between (F') and F' to generate the refined feature map F", and the calculation model is:

[0021] M s (F')=σ(f([AvgPool(F');MaxPool(F')])), Where σ represents the sigmoid function and f represents the convolution operation with the filter size.

[0022] Preferably, the spatiotemporal sequence trajectory planning feature learning module uses a multi-head attention mechanism to learn the extracted driving trajectory spatial features through linear transformations of different dimensions, and adds an error compensation layer. The calculation model is: in d n ,d k ,d v Representing different dimensions, the error compensation formula is: where t 0 The first sequence point of the true value, p 0 is the first sequence point predicted, Δa is the difference between the true value and the predicted point in the first sequence point, w i and b i is a trainable parameter, p i is the prediction point, p l is the predicted point after compensation.

[0023] Preferably, the spatiotemporal sequence trajectory planning feature parameter dependency analysis module uses an LSTM network as the basic structure to capture and memorize long-term dependencies in the time series. The specific calculation model is as follows:

[0024] i t =σ(W ic ⊙c t-1 +W ih h t-1 +W ix x t +b i ),

[0025] f t =σ(W ft ⊙c t-1 +W fh h t-1 +W fx x t +b f ),

[0026] c t =f t ⊙c t-1 +i t ⊙tanh(Wch h t-1 +W cx x t +b c ),

[0027] o t =σ(W oc ⊙c t +W oh h t-1 +W ox x t +b o ),

[0028] h t =o t ⊙tanh(c t ),

[0029] where i t ,f t ,o t They are input gate, forget gate, and output gate; c t is the state cell; W uv (u,v∈{x,i,f,c,h,o}) represents the weight matrix; b u (u∈{i,f,c,o}) is the bias vector; ⊙ is the element-wise product; σ(x)=1 / (1+e -x ) is the sigmoid activation function; h t Output for short-term memory.

[0030] Technical effects and advantages of the present invention:

[0031] 1. The present invention divides the drivable area by elevation data, calculates the danger level of different parts of the drivable area, constructs a radar map with the danger level of each part as an independent dimension, and inputs the radar map into the deep network together with the original image as auxiliary information of the driving scene to generate human-like control trajectory information, and filters the data by the energy consumption level of different routes, thereby constructing a human control behavior data set;

[0032] 2. The present invention uses the vehicle dynamics characteristic parameters and the power consumption during driving as constraints, introduces a residual structure, thereby avoiding the occurrence of network degradation, and integrates the attention mechanism from different dimensions into the feature extraction network to focus on useful feature channels and local areas of interest, which helps to better extract features from the driving trajectory space;

[0033] 3. The present invention takes the future trajectory points, electric quantity and dynamic characteristic parameters output by the compensation network as the basis, and has a strong correlation in time. It uses the LSTM network as the basic structure to capture and memorize the long-term dependencies in the time series, and controls the output of the information flow by introducing special storage units and gating mechanisms, thereby improving the planning ability for future long sequence trajectory points. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a schematic diagram of the overall process of the present invention.

[0035] Figure 2 This is a diagram of the cumulative power consumption of vehicles on different routes. DETAILED DESCRIPTION

[0036] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0037] See also Figure 1 As shown, the present invention provides a humanoid intelligent driving trajectory planning method based on RCSM-PL network, including: a human control raw data acquisition module, a human control data analysis module, a spatiotemporal sequence trajectory planning feature extraction module, a spatiotemporal sequence trajectory planning feature learning module, and a spatiotemporal sequence trajectory planning feature parameter dependency analysis module.

[0038] The human manipulation raw data acquisition module is connected to the human manipulation data analysis module, the spatiotemporal sequence trajectory planning feature extraction module is respectively connected to the spatiotemporal sequence trajectory planning feature learning module and the human manipulation data analysis module, and the spatiotemporal sequence trajectory planning feature learning module is respectively connected to the spatiotemporal sequence trajectory planning feature parameter dependency analysis module and the spatiotemporal sequence trajectory planning feature extraction module.

