Personalized end-to-end automatic driving control method considering spatio-temporal context

Abstract

The invention discloses a personalized end-to-end automatic driving control method considering spatio-temporal context, and belongs to the technical field of intelligent driving. The method comprises the following steps: firstly, for a continuous multi-frame RGB image sequence and vehicle state information, carrying out multi-modal spatial-temporal feature extraction based on a convolutional neural network and a long-short-term memory network to realize efficient fusion of vision and state information; meanwhile, the driving preference characteristics of the user are efficiently coded and integrated by adopting an embedded representation method; further, by sharing the spatio-temporal context features, a personalized trajectory prediction network and a security constraint control network based on a hybrid expert network are established and are respectively used for generating trajectories of different driving styles and performing security evaluation on the trajectories; and finally, constructing a multi-task deep learning framework, and performing supervised learning in combination with expert trajectory data, so that the model can directly output a vehicle control instruction after inputting multi-modal data and a driving style label, thereby realizing personalized end-to-end automatic driving control on the premise of ensuring safety.

Description

Technical Field

[0001] This invention relates to the field of intelligent driving technology, specifically to a personalized end-to-end autonomous driving control method that takes into account spatiotemporal context. Background Technology

[0002] As autonomous driving technology continues to mature and vehicles become increasingly intelligent, standardized driving models and personalized driving styles will coexist in the road environment for a considerable period of time, forming complex hybrid driving scenarios.

[0003] Traditional autonomous driving control methods are mostly based on modular architectures or end-to-end neural networks with unified behavioral assumptions. The former suffers from error propagation and response delay issues, while the latter trains a single model with mixed data from different driving styles, lacking explicit modeling of heterogeneous driving preferences such as aggressive, stable, and conservative driving styles. This makes it difficult to meet the real-time, efficient, and personalized trajectory prediction and control requirements in complex environments. Furthermore, existing safety constraint methods use fixed thresholds, which cannot adaptively adjust according to driving style and scenario complexity, making it difficult to dynamically balance personalization and safety, thus limiting the accuracy, safety assurance capabilities, and generalization efficiency of end-to-end systems. Especially in the field of electric vehicles, their fast power response, high precision of drive-by-wire execution, and flexible energy management provide a superior platform for personalized autonomous driving control. The vehicle motor, electric drive system, and battery management system onboard electric vehicles also offer high dynamic response capabilities and potential for energy efficiency optimization in the execution of end-to-end control commands. Summary of the Invention

[0004] To address the aforementioned issues, this invention provides a personalized end-to-end autonomous driving control method that considers spatiotemporal context. By introducing explicit driving style embedding representations and a hybrid expert network architecture, this invention achieves accurate modeling and end-to-end control of users' personalized driving preferences. Through the fusion of spatiotemporal context information and a hierarchical dynamic safety constraint mechanism, a complete personalized autonomous driving framework is constructed, encompassing environmental perception and control command generation. This method can generate trajectories and control strategies that conform to the user's driving style while ensuring driving safety. It effectively solves the technical challenges of insufficient personalization, inadequate utilization of spatiotemporal information, and the difficulty in balancing safety and personalization in existing end-to-end autonomous driving systems, significantly improving the user experience and acceptance of autonomous driving systems.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a personalized end-to-end autonomous driving control method considering spatiotemporal context, specifically including the following steps: S1. Construct a personalized end-to-end autonomous driving control system architecture that takes into account the spatiotemporal context; The construction methods include: designing an end-to-end deep neural network architecture, including an input layer, a feature extraction layer, a personalized modeling layer, a trajectory prediction layer, a safety constraint layer, and an output layer; The input structure of the input layer includes visual input, status input, and style input.

[0006] S2, Multimodal Data Perception and Spatiotemporal Feature Extraction Module: Based on the input structure of the input layer, visual feature sequences are extracted using ResNet-50, temporal dependencies are learned using a bidirectional LSTM temporal feature encoder, and temporal feature vectors are output. A multilayer perceptron (MLP) is used to encode state information, and after outputting state feature vectors, the temporal feature vectors and state feature vectors are concatenated to perform feature fusion, outputting spatiotemporal context features. This invention constructs a multi-frame image perception and spatiotemporal feature extraction module based on the input structure in step S1, taking as input a continuous T-frame RGB image sequence and a vehicle state vector (position coordinates). , ,speed acceleration Heading angle Lane information The system extracts visual features and user driving style labels (aggressive, stable, conservative) using ResNet-50, learns temporal dependencies using bidirectional LSTM, encodes state information using multilayer perceptron, and fuses them to form a 640-dimensional spatiotemporal context feature representation.

[0007] S3. Personalized Driving Style Modeling and Expert Hybrid Network Architecture: Based on spatiotemporal context features and style input, personalized enhanced context feature vectors are obtained through an embedding layer. Then, three types of expert networks—aggressive, stable, and conservative—are constructed based on an expert knowledge base. An expert data fusion mechanism is designed to dynamically combine the personalized context feature vectors and the expert knowledge base through feature enhancement formulas to generate differentiated expert-specific enhanced features. A gating network is introduced to calculate gating weights. Based on the original trajectory vectors output by the three types of experts and the gating weights, personalized predicted trajectories are obtained through trajectory fusion and trajectory optimization.

[0008] S4. Construct a hierarchical dynamic safety constraint and predictive intervention mechanism; construct the overall control architecture through personalized trajectory control mapping, perform PID trajectory tracking control calculation, conduct multimodal risk assessment, construct hierarchical dynamic safety constraints and three-layer safety assurance settings, and output vehicle control commands and system status.

[0009] S5. Design an end-to-end multi-task joint training strategy; by defining a multi-task loss function and security constraints, output training data stream storage to achieve offline batch training and online incremental learning.

[0010] S6. Construct an online deployment and real-time inference framework for personalized autonomous driving models, save the trained model parameters, deploy the end-to-end prediction and control models to the vehicle computing platform, and complete the construction of a personalized end-to-end autonomous driving control system that considers spatiotemporal context.

[0011] As a preferred embodiment of the present invention, step S1 includes the following steps: S1.1 Define the input structure of the input layer: including visual input, state input and style input, and construct the input fusion formula; S1.2 Define the overall architecture of the end-to-end deep neural network: A multi-layer neural network architecture combined with safety constraints is used for the final vehicle command output; The multi-layer neural network architecture includes: ResNet-50 as the visual backbone, bidirectional LSTM for temporal processing, MLP for state encoding, embedding layer for driving style embedding, hybrid expert network (MoE) for personalized trajectory generation, and PID controller combined with safety constraints for final vehicle command output.

[0012] As a preferred embodiment of the present invention, step S2 includes the following steps: S2.1 Data Preprocessing: Based on the input structure of the input layer, preprocessing operations are performed through the data flow pipeline to output the preprocessed image sequence and the preprocessed vehicle state vector; S2.2 Visual Feature Extraction: Input the preprocessed image sequence and extract visual features through a pre-trained ResNet-50 convolutional neural network to obtain a visual feature sequence; The method of pre-training the ResNet-50 convolutional neural network is as follows: the ResNet-50 convolutional neural network is pre-trained on the ImageNet dataset, the fully connected classification layer is removed, the convolutional feature extraction part and the global average pooling layer are retained, and a 1024-dimensional feature vector is output. S2.3 Temporal Feature Encoding: Based on the visual feature sequence and the preprocessed vehicle state vector, the temporal dependency is learned by fusing the features with a bidirectional LSTM temporal feature encoder to obtain the temporal feature vector; S2.4 State Information Encoding: Based on the preprocessed vehicle state vector, state information is encoded through a multilayer perceptron (MLP) to obtain a state feature vector; S2.5 Feature Fusion: Input temporal feature vectors and state feature vectors, and perform feature fusion by concatenating the vectors to obtain spatiotemporal context features.

[0013] As a preferred embodiment of the present invention, step S3 includes the following steps: S3.1 Driving Style Embedding and Adaptation: Based on driving style labels and spatiotemporal context features, personalized enhanced context feature vectors are obtained through driving style embedding and style adaptation. S3.2. Construct an expert knowledge base based on trajectory templates, driving modes, and decision preferences; S3.3 Hybrid Expert Network Reasoning: Based on the expert knowledge base and personalized enhanced context feature vectors, gating weights are obtained through feature enhancement mechanisms, forward propagation of hybrid expert networks, and gating networks. Specifically, the following steps are included: S3.3.1 Based on the expert knowledge base and personalized enhanced contextual feature vectors, enhanced input features are obtained through a feature enhancement mechanism; S3.3.2 Constructing three types of expert networks—radical, stable, and conservative—based on an expert knowledge base: Inputting enhanced input features, the original trajectory vectors of the three types of expert outputs are obtained through a hybrid expert network; The hybrid expert network consists of parallel radical expert networks, stationary expert networks, conservative expert networks, and a gated network. The original trajectory vectors output by the three types of experts include: the original trajectory vector output by radical experts, the original trajectory vector output by stationary experts, and the original trajectory vector output by conservative experts. S3.3.3 Based on the enhanced input features, the gating weights are obtained through a gating network; S3.4 Based on the original trajectory and gating weights output by three types of experts, the trajectory is optimized after weighted trajectory fusion to output a personalized predicted trajectory.

