A vehicle longitudinal intention and working condition prediction method, device and equipment

Abstract

This invention relates to the field of data processing technology, and discloses a method, apparatus, and device for predicting vehicle longitudinal intent and operating conditions. The method includes: acquiring original traffic flow features and splitting them into historical trajectory time-series data and current operating condition feature sequences; inputting the historical trajectory time-series data into a style encoder to generate a theoretical physical baseline, and calculating the mirror residual sequence between the actual action and the baseline; aggregating and reducing the dimensionality of the mirror residual sequence to extract a driving style vector; inputting the current operating condition feature sequence into a unidirectional shared network to extract the instantaneous physical momentum state feature vector; generating modulation parameters based on the driving style vector, linearly modulating the state features to output a modulated state feature vector; inputting the driving style vector into a gating network to output expert routing distribution weights, and inputting the modulated state feature vector into multiple expert subnetworks to obtain candidate prediction baseline values, which are then weighted and aggregated to obtain the predicted value of the vehicle's longitudinal kinematic state. This invention can improve the timeliness and accuracy of vehicle operating condition prediction.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, and in particular relates to a method, apparatus and equipment for predicting the longitudinal intent and operating conditions of a vehicle. Background Technology

[0002] In the current field of autonomous driving and advanced driver assistance systems (ADAS), vehicle longitudinal intent and condition prediction (such as acceleration or jerk prediction in adaptive cruise control (ACC) systems) are key technologies for achieving safe and smooth control of intelligent vehicles. Currently, traditional longitudinal intent prediction methods mostly adopt end-to-end regression architectures based on deep learning.

[0003] Existing prediction schemes typically collect raw dynamic data from vehicle sensors (including but not limited to vehicle speed, acceleration, and relative distance and speed to vehicles ahead). During data processing and model prediction, current technologies often simply concatenate these temporal features into a high-dimensional vector and directly input it into standard recurrent neural network (RNN) or long short-term memory network (LSTM) processing modules for feature extraction. Simultaneously, during model optimization and training, mean squared error (MSE) is typically used as the loss function, and all collected traffic flow data are subjected to undifferentiated full training.

[0004] However, through long-term research and development and practice, it has been found that the aforementioned existing technologies have the following obvious technical shortcomings when dealing with complex and ever-changing real-world driving scenarios: First, the deep coupling between driving intention and vehicle physical motion characteristics results in an extremely low signal-to-noise ratio for the predicted signal. Existing end-to-end models typically input historical trajectory features directly as a "black box." This approach cannot effectively distinguish whether changes in vehicle motion are due to the vehicle's own physical inertia (e.g., natural deceleration or inertial roll caused by road slope) or the driver's subjective intention (e.g., acceleration intentionally applied for overtaking). Because general physical motion laws are highly coupled with individual subjective driving intentions in the feature space, a large amount of physical noise is introduced during the feature extraction stage. When the vehicle is in a complex game situation or its state changes drastically, the model is prone to misinterpreting normal physical feedback as intentional steering, resulting in an "impure" predicted signal that severely affects prediction accuracy.

[0005] Second, static mapping and a single spatiotemporal architecture struggle to balance "long-term style stability" and "instantaneous decision sensitivity." Existing prediction architectures typically fail to distinguish between temporal and spatial scales, attempting to capture both long-term driver characteristics (such as aggressive or conservative driving habits) and immediate physical constraints (such as millisecond-level emergency obstacle avoidance maneuvers) within the same network weight space (e.g., a single LSTM backbone). This forced fusion and static mapping of long-term qualitative features with short-term quantitative features is highly prone to feature conflicts during model parameter updates. This causes existing models to lose sensitivity to sudden conditions when pursuing smooth predicted trajectories, while causing severe fluctuations in predicted trajectories when pursuing rapid responses, ultimately resulting in significant time lag effects in the prediction system.