[0039] The human control raw data acquisition module is used to collect the driving control trajectory data of the experimenter controlling the vehicle to avoid obstacles according to his own control habits, and transmit the driving control trajectory data to the human control data analysis module;

[0040] The human control data analysis module is used to analyze the control trajectory data, obtain the humanoid intelligent driving trajectory power analysis results and the driving space scene processing and analysis results, and transmit the analysis results to the spatiotemporal sequence trajectory planning feature extraction module;

[0041] The spatiotemporal sequence trajectory planning feature extraction module is used to extract features according to the analysis results processed in the data analysis module, and transmit the feature extraction results to the spatiotemporal sequence trajectory planning feature learning module;

[0042] The spatiotemporal sequence trajectory planning feature learning module is used to learn according to the features extracted by the feature extraction module, and transmit the learning results to the spatiotemporal sequence trajectory planning feature parameter dependency relationship analysis module;

[0043] The spatiotemporal sequence trajectory planning feature parameter dependency analysis module is used to perform feature parameter dependency analysis based on the results learned by the above feature learning module.

[0044] A humanoid intelligent driving trajectory planning method based on RCSM-PL network includes the following steps: step 1: collecting driving control trajectory data, step 2: analyzing the driving control trajectory data collected in step 1, step 3: extracting trajectory data features in the radar map processed in step 2 and the original image, step 4: using a multi-head attention mechanism to learn the driving trajectory spatial features extracted in step 3 through linear transformations of different dimensions, thereby enhancing the mapping ability of the network, and step 5: based on the future trajectory points, power and dynamic characteristic parameters output by the compensation network obtained in step 4, which have a strong correlation in time, an LSTM network is used as the basic structure to capture and memorize long-term dependencies in time series.

[0045] Step 1: The tester makes a decision based on the scene displayed on the screen, and then controls the vehicle to avoid obstacles according to his own control habits, while collecting driving control trajectory data. The driving control trajectory data includes driving trajectory power parameters and driving space scene parameters, where the driving space scene parameters are radar coordinate data and driving area image pixel data;

[0046] What needs to be specifically explained in this embodiment is that this study invited multiple participants to conduct the test. They have been engaged in the operation of engineering equipment and have more than thousands of hours of operation experience in the field. They can operate the equipment smoothly, and each experimenter is informed of the data collection process and how to use the sensor equipment.

[0047] Step 2: Analyze the driving control trajectory data collected in step 1, including a human control trajectory and power analysis unit and a driving space scene processing and analysis unit, wherein the human control trajectory and power analysis unit is used to analyze the driving trajectory power parameters and the driving space scene processing and analysis unit is used to analyze the driving space scene parameters, and the analysis includes the following steps;

[0048] See also Figure 2As shown, step 2.1: the human control trajectory and power analysis unit analyzes the power parameters of the driving trajectory. First, in a scene with densely distributed obstacles, 5 driving routes are selected, and then the power consumption of the left and right wheels of the vehicle during driving under different driving routes is accumulated. According to the results of the accumulated power consumption, a cumulative power consumption diagram of the vehicle under the 5 routes is drawn. It is concluded that as the vehicle mileage increases, greater energy consumption will occur. Therefore, according to the power loss situation, the collected data is screened to select Figure 2 The trajectory data corresponding to the ID1 route in the figure is grouped and processed for deep network training.