[0014] As a preferred embodiment of the present invention, step S4 includes the following steps: S4.1 Personalized trajectory control mapping: Based on personalized predicted trajectory, preprocessed vehicle state vector and driving style label, construct the overall control architecture; S4.2 Based on the overall control architecture, input the personalized predicted trajectory and vehicle state vector, and obtain the preliminary vehicle control command through motion parameter derivation and PID trajectory tracking controller calculation; S4.3 Constructing hierarchical dynamic safety constraints: Based on driving style labels and basic safety thresholds, personalized safety boundaries are obtained through personalized safety boundary adjustment strategies; S4.4 Multimodal Risk Assessment: Based on personalized predicted trajectory, vehicle state vector, environmental information and personalized safety boundary, a comprehensive risk assessment value is obtained through a multi-dimensional risk assessment model; The multi-dimensional risk assessment model includes: collision risk assessment, comfort risk assessment, traffic rule risk assessment, and pedestrian risk assessment. S4.5 Predictive safety intervention: Based on comprehensive risk assessment values ​​and personalized safety boundaries, safety intervention weights and minimum mandatory safety intervention weights are obtained through trajectory extrapolation prediction, safety intervention weight calculation, and hard constraint safety assurance mechanisms. S4.6, Layered Dynamic Safety Constraints: Based on safety intervention weights, the final vehicle control commands and system status are obtained through trajectory smoothing and trajectory fusion.

[0015] As a preferred embodiment of the present invention, step S5 includes the following steps: S5.1 End-to-end Prediction and Control; Based on the constructed personalized end-to-end prediction and control system, the optimized model parameters are obtained by defining the multi-task joint loss parameters and end-to-end forward propagation. S5.2 Calculation of the joint loss function for multiple tasks: After parallel calculation of five sub-loss terms, a weighted aggregation is performed to obtain the total loss function for multiple tasks; The five loss terms include: trajectory loss function, control loss function, safety loss function, style loss function, and expert differentiation loss function; S5.3, Adaptive Weight Adjustment: Based on the multi-task total loss function, backpropagation updates are performed through adaptive weight adjustment and learning rate scheduling strategies; S5.4 Training Data Organization and Incremental Learning: Through training data organization and annotation mechanisms, differentiated storage strategies, online learning triggering strategies, annotation quantity and quality control, and driving style label annotation, abnormal data is deleted to obtain training data with optimized model parameters.

[0016] Compared with existing technologies, this invention provides a personalized end-to-end autonomous driving control method that considers spatiotemporal context, and has the following beneficial effects: This invention accurately captures personalized behavioral patterns of different driving styles by constructing a spatiotemporal contextual feature system that integrates multi-frame images, vehicle status, and driving style, and designs a differentiated hybrid expert network architecture, achieving end-to-end precise mapping from environmental perception to control commands. Secondly, it introduces a feature distillation mechanism based on an expert knowledge base and hierarchical dynamic safety constraints to ensure driving safety while maintaining personalized driving styles. This achieves personalized customization of driving behavior while guaranteeing absolute safety boundaries through predictive safety intervention mechanisms. Finally, by using a multi-task joint loss function, the training efficiency and stability of the personalized expert network are optimized, and the control strategy can be flexibly adjusted according to different users' driving preferences, thereby effectively balancing driving personalization and driving safety, improving system practicality and user acceptance. Attached Figure Description

[0017] Figure 1A flowchart of the personalized end-to-end autonomous driving control method considering spatiotemporal context of the present invention; Figure 2 This is a schematic diagram of the personalized driving style modeling and hybrid expert network architecture of the present invention; Figure 3 This is a schematic diagram of the hierarchical dynamic security constraint and predictive intervention mechanism of the present invention; Figure 4 This is a schematic diagram of the training process of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figures 1-4 A personalized end-to-end autonomous driving control method considering spatiotemporal context, characterized by comprising the following steps: S1. Construct a personalized end-to-end autonomous driving control system architecture that takes into account the spatiotemporal context; The construction methods include: designing an end-to-end deep neural network architecture, including an input layer, a feature extraction layer, a personalized modeling layer, a trajectory prediction layer, a safety constraint layer, and an output layer; The input structure of the input layer includes visual input, status input, and style input; This invention constructs a personalized end-to-end autonomous driving control system that considers spatiotemporal context, including an input structure that includes multi-frame image perception, vehicle state information and user driving style labels, and a deep learning architecture that directly generates vehicle control commands from raw sensor data.

[0020] The overall system architecture of this invention adopts an end-to-end deep neural network architecture, including an input layer, a feature extraction layer, a personalized modeling layer, a trajectory prediction layer, a safety constraint layer, and an output layer. The input structure of the input layer includes multi-frame image perception (a continuous T-frame RGB image sequence), vehicle state information (real-time physical vectors), and user driving style labels (discrete encoding); the output of the output layer is the vehicle control commands (steering angle, throttle opening, braking intensity).

[0021] S1.1 Define the input structure of the input layer: including visual input, state input, and style input, and construct the overall input; This invention designs the input structure of the input layer, including the following steps: Visual input: A continuous sequence of T RGB images, each with a resolution of 224×224×3 pixels, represented as follows: ,in ; State input: Vehicle state vector ,in , Position coordinates (unit: ), Speed ​​(unit: ), Acceleration (unit: ), The heading angle (unit: radians). Lane information (encoded as integers 0-5, representing different lane types); Style Input: User's driving style tags The values ​​are encoded as discrete values: 0 (aggressive), 1 (stable), and 2 (conservative). Constructing the input fusion formula: The overall input is represented as Through subsequent layer processing, multimodal data fusion is achieved. This representation ensures that the system captures visual dynamics (image sequences), physical constraints (vehicle state vectors), and behavioral preferences (driving style labels), providing a complete information foundation for end-to-end control.

[0022] S1.2 Define the overall architecture of the end-to-end deep neural network: A multi-layer neural network architecture combined with safety constraints is used for the final vehicle command output; The multi-layer neural network architecture in this invention includes: ResNet-50 as the visual backbone, bidirectional LSTM for temporal processing, MLP for state encoding, embedding layer for driving style embedding, hybrid expert network (MoE) for personalized trajectory generation, and PID controller combined with safety constraints for final vehicle command output. The expression for a multi-layer neural network architecture is:

[0023] In the formula, For the feature extraction layer, Convert to a spatiotemporal context representation; For the personalization layer, integrate style embedding; This is the trajectory prediction layer, which generates predicted trajectories. For the safety constraint layer, apply layered boundaries; Generate instructions for the control layer.

[0024] S2, Multimodal Data Perception and Spatiotemporal Feature Extraction Module: Based on the input structure of the input layer, visual feature sequences are extracted through ResNet-50, temporal dependencies are learned through a bidirectional LSTM temporal feature encoder, and temporal feature vectors are output. State information is encoded through a multilayer perceptron (MLP), and state feature vectors are output. After outputting state feature vectors, the temporal feature vectors and state feature vectors are fused by vector concatenation to output spatiotemporal context features. This invention constructs a multi-frame image perception and spatiotemporal feature extraction module based on the input structure defined in step S1. The input consists of a continuous T-frame RGB image sequence, the current vehicle state vector (position coordinates, speed, acceleration, heading angle, lane information), and the user's driving style label (aggressive, moderate, conservative). Visual features are extracted using ResNet-50, temporal dependencies are learned using bidirectional LSTM, and state information is encoded using an MLP (Multilayer Perceptron). These features are then fused to form a 640-dimensional spatiotemporal context feature representation, providing a unified environmental perception foundation for subsequent personalized decision-making.

[0025] Specifically, the following steps are included: S2.1 Data Preprocessing: Based on the input structure of the input layer, preprocessing operations are performed through the data flow pipeline to output the preprocessed image sequence and the preprocessed vehicle state vector; This invention first collects and preprocesses input data to ensure data format standardization and noise filtering. The input data is then transmitted to the feature extraction layer through a data pipeline. A batch processing mechanism is used to achieve efficient parallel computing. After the input structure defined in step S1.2, the data is processed through the data stream pipeline.

[0026] In the formula, This is a preprocessing function, specifically defined as: The image is after preprocessing; This is the preprocessed current vehicle state vector; For the original image, perform pixel value normalization, i.e. This maps pixel values ​​to the [0,1] range. For the current vehicle state vector, perform Z-score standardization, i.e. ,in It is the mean vector. The standard deviation vector; For style tags, no numerical transformation will be performed for now; they will be passed directly. .

[0027] S2.2 Visual Feature Extraction: Input the preprocessed image sequence and extract visual features through a pre-trained ResNet-50 convolutional neural network to obtain a visual feature sequence; In this invention, a pre-trained ResNet-50 convolutional neural network is used as the visual feature extractor to encode deep features for each frame of the image. The ResNet-50 structure includes five stages ( to The ResNet-50 convolutional neural network, with a total of 50 layers, uses residual connections to avoid gradient vanishing. It is pre-trained on the ImageNet dataset, removing fully connected classification layers while retaining the convolutional feature extraction part and the global average pooling (GAP) layer, outputting a 1024-dimensional feature vector. (Preprocessed image sequence) Each frame is fed into the ResNet-50, and features are extracted independently from each frame to form a sequence. Then the entire sequence is transmitted to the timing feature encoder in S2.3.