[0006] Third, the severe imbalance in traffic flow data leads to serious "mean regression" and mediocre predictions in the models. In natural driving scenarios (such as highways or urban expressways), the vast majority of traffic flow data (usually over 90%) falls under smooth or constant-speed conditions (with acceleration approaching zero). Existing technologies use undifferentiated full-data training, forcing the model to overfit these predominantly "mediocre samples" during optimization. To reduce overall training error, the model tends to learn a conservative strategy that consistently outputs near zero. This "central tendency effect" causes the model to produce overly mediocre predictions when faced with high-value, high-dynamic extreme conditions such as emergency lane changes or rapid acceleration. The system exhibits sluggish response and severely insufficient predictive ability, failing to meet the control requirements of autonomous driving in critical moments.

[0007] In summary, existing technologies result in lower timeliness and accuracy in predicting vehicle operating conditions. Summary of the Invention

[0008] This invention provides a method, apparatus, and equipment for predicting the longitudinal intent and operating conditions of a vehicle, which can improve the timeliness and accuracy of vehicle operating condition prediction.

[0009] To achieve the above objectives, the present invention provides a method for predicting the longitudinal intent and operating conditions of a vehicle, comprising: Obtain the original traffic flow characteristics of vehicles, and split the original traffic flow characteristics into historical trajectory time series data based on historical preset time windows and current operating condition characteristic sequence based on the current moment; Historical trajectory time series data is input into a style encoder driven by a physical mirror. The theoretical physical reference sequence is generated through the dynamic mirror layer of the style encoder, and the mirror residual sequence between the actual action sequence corresponding to the historical trajectory time series data and the theoretical physical reference sequence is calculated. Feature aggregation and dimensionality reduction are performed on the mirror residual sequence to extract driving style vectors that represent the subjective operating habits of the vehicle driver. The current operating condition feature sequence is input into a one-way shared network for feature compression, and a state feature vector representing the current instantaneous physical momentum state is extracted. Modulation parameters are generated based on driving style vectors, and linear feature modulation is performed on the state feature vector using the modulation parameters to output the modulated state feature vector. The driving style vector is input into the gating network to output the expert routing distribution weights, and the modulation state feature vector is input into multiple expert sub-networks set in parallel to output the candidate prediction benchmark value of each expert. The predicted value of the vehicle's longitudinal kinematic state is obtained by weighting and aggregating multiple candidate prediction benchmark values ​​according to the expert route distribution weights.

[0010] To address the above problems, the present invention also provides a vehicle longitudinal intention and operating condition prediction device, the device comprising: The driving style vector acquisition module is used to acquire the original traffic flow features of vehicles and split the original traffic flow features into historical trajectory time series data based on a historical preset time window and current operating condition feature sequence based on the current moment. The historical trajectory time series data is input into the style encoder driven by physical mirror. The theoretical physical reference sequence is generated through the dynamic mirror layer of the style encoder, and the mirror residual sequence between the actual action sequence corresponding to the historical trajectory time series data and the theoretical physical reference sequence is calculated. The mirror residual sequence is subjected to feature aggregation and dimensionality reduction to extract the driving style vector representing the subjective operating habits of the vehicle driver. The state feature vector extraction module is used to input the current working condition feature sequence into a one-way shared network for feature compression and extract the state feature vector representing the current instantaneous physical momentum state. The vehicle longitudinal intent and operating condition prediction module is used to generate modulation parameters based on the driving style vector, perform linear feature modulation on the state feature vector using the modulation parameters, and output the modulated state feature vector. The driving style vector is input into the gating network to output the expert routing distribution weights, and the modulated state feature vector is input into multiple expert sub-networks set in parallel to output the candidate prediction benchmark value of each expert. The multiple candidate prediction benchmark values ​​are weighted and aggregated according to the expert routing distribution weights to obtain the predicted value of the vehicle's longitudinal kinematic state.

[0011] To address the above problems, the present invention also provides an electronic device, the electronic device comprising: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the vehicle longitudinal intent and operating condition prediction method described above.

[0012] To address the aforementioned problems, the present invention also provides a computer-readable storage medium storing at least one computer program, which is executed by a processor in an electronic device to implement the vehicle longitudinal intent and operating condition prediction method described above.