[0049] Step 2.2: The driving space scene processing and analysis unit analyzes the driving space scene parameters, including the following steps:

[0050] Step 2.2.1: When the tester collects data, he makes driving decisions based on the scene displayed on the screen. However, the laser radar can collect elevation point cloud information, which can be used to distinguish whether it is an obstacle. Therefore, the drivable area is segmented through elevation data, and the danger level of different parts of the drivable area is calculated. The radar map is constructed with the danger level of each part as an independent dimension, which is then input into the deep network together with the original image as auxiliary information of the driving scene to generate human-like control trajectory information. The coordinate transformation model is: where x l ,y l ,z l They represent the point cloud coordinates in the laser radar coordinate system, λ represents the scale factor, u i ,v i represents the point cloud coordinates in the pixel coordinate system, M represents the camera intrinsic parameters, R and T represent the rotation matrix and offset vector between the camera and the lidar, respectively;

[0051] Step 2.2.2: Project the point cloud data collected by the LiDAR in the height direction and process the binary drivable area. The black pixels are the undrivable areas and the white pixels are the drivable areas. Then divide the feasible area into 8 parts. Calculate the percentage of the undrivable area in each part to the total area, and draw a radar map with the proportion of each part as an independent dimension. The calculation model of the proportion of the undrivable area is: where n i Indicates the total number of pixels in different areas, n b represents the number of black pixels in the corresponding area, and p represents the proportion of black pixels;

[0052] What needs to be specifically explained in this embodiment is that the spatial information, trajectory information and power information processed above are grouped for use in the subsequent improvement of the deep network to perform trajectory mapping.

[0053] Step 3: Extracting the trajectory data features from the radar image processed in step 2 and the original image, including the following steps:

[0054] Step 3.1: Based on the time series relationship between the driving space coordinates and the trajectory points, a multi-network fusion method is used to plan the future trajectory points. First, the improved CNN network is used to extract spatial features from the driving scene. Then, the improved LSTM network is used to plan the future trajectory route with less energy consumption using the vehicle dynamics characteristic parameters and power information as constraints.

[0055] Step 3.2: In order to better extract features from the driving trajectory space, the residual structure in the resnet50 network is used to ensure that the L-th layer network contains more feature information than the L-1-th layer network, thereby avoiding the occurrence of network degradation. The residual model is: y = R(x, {W i})+x, where x and y represent the input and output of the residual block, R(·) represents the residual mapping to be learned, and W i represents the parameter to be learned, and R(·)+x represents the sum of the corresponding positions;

[0056] Step 3.3: Introduce channel attention in the residual block, pay attention to useful feature channels, ignore other channels, aggregate the spatial information of the input feature map F through maximum pooling and average pooling, and generate two different types of features: maximum pooling features and average pooling features; then, transmit the two different types of features to a weight-sharing network composed of a multi-layer perceptron with one hidden layer, and then generate a channel attention map M after further maximum pooling and average pooling and a nonlinear activation unit. c (F), the generated M c (F) and the input feature map F are combined through element-wise multiplication to obtain F'. The calculation model is: c (F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F))), Where σ represents the activation function, Represents element-wise multiplication;

[0057] Step 3.4: Input F' into the spatial attention, focus on the local area of ​​interest, perform maximum pooling and average pooling operations along the channel dimension to obtain the prominent information area, and then merge and convolve through the convolution layer with a convolution kernel size of (3×3) to obtain the spatial attention map M. s (F'), then, through the spatial feature map M sMultiply the elements between (F') and F' to generate the refined feature map F", and the calculation model is: M s (F')=σ(f([AvgPool(F');MaxPool(F')])), Where σ represents the sigmoid function and f represents the convolution operation with the filter size;

[0058] Step 4: Use the multi-head attention mechanism to learn the spatial features of the driving trajectory extracted in step 3 through linear transformations of different dimensions, thereby enhancing the mapping ability of the network, and add an error compensation layer to compensate for excessive mapping errors. The calculation model is: in d n ,d k ,d v Representing different dimensions, the error compensation formula is: where t 0 The first sequence point of the true value, p 0 is the first sequence point predicted, Δa is the difference between the true value and the predicted point in the first sequence point, w i and b i is a trainable parameter, p i is the prediction point, p l is the predicted point after compensation;