[0028] Visual extraction formula: For each frame The calculation expression is as follows:

[0029] in, Represented as the forward propagation of ResNet-50, it can be decomposed into:

[0030] The residual block ResBlock is defined as:

[0031] In the formula, It is a transformation sequence consisting of convolution, batch normalization, and ReLU activation function; These are the input features for the residual block; The output characteristics of the residual block; This is the final extracted visual feature vector; For each frame of the image; The initial convolutional layer and downsampling layer (7×7 kernel, strude=2) are followed by a batch normalization layer and a linear rectified (ReLU) activation function. Immediately after the initial convolutional layer, a max pooling operation with a convolutional kernel size of 3×3 and a stride of 2 is used to perform a second downsampling on the feature map to increase the receptive field of subsequent layers and reduce computational complexity. It contains 3 bottleneck blocks. It contains 4. It contains 6. It contains 3, and each residual block uses a bottleneck structure (1×1 conv for dimensionality reduction, 3×3 conv for main computation, and 1×1 conv for dimensionality increase). Global average pooling is performed on the feature map output by the adaptive feature projection layer in its spatial dimensions (height and width), and all activation values ​​of each channel are aggregated into a single scalar, thereby compressing the feature map of arbitrary spatial size into a feature vector of fixed dimension. As an adaptive feature projection layer, this layer uses a 1×1 convolutional kernel, which optimizes the final feature representation and projects the number of feature channels to a preset target dimension.

[0032] The above process and (Low-level) Extract basic features such as edges and textures (capture local details, such as road lines); (High-level) Extract high-level semantics such as object categories (identify vehicle and pedestrian locations); The compressed activation map is a fixed-dimensional vector, ensuring that the output is independent of the input resolution. Residual connections. To ensure stable training of deep networks, gradient flow is preserved through identity mapping to avoid the vanishing gradient problem. This process is performed for each frame. Independently extracted feature vectors This constitutes a visual feature sequence containing complete spatial semantic information. .

[0033] S2.3 Temporal Feature Encoding: Based on the visual feature sequence and the preprocessed vehicle state vector, the temporal dependency is learned by fusing the features with a bidirectional LSTM temporal feature encoder to obtain the temporal feature vector; Visual feature sequences and the preprocessed vehicle state vector from step S2.1 By employing parallel encoding and feature fusion, a comprehensive spatiotemporal context feature representation is constructed, providing an information foundation for downstream personalized decision-making. The network structure uses a two-layer bidirectional long short-term memory network (BiLSTM) as the temporal encoder. This network, through its internal gating mechanism, can effectively capture long-range dependencies in sequential data. Input dimensions and visual feature vectors... The dimensionality is consistent, with a hidden layer dimension of 256 (each unidirectional LSTM), and two layers. The output of the first LSTM layer serves as the input to the second LSTM layer. Each unidirectional LSTM has a hidden layer dimension of 256, and the bidirectional LSTMs are concatenated to output a 512-dimensional array. 512 dimensions were chosen because this dimension provides sufficient capacity to capture complex spatiotemporal dependencies without over-parameterizing, which could lead to overfitting or high computational cost. The data is processed as a visual feature sequence. Input the data into the BiLSTM in chronological order, and then concatenate the second-layer bidirectional hidden states after the entire sequence has been processed to form a summary representation of the entire sequence.

[0034] The formula for temporal feature encoding is:

[0035] The forward propagation process can be broken down into the following steps: 1. Forward and backward propagation:

[0036] In the formula, In time step Hidden states of the forward LSTM; In time step The hidden state of the backward LSTM; In time step The input visual feature vector comes from the sequence ; For the forward LSTM at time step The hidden state, with the initial state being the zero vector; Backward LSTM at time step The hidden state is initially a zero vector.

[0037] 2. Two-way state splicing

[0038] In the formula, In time step A bidirectional hidden state formed by combining forward and backward hidden states.

[0039] 3. Sequence summary extraction:

[0040] In the formula, As the final temporal feature vector summarizing the entire sequence; After the entire sequence has been processed by the forward LSTM, at the final time step The hidden state; For backward LSTM from Reverse processing to time step The hidden state.

[0041] Detailed explanation of LSTM unit computation for any time step in the sequence The internal state update of its LSTM unit is as follows:

[0042] In the formula, Use the Sigmoid activation function; It is the hyperbolic tangent activation function; This is the element-wise multiplication of vectors (Hadamard product). , , , For the corresponding input The learnable weight matrix; , , , To correspond to the previous hidden state The learnable weight matrix; , , , It is a trainable bias vector; , , These are the activation vectors for the input gate, forget gate, and input gate, respectively. This represents the candidate cell state vector; This is the cell state vector updated at the current time step, with the initial state... It is a zero vector; This is the hidden state vector output at the current time step.

[0043] S2.4 State Information Encoding: Based on the preprocessed vehicle state vector, state information is encoded through a multilayer perceptron (MLP) to obtain a state feature vector; This invention employs a multilayer perceptron (MLP) to nonlinearly map low-dimensional vehicle state vectors to a high-dimensional semantic space, facilitating fusion with temporal features. The input layer consists of six neurons, corresponding to the state vectors. The dimension is [dimension missing], the hidden layer has 64 neurons using the ReLU activation function, and the output layer has 128 neurons using the ReLU activation function. The state information encoding formula is:

[0044] The forward propagation process can be broken down as follows:

[0045] In the formula, The input is a 6-dimensional vehicle state vector; This is the learnable weight matrix of the first fully connected layer; The learnable bias vector of the first fully connected layer; The linear rectified activation function is defined as follows: ; This is the output feature vector of the first fully connected layer after ReLU activation; This is the learnable weight matrix for the second fully connected layer; This is the learnable bias vector for the second fully connected layer; This results in a 128-dimensional state feature vector as the final output. This MLP implements a non-linear projection. The first layer expands the 6-dimensional physical input to 64 dimensions to capture basic interactions; the second layer further maps to 128 dimensions, enhancing its ability to encode complex states.

[0046] S2.5 Feature Fusion: Input temporal feature vector and state feature vector, and perform feature fusion by concatenating the vectors to obtain spatiotemporal context features; In this invention, the feature fusion method is to combine temporal feature vectors With state feature vector Vector concatenation is performed to form a comprehensive spatiotemporal context feature. The feature fusion formula is as follows:

[0047] In the formula, The fused output is a 640-dimensional comprehensive spatiotemporal context feature vector; This is the output vector from the temporal feature encoding module; Output vector from the state information encoding module; This is the vector concatenation operator.

[0048] This formula achieves direct aggregation of dynamic temporal information and static spatial information at the feature level. (Temporal part) Provides the foundation for motion prediction, state part This injects the current physical constraints. User driving style tags. This module will not participate in the integration at this time and will be directly transmitted to the S3 module for personalized embedding.

[0049] S3. Personalized Driving Style Modeling and Expert Hybrid Network Architecture: Based on spatiotemporal context features and style input, personalized enhanced context feature vectors are obtained through an embedding layer. Then, three types of expert networks—aggressive, stable, and conservative—are constructed based on an expert knowledge base. An expert data fusion mechanism is designed to dynamically combine the personalized context feature vectors and the expert knowledge base through feature enhancement formulas to generate differentiated expert-specific enhanced features. A gating network is introduced to calculate gating weights. Based on the original trajectory vectors output by the three types of experts and the gating weights, personalized predicted trajectories are obtained through trajectory fusion and trajectory optimization. like Figure 2As shown, this invention designs a personalized driving style modeling module and a hybrid expert network architecture, fusing the spatiotemporal context features from step S2 with the driving style embedding (64-dimensional vector) to form a 704-dimensional personalized enhanced context representation; the style vector is adjusted in real time through a driving style adaptive mechanism, and three types of expert networks—aggressive, stable, and conservative—are constructed based on an expert knowledge base (using shallow wide networks, standard networks, and deep progressive networks, respectively); an expert data fusion mechanism is designed to dynamically combine personalized context features with static expert knowledge through feature enhancement formulas, generating differentiated expert-specific enhanced features; a gating network is introduced to intelligently calculate the expert weight distribution based on the statistical representation of the enhanced features, achieving an effective balance between dynamic style adaptation and static expert knowledge, and generating a predicted trajectory (2-second span, 20 trajectory points) that conforms to the user's personalized driving style. Specifically, the following steps are included: S3.1 Driving Style Embedding and Adaptation: Based on driving style labels and spatiotemporal context features, personalized enhanced context feature vectors are obtained through driving style embedding and style adaptation. like Figure 2 As shown, due to differences in user driving styles in terms of attributes and preference dimensions (such as aggressive quick response and conservative smooth caution), a personalized style embedding and fusion encoder is designed to process different features of different style labels, transforming the abstract concept of driving style into a computable high-dimensional vector representation, and deeply fusing it with spatiotemporal context information to provide personalized guidance for trajectory prediction. Through driving style embedding and adaptive mechanisms, a trainable parameter embedding layer works in conjunction with a style classifier.

[0050] Discrete style labels are mapped to continuous style feature vectors to capture semantic relationships between styles. The driving style embedding formula is as follows:

[0051] In the formula, The input driving style label is 0, which represents aggressive driving (preferring high speed, frequent lane changes, and short following distance), 1 represents moderate driving (preferring medium speed, moderate following distance, and standard driving behavior), and 2 represents conservative driving (preferring low speed, long following distance, and cautious driving behavior). The embedding matrix is ​​a trainable parameter matrix, which is optimized through end-to-end training. This produces a 64-dimensional style embedding vector. Style similarity is then calculated using Euclidean distance within this vector space.

[0052] The expression for the quantitative indicator vector is:

[0053] In the formula, Quantification of speed preference (slow: 0, normal: 1, fast: 2); This is a quantified value for following distance (long distance: 0, medium distance: 1, short distance: 2). This is the quantization value for lane change frequency (conservative: 0, normal: 1, aggressive: 2).