[0013] This invention acquires the original traffic flow features of vehicles and decomposes these features into historical trajectory time-series data based on a preset historical time window and current operating condition feature sequences based on the current moment, enabling the extraction of long-term behavioral features and instantaneous state features. Furthermore, by inputting the historical trajectory time-series data into a physically mirror-driven style encoder, a theoretical physical reference sequence is generated through the style encoder's dynamic mirror layer. The mirror residual sequence between the actual action sequence corresponding to the historical trajectory time-series data and the theoretical physical reference sequence is calculated, thus filtering out physical noise. Moreover, feature aggregation and dimensionality reduction are performed on the mirror residual sequence to extract driving style vectors representing the subjective operating habits of the vehicle driver. This quantifies the driver's personalized operations into a vector format that the model can understand, eliminating interference from single random operations by the driver. Additionally, the current operating condition feature sequence is input into a unidirectional shared network for feature compression, extracting state feature vectors representing the current instantaneous physical momentum state, which can transform the current complex... The complex driving condition features are mapped to a unified state representation. Furthermore, modulation parameters are generated based on the driving style vector, and these parameters are used to linearly modulate the state feature vector, outputting a modulated state feature vector. This allows the current instantaneous state features to adaptively adjust according to the long-term driving styles of different drivers. Additionally, the driving style vector is input into a gating network to output expert route distribution weights, and the modulated state feature vector is input into multiple parallel expert subnetworks to output candidate prediction benchmark values ​​for each expert. A hybrid expert prediction mechanism is constructed through the gating network and multiple expert subnetworks, enabling the model to assign different expert prediction paths for different driving styles. Finally, the multiple candidate prediction benchmark values ​​are weighted and aggregated according to the expert route distribution weights to obtain the predicted value of the vehicle's longitudinal kinematic state. Adaptive weighted aggregation of multiple candidate prediction benchmark values ​​using expert route weights allows the prediction results to be dynamically adjusted according to driving style and current driving conditions, improving the timeliness and accuracy of vehicle condition prediction. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating a method for predicting the longitudinal intent and operating conditions of a vehicle according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the overall algorithm flow of an example of a vehicle longitudinal intent and operating condition prediction method provided in an embodiment of the present invention; Figure 3 A model network architecture diagram of a vehicle longitudinal intent and operating condition prediction method provided in an embodiment of the present invention; Figure 4 This is a functional block diagram of a vehicle longitudinal intention and operating condition prediction device provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of an electronic device for implementing the vehicle longitudinal intent and operating condition prediction method according to an embodiment of the present invention.

[0015] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0016] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0017] This application provides a method for predicting vehicle longitudinal intent and operating conditions. The execution entity of this method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the method for predicting vehicle longitudinal intent and operating conditions can be executed by software or hardware installed on a terminal device or a server device, and the software can be a blockchain platform. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0018] Reference Figure 1 The diagram shown is a flowchart illustrating a vehicle longitudinal intent and operating condition prediction method according to an embodiment of the present invention. In this embodiment, the vehicle longitudinal intent and operating condition prediction method includes: S1. Obtain the original traffic flow features of vehicles and split the original traffic flow features into historical trajectory time series data based on historical preset time windows and current operating condition feature sequence based on the current moment.

[0019] Understandably, traffic flow primitive characteristics refer to various physical quantities acquired through onboard sensors or external sensing devices that describe the vehicle's motion state and the spatial interaction between the vehicle and surrounding traffic participants. For example, traffic flow primitive characteristics include, but are not limited to, the vehicle's longitudinal velocity, lateral velocity, longitudinal acceleration, lateral acceleration, and the actual jerk obtained by differentiating the acceleration, which reflects the micro-control operations.

[0020] Understandably, the historical preset time window refers to a continuous sampling period preceding the current moment.

[0021] Understandably, the current operating condition feature sequence refers to a subset of data describing the instantaneous physical state at the current moment.