[0059] Step 5: Based on the future trajectory points, power and dynamic characteristic parameters output by the compensation network obtained in step 4, which have strong correlation in time, the LSTM network is used as the basic structure to capture and memorize the long-term dependencies in the time series. The LSTM network controls the output of the information flow by introducing special storage units and gating mechanisms. In order to increase the accuracy of extracting input information, the state cell information of the previous layer is considered when making decisions at each gate layer, thereby planning the trajectory points of the next 14 time steps. The specific calculation model is as follows:

[0060] i t =σ(W ic ⊙c t-1 +W ih h t-1 +W ix x t +b i ),

[0061] f t =σ(W ft ⊙c t-1 +W fh h t-1 +W fx x t +b f ),

[0062] c t =f t ⊙c t-1 +i t ⊙tanh(W ch h t-1 +W cx x t +b c ),

[0063] o t =σ(W oc ⊙c t +W oh h t-1 +W ox x t +b o ),

[0064] h t =o t ⊙tanh(c t ),

[0065] where i t ,f t ,o t They are input gate, forget gate, and output gate; c t is the state cell;

[0066] W uv (u,v∈{x,i,f,c,h,o}) represents the weight matrix; b u (u∈{i,f,c,o}) is the bias vector; | is the element-wise product;

[0067] σ(x)=1 / (1+e -x ) is the sigmoid activation function; h t Output for short-term memory.

[0068] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

[0069] The above contents are merely examples and explanations of the concept of the present invention. The technicians in this technical field may make various modifications or additions to the specific embodiments described or replace them in a similar manner. As long as they do not deviate from the concept of the invention or exceed the scope defined by the claims, they should all fall within the protection scope of the present invention.

Claims

1. A humanoid intelligent driving trajectory planning method based on RCSM-PL network, characterized in that: The following steps are involved: Step 1: Collect driving control trajectory data; Step 2: Analyze the driving control trajectory data collected in step 1; Step 3: Extract the trajectory data features from the radar image processed in step 2 and the original image; Step 4: Use the multi-head attention mechanism to learn the spatial features of the driving trajectory extracted in step 3 through linear transformations of different dimensions; Step 5: Based on the future trajectory points, power, and dynamic characteristic parameters output by the compensation network obtained in step 4, which have a strong correlation in time, the LSTM network is used as the basic structure to capture and memorize the long-term dependencies in the time series.

2. The humanoid intelligent driving trajectory planning method based on RCSM-PL network according to claim 1 is characterized by: The driving control trajectory data in the human control raw data acquisition module includes driving trajectory power parameters and driving space scene parameters, wherein the driving space scene parameters are radar coordinate data and driving area image pixel data.

3. The humanoid intelligent driving trajectory planning method based on RCSM-PL network according to claim 1 is characterized by: The analysis in the human manipulation data analysis module includes the following steps: Step 2.1: The human control trajectory and power analysis unit analyzes the power parameters of the driving trajectory, selects 5 driving routes, and then accumulates the power consumption of the left and right wheels of the vehicle during driving under different driving routes. According to the accumulated power consumption results, a cumulative power consumption diagram of the vehicle under the 5 routes is drawn, so as to select the trajectory data corresponding to the ID1 route in Figure 2, and finally the collected data is grouped and processed for deep network training; Step 2.2: The driving space scene processing and analysis unit analyzes the driving space scene parameters, including the following steps: Step 2.2.1: Simulate human-controlled trajectory information, the coordinate transformation model is: where x l ,y l ,z l They represent the point cloud coordinates in the laser radar coordinate system, λ represents the scale factor, u i ,v i represents the point cloud coordinates in the pixel coordinate system, M represents the camera intrinsic parameters, R and T represent the rotation matrix and offset vector between the camera and the lidar, respectively; Step 2.2.2: Project the point cloud data collected by the LiDAR in the height direction and process the binary drivable area. The black pixels are the undrivable areas and the white pixels are the drivable areas. Then divide the feasible area into 8 parts. Calculate the percentage of the undrivable area in each part to the total area, and draw a radar map with the proportion of each part as an independent dimension. The calculation model of the proportion of the undrivable area is: where n i Indicates the total number of pixels in different areas, n b It represents the number of black pixels in the corresponding area, and p represents the proportion of black pixels.