[0054] Real-time style recognition and adaptation dynamically adjusts the user's style vector representation by analyzing recent driving behavior data to adapt to the user's real-time state. The expression for the style classifier is:

[0055] In the formula, Quantify the user's recent driving behavior into a data vector, including indicators such as speed preference, following distance, and lane change frequency; , These represent the weights and biases of the first fully connected layer of the style classifier. , The weights and biases of the second fully connected layer of the style classifier; This is the predicted current driving style label.

[0056] Style Adaptive Formula:

[0057] In the formula, Embed the initial tag; For real-time identification of embedded tags; The adaptive weight hyperparameter is set to 0.1 through cross-validation to ensure that the real-time adjustment is gentle and stable, and to control the adjustment range. This is the final style embedding vector after adaptive adjustment.

[0058] The adjusted style information is fused with environmental information to form a personalized enhanced context. The feature fusion formula is as follows:

[0059] In the formula, Input features from the spatiotemporal context building module; This is a vector concatenation operation; This outputs a 704-dimensional personalized enhanced contextual feature vector.

[0060] S3.2. Construct an expert knowledge base based on trajectory templates, driving modes, and decision preferences; To achieve rapid convergence and differentiated initialization of the expert sub-network, this invention constructs an expert knowledge base for each driving style based on a large amount of historical driving data of different styles. This knowledge base provides supervision signals to the feature encoder during the training phase and achieves knowledge transfer through a pre-trained feature projection matrix during the inference phase. Based on a hybrid expert network architecture, predicted trajectories that conform to users' personalized preferences are generated, and the expert knowledge base is constructed. The trajectory data set is as follows:

[0061] In the formula, For style Trajectory dataset, For the first The input feature matrix of each trajectory; For the first A label matrix for each trajectory; Label the behavior (acceleration, deceleration, turning, etc.); For style The number of samples.

[0062] The formula for extracting trajectory templates is:

[0063] In the formula, Principal component analysis reduces the dimensionality of the trajectory to dimension, For style The trajectory template feature vector has a dimension of 10.

[0064] Driving mode feature formula:

[0065] In the formula, , The average speed and the standard deviation of the speed; , The average acceleration and the standard deviation of acceleration; This refers to jerk statistics; Trajectory curvature characteristics; This is the driving mode feature vector, with a dimension of 6.

[0066] Formula for modeling decision preferences:

[0067] In the formula, This refers to the scene state; In the state The conditional probability of taking a certain action; This is a decision preference vector with 4 dimensions.

[0068] The feature encoder pre-constructs a feature space that maps input features to a knowledge base during offline training. Its training objective is to minimize the difference between the encoder output and the corresponding style features in the knowledge base. The feature encoding formula is as follows:

[0069] In the formula, For activation functions; For the first Each trajectory template feature vector; This is the weight matrix of the trajectory template encoder; This is the bias vector of the trajectory template encoder; Trajectory template features The 64-dimensional hidden feature vector obtained after encoding; For the first Each driving mode feature vector; This is the weight matrix of the driving mode encoder; This is the bias vector of the driving mode encoder; Driving mode features The 32-dimensional hidden feature vector obtained after encoding; For the first A decision preference feature vector; This is the weight matrix of the decision preference encoder; This is the bias vector of the decision preference encoder; Decision preference characteristics The resulting 16-dimensional hidden feature vector is obtained after encoding.

[0070] Expert knowledge base construction:

[0071] In the formula, For style Trajectory template feature vector, For driving mode feature vectors, This represents the decision preference vector.

[0072] S3.3 Hybrid Expert Network Reasoning: Based on the expert knowledge base and personalized enhanced context feature vectors, gating weights are obtained through feature enhancement mechanisms, forward propagation of hybrid expert networks, and gating networks. Specifically, the following steps are included: S3.3.1 Based on the expert knowledge base and personalized enhanced contextual feature vectors, enhanced input features are obtained through a feature enhancement mechanism; Based on personalized enhanced contextual feature vectors and a constructed expert knowledge base, expert data is fused with personalized enhanced contextual feature vectors. Through an expert data fusion mechanism, the formula for personalized contextual projection and real-time style feature generation is as follows:

[0073]

[0074] Feature enhancement formula:

[0075] In the formula, The trajectory feature projection matrix; The projection matrix is ​​the feature projection matrix for driving modes, which is obtained through supervised pre-training using a knowledge base. , For fusion weighting coefficients; For enhanced input features; Enhance contextual feature vectors for personalization.

[0076] S3.3.2 Constructing three types of expert networks—radical, stable, and conservative—based on an expert knowledge base: Inputting enhanced input features, the original trajectory vectors of the three types of expert outputs are obtained through a hybrid expert network; Because different driving styles differ in decision-making preferences and response characteristics, a hybrid expert network architecture is designed to differentiate between different types of experts, enabling them to uniformly output personalized predicted trajectories for subsequent control. Through a hybrid expert network, personalized contextual information is transformed into a vehicle motion trajectory within the next 2 seconds that conforms to the user's specific driving style. This trajectory consists of 20 consecutive trajectory points, each containing the vehicle's longitudinal and lateral positions. The overall architecture formula is:

[0077] In the formula, The forward propagation function representing the entire hybrid expert network; For enhanced input features, specifically designed for experts custom made; For personalized trajectory prediction matrices.

[0078] Hybrid expert networks consist of multiple parallel expert networks and a gating network. The core idea is "divide and conquer," where different experts learn different driving strategies, and the gating network intelligently combines the opinions of all experts based on the current situation. The three expert networks process the same enhanced input features in parallel. However, due to their different structures and objectives, different trajectory proposals are generated. The radical expert adopts a shallow, wide structure, and its LeakyReLU activation function allows negative gradients to flow, reducing network response time and making it more suitable for high-speed scenarios; the conservative expert adopts a deep, progressive structure, and its Tanh activation function has bounded outputs, prioritizing safety.

[0079] The aggressive expert network simulates aggressive driving behavior, generating trajectories that are fast-responding, efficient, and more aggressive. The aggressive expert network employs a shallow, wide + LeakyReLU structure, with a wide network width and large activation function gradients, facilitating rapid learning and response, aligning with its "efficiency-first" positioning. The forward propagation formula is:

[0080] In the formula, The weight matrix for the first layer is the 704-dimensional input. Transform to a 512-dimensional space; The input features are those that have been radically enhanced; This is the bias vector of the first layer, multiplied by the weight matrix and added to adjust the network output. As the activation function, it allows very small negative values ​​to pass through (with a slope of 0.01), thus enabling the network to respond quickly to changes in input; This is the hidden feature vector obtained after the first layer calculation. The weight matrix for the second layer maps the features from 512 dimensions to 256 dimensions; This is the hidden feature vector of the second layer; The weight matrix for the third layer (output layer) compresses the 256-dimensional features to 40 dimensions; The original trajectory vector output for the radical expert is a 40-dimensional vector, which will be reshaped into a 20-point trajectory.

[0081] Stationary expert networks generate balanced, safe, and standard trajectories that conform to most scenarios. The stationary expert network employs a standard network + ReLU structure, achieving structural balance and using the most stable components as a reliable benchmark, consistent with its "balanced and robust" positioning. The forward propagation formula is:

[0082] In the formula, The weight matrix for the first layer of a stationary expert network is a 704-dimensional input. Transformed into a 256-dimensional space; For input features that are stable and enhanced; This is the bias vector for the first layer of the stationary expert model; It is a standard linear rectified function. It sets all negative values ​​to zero and retains positive values, providing good non-linear properties and is one of the most commonly used activation functions in deep learning; The hidden feature vectors obtained after the first layer of a stationary expert network are computed. The weight matrix for the second layer of a stationary expert network maps the features from 256 dimensions to 128 dimensions. The hidden feature vectors of the second layer of a stationary expert network; To compress the 128-dimensional features to 40 dimensions for the weight matrix of the output layer of a stationary expert network; The original trajectory vector output by the stationary expert.

[0083] Conservative expert networks generate cautious, smooth outputs that prioritize safety margins. They employ a deep, progressive approach combined with Tanh, where network depth facilitates complex reasoning, and Tanh's outputs are smooth and bounded, naturally suppressing aggressive outputs. This perfectly embodies their "safety first" principle. The forward propagation formula is:

[0084] In the formula, The weight matrix for the first layer of a conservative expert network; The input features are conservatively enhanced. This represents the bias vector of the first layer of a conservative expert network. It is a hyperbolic tangent activation function, whose output range is compressed between (-1,1), which can produce smooth and bounded outputs, making it very suitable for generating stable, abrupt trajectories; The hidden feature vectors of the first layer of a conservative expert network; This is the weight matrix for the second layer of a conservative expert network. The hidden feature vectors of the second layer of the conservative expert network; The weight matrix for the third layer of a conservative expert network; The hidden feature vectors of the third layer of a conservative expert network; This is the weight matrix for the output layer of a conservative expert network. The original trajectory vector output for conservative experts.

[0085] Output adjustment formula:

[0086] In the formula, This is the expert's original output vector; The decision preference projection matrix; To adjust the weighting coefficients; This is the adjusted expert output.

[0087] S3.3.3 Based on the enhanced input features, the gating weights are obtained through a gating network; The role of gating networks is to act as intelligent schedulers, analyzing the current situation. The gating network formula is as follows: (The formula is missing from the original text.)