[0022] Specifically, before obtaining the original traffic flow characteristics of vehicles, the process also includes: Acquire raw data including the vehicle's longitudinal velocity, longitudinal acceleration, actual jerk, and relative distance and relative velocity representing the relationship between the vehicle and the lead vehicle; The original data is input into a local polynomial least squares fitting filter. After smoothing and denoising the high-frequency noise of the third derivative signal in the local polynomial least squares fitting filter, the original traffic flow characteristics are obtained.

[0023] It is understandable that jerk refers to the rate of change of acceleration with respect to time, that is, the third derivative of displacement with respect to time (or the second derivative of velocity with respect to time). In the field of vehicle longitudinal dynamics, it represents how quickly the longitudinal acceleration of a vehicle changes.

[0024] S2. Input the historical trajectory time series data into the style encoder driven by the physical mirror, generate the theoretical physical reference sequence through the dynamic mirror layer of the style encoder, and calculate the mirror residual sequence between the actual action sequence corresponding to the historical trajectory time series data and the theoretical physical reference sequence.

[0025] Understandably, a physical mirror-driven style encoder refers to a neural network feature extraction module that integrates a physical prior law constraint mechanism. It contains at least a dynamic mirror layer and a temporal feature extraction network (such as a bidirectional long short-term memory network Bi-LSTM). Its core logic is to simulate the benchmark response of a vehicle under pure physical constraints and compare it with the control signals actually generated by the driver. Through the physical mirror-driven style encoder, the common parts affected by the physical characteristics of the vehicle can be extracted from the mixed sensor signals, thereby extracting a driving style vector that reflects the driver's personalized operating tendencies.

[0026] Understandably, the dynamics mirror layer refers to a physical response simulation component within the style encoder, used to calculate in real time the theoretical physical reference motion signal that conforms to the common sense of dynamics physics based on the input vehicle state. The dynamics mirror layer maps the input historical trajectory time series data through learnable linear neurons combined with physical boundary truncation operators (such as the torch.clamp operator) to generate a theoretical physical reference motion signal within a preset physical boundary range. The physical boundary truncation operator is a hard clipping operation based on physical constraints, which strictly clips the theoretical physical reference motion signal calculated by the linear neurons to the preset driving physical boundary through operators such as torch.clamp, thereby ensuring that the output theoretical physical reference motion signal meets the dynamic boundary conditions such as acceleration and jerk.

[0027] Understandably, the mirror residual sequence refers to the time-series signal that reflects the difference between the actual driving actions of the driver and the theoretical physical reference motion.

[0028] Specifically, a theoretical physical reference sequence is generated through the dynamic mirror layer of the style encoder, and the mirror residual sequence between the actual action sequence corresponding to the historical trajectory time series data and the theoretical physical reference sequence is calculated, including: Extract the actual jerk sequence from historical trajectory time series data using indexing operators as the actual motion sequence; By inputting historical trajectory time series data into a linear neuron containing a physical boundary truncation operator, a theoretical physical benchmark sequence that conforms to the consensus boundary of dynamic physics and is within a preset value range is obtained. The actual action sequence and the theoretical physical reference sequence are input into the subtraction node to calculate the mirror residual sequence.

[0029] Understandably, an indexing operator is a data extraction module used in tensor computation or high-dimensional data processing to accurately locate, peel off, and extract specific dimension data from a multidimensional feature matrix. Through the indexing operator, the system can actively discard other physical state quantities that are not directly related to the calculation of the intention residual. For example, historical trajectory time series data contains a 17-dimensional feature matrix including the vehicle's speed, acceleration, and relative distance. The indexing operator can accurately extract the actual acceleration data representing the vehicle's control micro-operations across time series dimensions based on preset dimension coordinates (such as locating to the 9th dimension feature).

[0030] S3. Perform feature aggregation and dimensionality reduction on the mirror residual sequence to extract the driving style vector that represents the subjective operating habits of the vehicle driver.

[0031] Understandably, feature aggregation and dimensionality reduction refer to the computational process of extracting evolutionary patterns with a time span from a high-dimensional temporal feature sequence, forcibly filtering out redundant instantaneous physical noise, and compressing it into a low-dimensional, high-purity representation vector.