4. The humanoid intelligent driving trajectory planning method based on RCSM-PL network according to claim 1 is characterized by: The spatial-temporal sequence trajectory planning feature extraction module adopts the residual structure in the resnet50 network, and the residual model is: y=R(x,{W i })+x, where x and y represent the input and output of the residual block, R(·) represents the residual mapping to be learned, and W i represents the parameter to be learned, and R(·)+x represents the sum of the corresponding positions.

5. The humanoid intelligent driving trajectory planning method based on RCSM-PL network according to claim 1 is characterized by: In the spatiotemporal sequence trajectory planning feature extraction module, channel attention is introduced in the residual block to focus on useful feature channels and ignore other channels. The spatial information of the input feature map F is aggregated through maximum pooling and average pooling to generate two different types of features: maximum pooling features and average pooling features. Then, the two different types of features are transmitted to a weight sharing network composed of a multi-layer perceptron with a hidden layer, and then further maximum pooling and average pooling and a nonlinear activation unit are performed to generate a channel attention map M. c (F), the generated M c (F) and the input feature map F are combined through element-wise multiplication to obtain F'. The calculation model is: c (F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F))), Where σ represents the activation function, Represents element-wise multiplication.

6. The humanoid intelligent driving trajectory planning method based on RCSM-PL network according to claim 1 is characterized by: In the spatiotemporal sequence trajectory planning feature extraction module, F' is input into the spatial attention, the local area of ​​interest is focused on, and the maximum pooling and average pooling operations are performed along the channel dimension to obtain the prominent information area. After merging and convolution by the convolution layer with a convolution kernel size of (3×3), the spatial attention map M is obtained. s (F'), then, through the spatial feature map M s Multiply the elements between (F') and F' to generate the refined feature map F", and the calculation model is: M s (F')=σ(f([AvgPool(F');MaxPool(F')])), Where σ represents the sigmoid function and f represents the convolution operation with the filter size.

7. The humanoid intelligent driving trajectory planning method based on RCSM-PL network according to claim 1 is characterized by: In the spatiotemporal sequence trajectory planning feature learning module, a multi-head attention mechanism is used to learn the extracted driving trajectory spatial features through linear transformations of different dimensions, and an error compensation layer is added. The calculation model is: in d n ,d k ,d v Representing different dimensions, the error compensation formula is: where t 0 The first sequence point of the true value, p 0 is the first sequence point predicted, Δa is the difference between the true value and the predicted point in the first sequence point, w i and b i is a trainable parameter, p i is the prediction point, p l is the predicted point after compensation.

8. The humanoid intelligent driving trajectory planning method based on RCSM-PL network according to claim 1 is characterized by: The feature parameter dependency analysis module of the spatiotemporal sequence trajectory planning features uses the LSTM network as the basic structure to capture and memorize the long-term dependencies in the time series. The specific calculation model is as follows: i t =σ(W ic ⊙c t-1 +W ih h t-1 +W ix x t +b i ), f t =σ(W ft ⊙c t-1 +W fh h t-1 +W fx x t +b f ), c t =f t ⊙c t-1 +i t ⊙tanh(W ch h t-1 +W cx x t +b c ), o t =σ(W oc ⊙c t +W oh h t-1 +W ox x t +b o ), h t =o t ⊙tanh(c t ), where i t ,f t ,o t They are input gate, forget gate, and output gate; c t is the state cell; W uv (u,v∈{x,i,f,c,h,o}) represents the weight matrix; b u (u∈{i,f,c,o}) is the bias vector; ⊙ is the element-wise product; σ(x)=1 / (1+e -x ) is the sigmoid activation function; h t Output for short-term memory.

Images