[0088] In the formula, This is the weight matrix of the first layer of the gated network; This is the bias vector of the first layer of the gated network; is the hidden feature vector of the gated network; The weight matrix of the output layer of the gated network maps the 128-dimensional features to 3 dimensions, corresponding to three experts; This is the bias vector of the output layer of the gated network; The original vector of expert scores; Softmax is the activation function. It transforms the original expert scores... Transform into a probability distribution This operation ensures the three weight values , , All are positive numbers, and their sum is 1; For the gating weight, intuitively represented as the gating weight at the th... The confidence or importance of the driving strategies represented by each expert in the current context, among which, .

[0089] S3.4 Based on the original trajectory and gating weights output by three types of experts, the trajectory is optimized after weighted trajectory fusion to output a personalized predicted trajectory; This invention, based on step S3.3.2, merges the trajectory proposals of the three experts according to their respective importance (weight) to form a unified trajectory that comprehensively considers multiple strategies. The trajectory fusion formula is as follows:

[0090] In the formula, For expert weights in aggressive gating networks, For expert weights in a stationary gating network, For conservative gating networks; Output for the adjusted radical expert This is for the adjusted, stable expert output. Output for the adjusted conservative expert; This is a scalar-vector multiplication; that is, each expert's trajectory vector is multiplied by its corresponding weight. The initial trajectory vector after fusion is a 40-dimensional vector containing information about all trajectory points after weighted averaging.

[0091] To ensure the physical feasibility and comfort of the trajectory, post-processing optimization must be performed on the initially fused trajectory. The trajectory optimization formula is as follows:

[0092] In the formula, The constraint function imposes physical and traffic rule constraints on the initial trajectory, and dynamic constraints: ensuring the velocity between trajectory points. and acceleration Within the limits allowed by the vehicle dynamics model; It is a smoothing function used to eliminate high-frequency jitter on the track and improve ride comfort; The reshaping function optimizes and smooths the 40-dimensional vector. Rearrange (reshape) into a 20×2 matrix. The output data structure was defined as follows: 20 time points, each with 2 coordinate values.

[0093] S4. Construct a hierarchical dynamic safety constraint and predictive intervention mechanism; construct the overall control architecture through personalized trajectory control mapping, perform PID trajectory tracking control calculation, conduct multimodal risk assessment, construct hierarchical dynamic safety constraints and three-layer safety assurance settings, and output vehicle control commands and system status; This invention designs a personalized safety boundary adjustment strategy, dynamically adjusting safety thresholds such as collision time and warning distance according to driving style (1.2 times more for aggressive driving and 0.8 times more for conservative driving), achieving style-appropriate safety margin. A multi-dimensional risk assessment model is constructed based on collision risk, comfort, traffic rules, and pedestrian risk to calculate the comprehensive risk value of the predicted trajectory. A three-layer safety assurance mechanism is designed, using the Sigmoid function to achieve a smooth mapping between risk and intervention intensity, and introducing hard constraints to ensure that the mandatory safety intervention weight is not less than 0.8 under extreme risks, achieving a dynamic balance between personalization and safety. The optimized trajectory is converted into vehicle control commands through physical constraints and smoothing, outputting steering wheel angle, throttle opening, and braking intensity, ensuring the executability of the trajectory and ride comfort.

[0094] Specifically, the following steps are included: S4.1 Personalized trajectory control mapping: Based on personalized predicted trajectory, preprocessed vehicle state vector and driving style label, construct the overall control architecture; like Figure 3As shown, this invention constructs a hierarchical dynamic safety constraint and predictive safety intervention mechanism suitable for personalized trajectories, converting the location trajectory into vehicle control commands to ensure trajectory smoothness, vehicle dynamics constraints, and traffic rule compliance. The input is the personalized predicted trajectory. Preprocessed vehicle state vector Driving style tags The output is vehicle control commands. ,in For steering angle, For throttle opening, For braking intensity, system status information ,in For expert weight distribution, This is the risk assessment value. The intensity of intervention. The overall control architecture expression is:

[0095] In the formula, For personalized trajectory control mapping process; The personalized predicted trajectory from step S3 contains only location coordinates; This is the current vehicle state vector; For the controller parameter set (PID gain, safety threshold, personalized adjustment factor); For vehicle control commands and system status information. Personalized predicted trajectory. The current vehicle status is directly transmitted from step S3 to the PID controller. Real-time data collected from sensors, and Timestamp alignment is achieved, and output commands are delivered via vehicle actuators.

[0096] S4.2 Based on the overall control architecture, input the personalized predicted trajectory and vehicle state vector, and obtain the preliminary vehicle control command through motion parameter derivation and PID trajectory tracking controller calculation; Based on the fundamental principle of the controller described in step S4.1, this invention designs a PID trajectory tracking controller, hierarchical dynamic safety constraints, and multimodal risk assessment modules as the main structure of the observation encoder for trajectory control. The derivation formulas for the motion parameters are as follows:

[0097] In the formula, For the first Reference velocity for each trajectory point; For the first Reference acceleration for each trajectory point; For the first Reference heading angles for each trajectory point; For time intervals.

[0098] The lateral control formula is:

[0099] In the formula, This refers to lateral positional deviation; The current vehicle's y-coordinate (from) (obtained from) The y-coordinate of the reference trajectory (from) (obtained from) This refers to the deviation in heading angle; The current heading angle (from) (obtained from) For reference heading angle (derived from position sequence); , , It is the PID gain function; For heading deviation gain; This is a turning instruction.

[0100] The longitudinal control formula is:

[0101] In the formula, For speed deviation; For the current speed (from) (obtained from) Reference velocity (calculated from the position sequence); , , For longitudinal PID gain; This refers to the throttle opening. The braking intensity is represented by the weight matrix and gain, which are learnable parameters that are fine-tuned end-to-end using simulated data.

[0102] S4.3 Constructing hierarchical dynamic safety constraints: Based on driving style labels and basic safety thresholds, a personalized safety boundary adjustment strategy is designed through the basic safety boundary to obtain the personalized safety boundary for driving style; Based on the two control methods described in step S4.2, the hierarchical dynamic safety constraint process of this invention is as follows: An adjustment relationship is established between the multi-level safety boundaries and personalized thresholds, representing the modification effect of driving style preferences on the basic thresholds; therefore, for any risk type, in a single constraint process, the hierarchical constraint process can be expressed as the basic safety boundary, with the specific expression as follows:

[0103] In the formula, Basic safety threshold, including collision time acceleration curvature ; The comfort boundary (ensuring driving comfort); This serves as a warning boundary (issuing a safety alert); Establish the intervention boundary (initiate safe intervention). Emergency boundary (mandatory security takeover).

[0104] Personalized security boundary adjustments: radical style ( ):

[0105] Stable style ( ):

[0106] Conservative style ( ):

[0107] S4.4 Multimodal Risk Assessment: Based on personalized predicted trajectory, vehicle state vector, environmental information and personalized safety boundary, a comprehensive risk assessment value is obtained through a multi-dimensional risk assessment model; The multi-dimensional risk assessment model includes: collision risk assessment, comfort risk assessment, traffic rule risk assessment, and pedestrian risk assessment. This invention performs multimodal risk assessment based on step S4.3, mainly involving collision risk, comfort risk, traffic rule risk, and pedestrian risk. It quantifies and analyzes trajectory safety from multiple dimensions, providing a decision-making basis for predictive safety intervention. The risk assessment aggregation formula is as follows:

[0108] In the formula, The comprehensive risk assessment value is obtained through weighted summation. This is a risk weight vector, reflecting the ranking of the importance of each risk dimension. The highest value is assigned to the collision risk weight, reflecting the principle of prioritizing safety. Weighting the risk based on comfort, focusing on the riding experience. Weighting traffic rule risks to ensure compliance. Pedestrian risk weights; To calculate the collision risk value, For comfort risks, For the overall risk of traffic rules, Risk to pedestrians; The collision risk assessment formula is:

[0109] In the formula, The collision time; Relative time; It is the relative velocity; As a basic security threshold, ; The calculated collision risk value; Comfort risk assessment formula:

[0110] Calculation of trajectory curvature and acceleration:

[0111] In the formula, The basic acceleration threshold is set based on human comfort. Based on the basic curvature threshold, and constrained by vehicle dynamics; Represents the velocity at different times; Represents the curvature at different times; For comfort-related risks, the first item is acceleration comfort, and the second item is curvature comfort. Indicates time interval, The x-axis is... The vertical axis is denoted as y.

[0112] Traffic rule assessment risk:

[0113] In the formula, For speed compliance assessment, ; For lane keeping assessment, ; Speed ​​limits on roads; This refers to the position of the lane centerline; Maximum permissible lane departure; The overall risk of traffic rules.

[0114] Pedestrian risk assessment:

[0115] In the formula, Minimum distance from pedestrians; This is the safe distance threshold; S4.5 Predictive safety intervention: Based on comprehensive risk assessment values ​​and personalized safety boundaries, safety intervention weights and minimum mandatory safety intervention weights are obtained through trajectory extrapolation prediction, safety intervention weight calculation, and hard constraint safety assurance mechanisms. Based on step S4.4, predict future risk trends and implement gradual safety interventions. While maintaining personalization, ensure absolute safety by constructing a predictive safety intervention mechanism. The trajectory extrapolation prediction formula is as follows:

[0116] In the formula, , The x and y coordinates of the current vehicle; For time step; This represents the number of extrapolation steps. For trajectory extrapolation and trajectory prediction, j This is the index for the extrapolation steps.