[0032] Understandably, the subjective operating habits of the driving subject refer to the driving behavior tendencies exhibited by the vehicle controller within a long historical observation window, which are highly individualized.

[0033] Specifically, feature aggregation and dimensionality reduction are performed on the mirror residual sequence to extract a driving style vector representing the subjective operating habits of the vehicle driver, including: The enhanced feature matrix is ​​obtained by concatenating the mirror residual sequence with the original historical trajectory time series data; The enhanced feature matrix is ​​input into a bidirectional long short-term memory network that is regularized by a layer normalization function; The evolution trend features of driving behavior are extracted using the forward network of the bidirectional long short-term memory network, and the backward consistency features of driving behavior are extracted using the backward network of the bidirectional long short-term memory network. The evolution trend features and the inverse consistency features are fused to obtain the initial fused features. The initial fused features are then input into a mapping network consisting of fully connected layers, nonlinear activation functions, and layer normalization functions to perform low-dimensional mapping and obtain the driving style vector.

[0034] Understandably, a bidirectional long short-term memory network is a deep neural network architecture specifically designed for processing sequential data. It consists of two parallel but oppositely directed long short-term memory (LSTM) network sublayers stacked together.

[0035] Understandably, evolutionary trend features refer to the hidden state representations that reflect the cumulative changes in driving operations over time, extracted from the feedforward network (i.e., the subnetwork that transmits information in chronological order) in a bidirectional long short-term memory network.

[0036] Understandably, reverse consistency features refer to the hidden state representation extracted from the backward network (i.e., the sub-network that transmits information in reverse chronological order) in a bidirectional long short-term memory network. This is achieved by backtracking from the current moment of the observation window to the historical moment to verify the logical connection between behavioral mutations and their historical states.

[0037] For example, feature aggregation and dimensionality reduction of the mirror residual sequence to extract a driving style vector representing the subjective operating habits of the vehicle driver can be achieved through the following implementation steps: Step 1: Generate the mirror residual sequence As the 18th feature, it is concatenated with the original 17-dimensional feature (the original historical trajectory time series data) using the torch.cat tensor function to form... The enhanced feature matrix.

[0038] Step 2: Deep Feature Extraction. The concatenated matrix is ​​input into a Bi-LSTM (Bidirectional Long Short-Term Memory) network. This network is configured with two layers, with a depth of 64 hidden layers. The forward LSTM learns the evolution trend of driving behavior, while the backward LSTM learns the inverse consistency of behavior. The combination of the two eliminates the delay in recognizing sudden actions by the unidirectional model. For regularization, a LayerNorm (layer normalization) layer is applied after the Bi-LSTM, resulting in stronger convergence stability when processing non-stationary driving data streams.

[0039] Step 3, Decoupling output of driving style vector S: Finally, through a mapping network containing fully connected layers, ReLU activation, and a second LayerNorm (layer normalization), the 128-dimensional (bidirectional 64+64) features are compressed into a 16-dimensional driving style vector S.

[0040] S4. Input the current operating condition feature sequence into a one-way shared network for feature compression, and extract the state feature vector representing the current instantaneous physical momentum state.

[0041] Understandably, a one-way shared network refers to a temporal feature extraction and compression module specifically designed to process the vehicle's kinematic state within the current short-term observation window. For example, a one-way shared network can be a single-layer shared long short-term memory network (Shared LSTM).

[0042] For example, inputting the current operating condition feature sequence into a one-way shared network for feature compression and extracting the state feature vector representing the current instantaneous physical momentum state can be achieved using the following implementation steps: The current working condition feature sequence is input into a Shared LSTM. The feature compression logic of the Shared LSTM uses a single-layer LSTM with a hidden layer dimension of 64, extracting only the hidden state of the last frame of the LSTM. .in It includes the current instantaneous physical momentum state after time-series filtering.