[0117] Formula for safe trajectory template:

[0118] In the formula, This is the safe displacement increment for straight-line mode. This is the safe displacement increment for parking on the side of the road; To reduce speed for safety; This is a safe trajectory template, containing a conservative path for deceleration to a stop. This is the index number of the trajectory point.

[0119] Safety intervention weight calculation:

[0120]

[0121] In the formula, To predict risk assessment values; The steepness parameter controls the smoothness of the transition. This refers to the personalized intervention boundary threshold; Weighting for safety intervention.

[0122] Hard constraint security mechanism:

[0123] In the formula, To minimize the mandatory safety intervention weight, ensure that the safety trajectory weight is not less than 80% under extreme risks; The critical safety threshold is set at 1.5 times the personalized intervention boundary. To predict the risk assessment value (from step S4.4); Temperature parameters are used to control the smoothness of the transition. This is the steepness coefficient; the larger it is, the closer it is to a hard handover. It is a numerically stable term; It is a hard sigmoid activation function, which achieves a smooth but near-step transition.

[0124] This mechanism achieves three layers of security: the first layer is through S4.3's personalized boundaries. The first layer provides reasonable safety margins for different driving styles; the second layer uses a sigmoid function to smoothly map risk assessment and safety intervention intensity, avoiding abrupt changes in control output; the third layer uses hard constraints. Ensure that a safety takeover is forcibly initiated when the risk exceeds a critical threshold, prioritizing driving safety.

[0125] The intervention level is:

[0126] S4.6, Layered Dynamic Safety Constraints: Based on safety intervention weights, the final vehicle control commands and system states are obtained through trajectory smoothing and trajectory fusion. Based on the safety intervention weight calculated in step S4.5 and minimum mandatory intervention weight The trajectory is optimized and the final control output is output. The trajectory smoothing formula is as follows:

[0127] In the formula, The smoothing coefficient controls the filtering strength; , The coordinates of adjacent trajectory points; The smoothed trajectory eliminates high-frequency jitter.

[0128] The trajectory fusion formula is:

[0129] In the formula, For a smoothed, personalized trajectory; For safe trajectory templates; The safety intervention weights from step S4.5; For the optimized final trajectory, when At that time, a fully personalized trajectory is used. The system prioritizes the use of safe trajectories to achieve a dynamic balance between safety and personalization. This triple safety mechanism ensures that the system prioritizes safety in extreme risk situations, while reserving reasonable operational space for personalized driving.

[0130] The final control output is as follows: The control commands and system status outputs are:

[0131] In the formula, The expert weight distribution comes from step S3; The optimized final trajectory; This is the current risk assessment value; Weighting for safety intervention.

[0132] S5. Design an end-to-end multi-task joint training strategy; by defining a multi-task loss function and security constraints, output training data stream storage to achieve offline batch training and online incremental learning; like Figure 4 As shown, this invention designs an end-to-end multi-task joint training strategy for the personalized expert network in step S3 and the safety constraint module in step S4. A balance between personalization and safety is achieved through the collaborative optimization of trajectory loss, control loss, safety loss, style loss, and expert differentiation loss. A safety-constrained expert differentiation loss function is designed to encourage experts to learn different strategies only under low-risk conditions, while allowing expert outputs to converge under high-risk conditions to prioritize safety. An adaptive weight adjustment mechanism is designed to dynamically balance the contribution of each sub-task, and the highest safety loss weight is determined through validation set optimization. To ensure safety as a priority, an expert-specific learning rate scheduling strategy is adopted, with an aggressive expert learning rate of 1.5 times the base learning rate to quickly adapt to highly dynamic scenarios and a conservative expert learning rate of 0.8 times to enhance stability. An end-to-end training data stream is constructed, which organizes input features, predicted trajectories, control commands, and labels into time-series data pairs and stores them in the dataset, supporting offline batch training and online incremental learning.

[0133] Specifically, the following steps are included: S5.1 End-to-end Prediction and Control; Based on the constructed personalized end-to-end prediction and control system, the optimized model parameters are obtained by defining the multi-task joint loss parameters and end-to-end forward propagation. like Figure 4 As shown, for the personalized end-to-end predictive and control system constructed in steps S3 and S4, a multi-task joint loss function is designed to achieve collaborative optimization of trajectory prediction, control output, safety constraints, style matching, and expert differentiation. This framework, through an end-to-end training strategy, enables optimization of all parameters across the entire system chain, from raw sensor data to final control commands, via gradient backpropagation, ensuring the consistency of personalized modeling, safety constraints, and overall performance.

[0134] In the formula, The optimized model parameters include all network weights (expert network weight matrix, gated network parameters, PID gain, etc.). For multi-task joint loss function; Input data; To predict the trajectory; For predictive control commands; Labels (including actual trajectories) Real instructions (Style tags). The end-to-end strategy avoids the accumulation of modular errors and learns the overall behavior directly from labeled data through supervised learning, ensuring a balance between personalization and security.

[0135] S5.2 Calculation of the joint loss function for multiple tasks: After parallel calculation of five sub-loss terms, a weighted aggregation is performed to obtain the total loss function for multiple tasks; Based on the fundamental principles of the training strategy described in step S5.1, modules such as a multi-task joint loss function, expert differential loss, and multi-objective balanced training are designed as the main structure of the end-to-end optimized observation encoder; wherein, the calculation process of the multi-task joint loss function can be described as follows:

[0136] In the formula, The trajectory loss function; To predict trajectory points; Labels for actual trajectories; It is 20; For trajectory loss weights; To control the loss function; For predictive control commands; These are actual control commands; It is 3; To control the loss weight; For the safety loss function; This is the risk assessment value; For personalized security boundaries; This is a ReLU penalty function that only penalizes violations of safety constraints; The safety loss weight is higher (emphasizing safety). Let style loss function be used. embedding vectors for predicting style; Embed vectors for real style tags; Style loss is weighted; trajectory and control losses focus on prediction accuracy; safety loss emphasizes zero-collision priority; style loss maintains personalized preferences. The weights are fixed hyperparameters, determined by grid search on the validation set.

[0137] Expert Differentiation Loss Function:

[0138] In the formula, , Trajectory fragments output by different experts; ; Cosine similarity between expert outputs; The total number of expert networks; The current predicted risk assessment value (from step S4.4) Personalize the warning boundary threshold (from step S4.3); this constraint is only applied in safe situations. When the risk exceeds the warning threshold, expert outputs are allowed to converge to prioritize safety. The negative sign ensures that similarity is minimized; the cosine similarity of expert outputs decreases, encouraging experts to learn differentiated strategies. This loss, by inversely incentivizing expert diversity, prevents all experts from converging to similar strategies, thus improving overall robustness.

[0139] For the trade-off optimization of multi-task objectives, an adaptive weight adjustment mechanism is designed to balance the contribution of each sub-loss. The total multi-task loss function is defined as follows:

[0140] In the formula, , , , , These are the weighting coefficients for each sub-loss, used to balance the importance of different optimization objectives; This is the final joint loss function.

[0141] S5.3, Adaptive Weight Adjustment: Based on the multi-task total loss function, backpropagation updates are performed through adaptive weight adjustment and learning rate scheduling strategies; Weight adaptive adjustment mechanism:

[0142] In the formula, For the first The initial weights of each loss; For the first The loss value for each training step; This is the initial loss value; The attenuation coefficient controls the speed of weight adjustment.

[0143] The base weights are determined through a grid search on the validation set: (Trajectory prediction is the primary task) (Control accuracy is secondary) (Highest security priority) (Appropriate weighting for style matching) (Expert differentiated assistance).

[0144] To achieve differentiated learning speeds among different experts, an expert-specific learning rate scheduling strategy is adopted:

[0145] In the formula, Based on the learning rate, it increases with the number of training steps. Dynamic decay; , , The learning rates are for aggressive, steady-state, and conservative experts, respectively. Aggressive experts use a higher learning rate to quickly adapt to highly dynamic scenarios, conservative experts use a lower learning rate to enhance stability and robustness, and steady-state experts maintain a standard learning rate as a baseline. This strategy maintains the numerical stability of network training by adjusting the parameter update rate rather than initializing the variance.

[0146] S5.4 Training Data Organization and Incremental Learning: Through training data organization and annotation mechanisms, differentiated storage strategies, online learning triggering strategies, annotation quantity and quality control, and driving style label annotation, abnormal data is deleted to obtain training data with optimized model parameters. For model deployment and real-time control, which have strong performance requirements, the training optimization strategy relies on the "input-output-loss-update" set of labeled data. Therefore, the end-to-end training strategy is designed as a dataset storage encoder, storing input data, predicted trajectories, control commands, and labeled parameters, and outputting iteratively optimized model parameters. The training data organization and labeling mechanism is as follows:

[0147] In the formula, for Input features at any given time (input image sequence + vehicle status + style label); The generated predicted trajectory; For label consistency loss, (Combined trajectory error and style consistency); This represents the input features for the next time step (used for sequence training).

[0148] The differentiated storage strategies are as follows: Aggressive style: storage every 3 steps to provide real-time feedback on high-frequency behaviors; Stable style: storage every 5 steps to balance learning efficiency; Conservative style: storage at the end of the trajectory (every 10 steps) to reduce noise. This strategy may result in relatively sparse conservative samples, but this is mitigated through data augmentation (DA) mechanisms to ensure the balance of the training process. Specifically, the aggressive style uses real-time augmentation (such as injecting random noise into the input), while the conservative style performs batch augmentation at the end of each cycle (such as flipping the trajectory to increase diversity), thereby maintaining dataset balance.

[0149] Online learning is triggered when the amount of accumulated new data exceeds a threshold (e.g., 1000 trajectories) or when the model prediction error continues to rise. Incremental training employs an experience replay mechanism, randomly sampling from historical data and mixing it with new data to avoid catastrophic forgetting.

[0150] Data quality control: The collected real trajectories are verified in multiple dimensions, and the following abnormal data are removed: (1) Position jumps caused by sensor failure (position change exceeding 5 meters); (2) Human violations (running red lights, driving against traffic, etc.); (3) Data with visibility less than 10 meters under extreme weather conditions; (4) Segments with cumulative errors exceeding 2 meters caused by GPS signal obstruction. Data that has passed quality inspection is entered into the training set to ensure that the model learns safe driving behaviors that comply with traffic rules.

[0151] Driving style labeling process: Based on the real-time driving style recognition results in step S3.1 Based on user's historical driving behavior statistics (average speed, acceleration distribution, lane change frequency), the data is initially classified by an automatic labeling algorithm, and then manually verified (10% of users are checked) to form the final style tags. .

[0152] S6. Construct an online deployment and real-time inference framework for personalized autonomous driving models, save the trained model parameters, deploy the end-to-end prediction and control models to the vehicle computing platform, and complete the construction of a personalized end-to-end autonomous driving control system that considers the spatiotemporal context. This invention constructs an online deployment and real-time inference framework for personalized autonomous driving models, saves the trained model parameters, and deploys the end-to-end prediction and control model to an onboard computing platform. The model receives multi-frame image data (camera image sequence), vehicle status (position, speed, heading angle), and user style labels as input. After spatiotemporal feature extraction in step S2, personalized expert network inference in step S3, and processing by the safety constraint module in step S4, it outputs vehicle control commands (steering angle, throttle opening, braking intensity) and system status information (expert weight distribution, current risk level, and safety intervention intensity).

[0153] Specifically, the following steps are included: S6.1 Design an online adaptive mechanism that triggers incremental learning to update model parameters when a new driving scenario is detected or prediction performance deteriorates; employ an experience replay strategy to avoid catastrophic forgetting and ensure the model continuously adapts to user driving habits and environmental changes; determine whether a new scenario has been encountered by monitoring the uncertainty of the model output, using the following detection formula:

[0154] In the formula, Entropy calculation based on expert weights is used to predict uncertainty. For the first The weights of each expert (from S3.3); It is a numerically stable term; The uncertainty threshold is 1.0; The average loss from the most recent 10 predictions; To validate the loss benchmark for the training set; when the uncertainty is too high or the loss continues to rise, it is determined that a new scenario has been encountered.

[0155] Experience replay sampling strategy:

[0156] In the formula, For newly collected datasets; This serves as a buffer for experience replay. This represents the number of new samples; This represents the number of historical samples; the batch size is 32.

[0157] Incremental update formula:

[0158] In the formula, For the first Layer parameters; To fine-tune the learning rate, set it to 1 / 10 of the base learning rate; For online batch loss, only the expert network of S3 and the control layer of S4 are updated, while the ResNet-50 and BiLSTM feature extractors of S2 are frozen to avoid excessive perturbation of the underlying features by the new scene.

[0159] Model hot update mechanism:

[0160] In this formula, incremental training in the background is triggered when the accumulated new data exceeds 1000 trajectories or when the loss increases by more than 20% for five consecutive control cycles. After training is complete, parameters are seamlessly switched through a double buffering mechanism.

[0161] This mechanism ensures that the main inference thread is not interrupted, the parameter update delay is less than 1ms, and the model iteration and upgrade are achieved without being noticed.

[0162] S6.2 Security Monitoring and Manual Takeover: The system continuously monitors the expert weight distribution and risk level. When the intensity of security intervention exceeds the critical threshold, a warning is issued through the human-computer interaction interface. The takeover judgment formula is:

[0163] In the formula, Weights for safety intervention (from S4.5); To predict risk values; The threshold is the critical safety threshold. When forced intervention lasts for 2 seconds and the driver does not respond, the system executes a conservative safety trajectory to achieve gradual safety degradation and ensure the unity of personalized driving and safety protection.

[0164] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A personalized end-to-end autonomous driving control method considering spatiotemporal context, characterized in that, Includes the following steps: S1. Construct a personalized end-to-end autonomous driving control system architecture that takes into account the spatiotemporal context; The construction methods include: designing an end-to-end deep neural network architecture, including an input layer, a feature extraction layer, a personalized modeling layer, a trajectory prediction layer, a safety constraint layer, and an output layer; The input structure of the input layer includes visual input, status input, and style input; S2, Multimodal Data Perception and Spatiotemporal Feature Extraction Module: Based on the input structure of the input layer, visual feature sequences are extracted through ResNet-50, bidirectional LSTM learns temporal dependencies and MLP encodes state information, outputs spatiotemporal context features, and completes the construction of the feature extraction layer; The multimodal data includes visual image sequences, vehicle state vectors, and driving style labels; S3. Personalized Driving Style Modeling and Expert Hybrid Network Architecture: Based on spatiotemporal context features and style input, personalized enhanced context feature vectors are obtained through an embedding layer. Then, three types of expert networks—aggressive, stable, and conservative—are constructed based on an expert knowledge base to complete the construction of the personalized modeling layer. An expert data fusion mechanism is designed to dynamically combine the personalized context feature vectors and the expert knowledge base through feature enhancement formulas to generate differentiated expert-specific enhanced features. A gating network is introduced to calculate gating weights. Based on the original trajectory vectors output by the three types of experts and the gating weights, personalized predicted trajectories are obtained through trajectory fusion and trajectory optimization to complete the construction of the trajectory prediction layer. S4. Construct a hierarchical dynamic safety constraint and predictive intervention mechanism; construct the overall control architecture through personalized trajectory control mapping, perform PID trajectory tracking control calculation, conduct multimodal risk assessment, construct hierarchical dynamic safety constraints and three-layer safety assurance settings, output vehicle control commands and system status, and complete the construction of the safety constraint layer and output layer; S5. Design an end-to-end multi-task joint training strategy; by defining a multi-task loss function and security constraints, output training data stream storage to achieve offline batch training and online incremental learning; S6. Construct an online deployment and real-time inference framework for personalized autonomous driving models, save the trained model parameters, deploy the end-to-end prediction and control models to the vehicle computing platform, and complete the construction of a personalized end-to-end autonomous driving control system that considers spatiotemporal context.

2. The personalized end-to-end autonomous driving control method considering spatiotemporal context according to claim 1, characterized in that, S3 specifically includes the following steps: S3.1 Driving style embedding and style adaptation; Based on driving style labels and spatiotemporal context features, after driving style embedding and style adaptation, feature fusion is performed by vector concatenation to obtain personalized enhanced context feature vectors; The formula for embedding driving style is: ； In the formula, The input driving style label is 0 for aggressive, 1 for moderate, and 2 for conservative. The embedding matrix contains trainable parameters; Embed the output 64-dimensional style vector; The formula for style adaptation is: ； In the formula, Embed the initial tag; For real-time identification of embedded tags; Preset adaptive weight hyperparameters; The final style embedding vector after adaptive adjustment; The formula for obtaining the personalized enhanced context feature vector is: ； In the formula, Input features from the spatiotemporal context building module; This is a vector concatenation operation; The output is a 704-dimensional personalized enhanced contextual feature vector; S3.

2. Construct an expert knowledge base based on trajectory templates, driving modes, and decision preferences; The expression for constructing an expert knowledge base is as follows: ； In the formula, The trajectory template feature vector for driving style. For driving mode feature vectors, Let be the decision preference vector. , For driving style type index, Radical type It is a stable type. Conservative type For expert knowledge base; S3.3 Hybrid Expert Network Reasoning: Based on the expert knowledge base and personalized enhanced context feature vectors, gating weights are obtained through feature enhancement mechanisms, forward propagation of hybrid expert networks, and gating networks. S3.4 Based on the original trajectory and gating weights output by three types of experts, personalized predicted trajectories are output after trajectory fusion and trajectory optimization; The trajectory fusion formula is: ； In the formula, For expert weights in aggressive gating networks, For expert weights in a stationary gating network, For conservative gating networks; Output for the adjusted radical expert This is for the adjusted, stable expert output. Output for the adjusted conservative expert; This is a scalar-vector multiplication. The trajectory optimization formula is: ； In the formula, For constraint functions; It is a smoothing function; For reshaping functions; After trajectory optimization and smoothing Rearrange to generate a 20×2 matrix. The data structure for outputting personalized predicted trajectories is as follows: 20 time points, each with 2 coordinate values.

3. The personalized end-to-end autonomous driving control method considering spatiotemporal context according to claim 2, characterized in that, S3.3 specifically includes the following steps: S3.3.1 Based on the expert knowledge base and personalized enhanced contextual feature vectors, enhanced input features are obtained through a feature enhancement mechanism; The expression for the enhanced input features is: ； In the formula, The trajectory feature projection matrix; The projection matrix is ​​the feature projection matrix for driving modes, which is obtained through supervised pre-training using a knowledge base. , For fusion weighting coefficients; For enhanced input features; To personalize and enhance contextual feature vectors, For style k The hidden representation of trajectory template features, style k Hidden representation of driving mode characteristics; S3.3.2 Constructing three types of expert networks—radical, stable, and conservative—based on an expert knowledge base: Inputting enhanced input features, the original trajectory vectors of the three types of expert outputs are obtained through a hybrid expert network; The hybrid expert network consists of three parallel radical expert networks, a stationary expert network, a conservative expert network, and a gated network. The original trajectory vectors output by the three types of experts include: the original trajectory vector output by radical experts, the original trajectory vector output by stationary experts, and the original trajectory vector output by conservative experts. S3.3.3 Based on the enhanced input features, the gating weights are obtained through a gating network; The formula for calculating the gated network is: ； In the formula, The input features are after radical enhancement. For input features that are enhanced for stationarity, For the conservatively enhanced input features, This is the weight matrix of the first layer of the gated network; This is the bias vector of the first layer of the gated network; is the hidden feature vector of the gated network; The weight matrix of the output layer of the gated network maps the 128-dimensional features to 3 dimensions; This is the bias vector of the output layer of the gated network; The original vector of expert scores; Softmax is the activation function; For the gating weight, intuitively represented as the gating weight at the th... The confidence or importance of the driving strategies represented by each expert in the current context, among which, .

4. The personalized end-to-end autonomous driving control method considering spatiotemporal context according to claim 1, characterized in that, S4 specifically includes the following steps: S4.1 Personalized trajectory control mapping: Based on personalized predicted trajectory, preprocessed vehicle state vector and driving style label, construct the overall control architecture; The overall control architecture expression is: ； In the formula, For personalized trajectory control mapping process; For personalized trajectory prediction, only location coordinates are included; This is the preprocessed vehicle state vector; This is a set of controller parameters, including PID gain, safety threshold, and personalized adjustment factors; For vehicle control commands and system status information; S4.2 Based on the overall control architecture, input the personalized predicted trajectory and vehicle state vector, and obtain the preliminary vehicle control command through motion parameter derivation and PID trajectory tracking controller calculation; The PID trajectory tracking controller calculation includes lateral control and longitudinal control calculations; S4.3 Constructing hierarchical dynamic safety constraints: Based on driving style labels and basic safety thresholds, a personalized safety boundary adjustment strategy is designed through the basic safety boundary to obtain the personalized safety boundary for driving style; The safety boundaries include comfort boundaries, warning boundaries, intervention boundaries, and emergency boundaries; S4.4 Multimodal Risk Assessment: Based on personalized predicted trajectory, vehicle state vector, environmental information and personalized safety boundary, a comprehensive risk assessment value is obtained through a multi-dimensional risk assessment model; The multi-dimensional risk assessment model includes: collision risk assessment, comfort risk assessment, traffic rule risk assessment, and pedestrian risk assessment. The aggregation formula for risk assessment is: ； In the formula, This is a comprehensive risk assessment value; To preset collision risk weights, To pre-determine comfort risk weights, To preset the risk weights of traffic rules, To preset pedestrian risk weights; To calculate the collision risk value, For comfort risks, For the overall risk of traffic rules, Risk to pedestrians; S4.5 Predictive safety intervention: Based on comprehensive risk assessment values ​​and personalized safety boundaries, safety intervention weights and minimum mandatory safety intervention weights are obtained through trajectory extrapolation prediction, safety intervention weight calculation, and hard constraint safety assurance mechanisms. The formula for calculating the weight of safety intervention is: ； In the formula, To predict risk assessment values; The steepness parameter controls the smoothness of the transition. This refers to the personalized intervention boundary threshold; Weighting for safety intervention; Sigmoid activation function S4.6, Layered Dynamic Safety Constraints: Based on safety intervention weights, the final vehicle control commands and system states are obtained through trajectory smoothing and trajectory fusion. ； In the formula, For a smoothed, personalized trajectory; For safe trajectory templates; The safety intervention weights from step S4.5; For the optimized final trajectory, when At that time, a fully personalized trajectory is used. The safe trajectory must be used first in a timely manner; The final control output is as follows: The formula for the control command and system status output is: ； In the formula, Expert weight distribution; The optimized final trajectory; This is the current risk assessment value; This is the current vehicle state vector; Weighting for safety intervention; For steering angle, For throttle opening, This refers to braking intensity.

5. A personalized end-to-end autonomous driving control method considering spatiotemporal context according to claim 1, characterized in that, S2 specifically includes the following steps: S2.1 Data Preprocessing: Based on the input structure of the input layer, preprocessing operations are performed through the data flow pipeline to output the preprocessed image sequence and the preprocessed vehicle state vector; S2.2 Visual Feature Extraction: Input the preprocessed image sequence and extract visual features through a pre-trained ResNet-50 convolutional neural network to obtain a visual feature sequence; The method of pre-training the ResNet-50 convolutional neural network is as follows: the ResNet-50 convolutional neural network is pre-trained on the ImageNet dataset, the fully connected classification layer is removed, the convolutional feature extraction part and the global average pooling layer are retained, and a 1024-dimensional feature vector is output. S2.3 Temporal Feature Encoding Module: Based on the visual feature sequence and the preprocessed vehicle state vector, the temporal dependency is learned by fusing the features with a bidirectional LSTM temporal feature encoder to obtain the temporal feature vector; S2.4, State Information Encoding Module: Based on the preprocessed vehicle state vector, the state information is encoded through a multilayer perceptron (MLP) to obtain the state feature vector; S2.5 Feature Fusion: Input temporal feature vector and state feature vector, and perform feature fusion by concatenating the vectors to obtain spatiotemporal context feature vector; The feature fusion formula is: ； In the formula, The fused output is a 640-dimensional comprehensive spatiotemporal context feature vector; This is the output vector from the temporal feature encoding module; Output vector from the state information encoding module; This is the vector concatenation operator.

6. A personalized end-to-end autonomous driving control method considering spatiotemporal context according to claim 1, characterized in that, S5 specifically includes the following steps: S5.1 End-to-end Prediction and Control; Based on the constructed personalized end-to-end prediction and control system, the optimized model parameters are obtained by defining the multi-task joint loss parameters and end-to-end forward propagation. S5.2 Calculation of the joint loss function for multiple tasks: After parallel calculation of five sub-loss terms, a weighted aggregation is performed to obtain the total loss function for multiple tasks; The five loss terms include: trajectory loss function, control loss function, safety loss function, style loss function, and expert differentiation loss function; The expression for calculating the total loss function for multiple tasks is: ； In the formula, The trajectory loss function; For trajectory loss weights; To control the loss function; To control the loss weight; For the safety loss function; For safety losses; Let style loss function be used. Weighting is applied to style; S5.3, Adaptive Weight Adjustment: Based on the multi-task total loss function, backpropagation updates are performed through adaptive weight adjustment and learning rate scheduling strategies; S5.4 Training Data Organization and Incremental Learning: Through training data organization and annotation mechanisms, differentiated storage strategies, online learning triggering strategies, annotation quantity and quality control, and driving style label annotation, abnormal data is deleted to obtain training data with optimized model parameters. The labeling mechanism for quantity and quality control and driving style labels is as follows: ； In the formula, for Input characteristics at any given time; The generated predicted trajectory; For label consistency loss, , This represents the input features at the next time step.

7. A personalized end-to-end autonomous driving control method considering spatiotemporal context according to claim 1, characterized in that, In the designed end-to-end deep neural network architecture, a multi-layer neural network architecture combined with safety constraints is used for the final vehicle command output. The expression for the multi-layer neural network architecture is: ； In the formula, For the feature extraction layer, Convert to a spatiotemporal context representation; For the personalization layer, integrate style embedding; This is the trajectory prediction layer, which generates predicted trajectories. For the safety constraint layer, apply layered boundaries; Generate instructions for the control layer.

8. A personalized end-to-end autonomous driving control method considering spatiotemporal context according to claim 4, characterized in that, Based on driving style tags and basic safety thresholds, a personalized safety boundary adjustment strategy is obtained, which is as follows: ； In the formula, To preset the basic security threshold; For comfort boundaries; Warning boundary; For intervention boundaries; This is an emergency boundary.

9. A personalized end-to-end autonomous driving control method considering spatiotemporal context according to claim 4, characterized in that, Among the collision risk assessment, comfort risk assessment, traffic rule risk assessment, and pedestrian risk assessment, the collision risk assessment formula is: ； In the formula, The collision time; Relative time; It is the relative velocity; As a basic security threshold, ; The calculated collision risk value; Comfort risk assessment formula: ； In the formula, The basic acceleration threshold is set based on human comfort. Based on the basic curvature threshold, and constrained by vehicle dynamics; Represents the velocity at different times; Represents the curvature at different times; For comfort risks; Indicates a time interval; The x-axis is the horizontal axis. The vertical axis is used as the coordinate. Traffic rule assessment risk: ； In the formula, For speed compliance assessment; Lane keeping assessment; The overall risk of traffic rules.

10. A personalized end-to-end autonomous driving control method considering spatiotemporal context according to claim 6, characterized in that, The constructed personalized end-to-end predictive and control system, by defining multi-task joint loss parameters and end-to-end forward propagation, obtains optimized model parameters. The expression for the optimized model parameters is as follows: ； In the formula, These are the optimized model parameters; For multi-task joint loss function; Input data; To predict the trajectory; For predictive control commands; For labeling.

Images