[0043] S5. Generate modulation parameters based on driving style vector, use the modulation parameters to perform linear feature modulation on the state feature vector, and output the modulated state feature vector.

[0044] Understandably, modulation parameters refer to a set of control factors generated by solving a driving style vector that represents global driving tendencies through a specific mapping network (such as parallel fully connected layers).

[0045] Understandably, linear feature modulation refers to a deep information fusion and semantic reshaping mechanism based on feature-level affine transformation. It achieves dynamic intervention of subjective style on objective state through online parameter reconstruction. For example, linear feature modulation can adopt feature-wise linear modulation based on FiLM (Feature-wise Linear Modulation).

[0046] Specifically, modulation parameters are generated based on the driving style vector, and linear feature modulation is performed on the state feature vector using the modulation parameters to output the modulated state feature vector, including: The driving style vector is input into two parallel fully connected layer mapping networks to generate scaling factors and offset factors with the same dimensions as the state feature vector. After performing element-wise multiplication of the state feature vector with the scaling factor, and then adding it with the offset factor, the spatial distribution modulation parameters of the state feature vector are reconstructed, and the modulation state feature vector is obtained.

[0047] For example, generating modulation parameters based on the driving style vector, using the modulation parameters to perform linear feature modulation on the state feature vector, and outputting the modulated state feature vector can be achieved through the following implementation steps: Step 1: A FiLM modulation layer is constructed, which receives a 16-dimensional driving style vector S from the left branch. Through mapping by two parallel fully connected layers, two sets of modulation parameters with dimensions completely consistent with the state feature vector F (64 dimensions) are generated. Step 2: Scaling factor γ and offset factor β, modulated features Follow the formula: ,in This is for element-wise multiplication.

[0048] Understandably, FiLM linear feature modulation enables real-time intervention in spatial distribution. For example, when the driving style vector S is characterized as "aggressive," γ amplifies the weight of "relative speed" in the state features, while β may reduce the threshold of "safe distance" at the feature level, thus allowing the model to produce drastically different responses in the same physical environment.

[0049] S6. Input the driving style vector into the gating network to output the expert routing distribution weights, and input the modulation state feature vector into multiple expert sub-networks set in parallel to output the candidate prediction benchmark value of each expert.

[0050] For example, inputting the driving style vector into a gating network outputs expert routing distribution weights, and inputting the modulation state feature vector into multiple expert sub-networks set in parallel outputs candidate prediction benchmark values ​​for each expert, can be achieved using the following implementation steps: The driving style vector S is synchronously input into a gating network. The gating network adopts a three-layer multilayer perceptron (MLP) architecture, with the middle layer containing 32 neurons and using the ReLU activation function; the terminal layer is normalized using the Softmax operator, and outputs a set of probability distributions. (where n is the number of experts, such as 4). The gating network is essentially a behavior classifier that does not directly participate in the calculation of physical quantities, but determines which "expert opinions" should dominate the current working condition based on style S.

[0051] S7. The multiple candidate prediction baseline values ​​are weighted and aggregated according to the expert routing distribution weights to obtain the predicted value of the vehicle's longitudinal kinematic state.

[0052] Specifically, multiple candidate prediction baseline values ​​are weighted and aggregated according to the expert routing distribution weights to obtain the predicted values ​​of the vehicle's longitudinal kinematic state, including: The gating network includes a multilayer perceptron and a normalization operator. The gating network maps the driving style vector to a probability distribution matrix of the activation of each expert subnetwork, which serves as the expert routing distribution weight. Each expert subnetwork independently performs dimensionality reduction mapping on the modulation state feature vector and outputs a single-dimensional candidate prediction benchmark value. The candidate prediction baseline value output by each expert subnetwork is multiplied by the corresponding expert route distribution weight, and the predicted jerk is output as the predicted value of the vehicle's longitudinal kinematic state.

[0053] For example, the predicted value of the vehicle's longitudinal kinematic state can be obtained by weighting and aggregating multiple candidate prediction benchmark values ​​according to the expert routing distribution weights using the following implementation steps: Step 1: Run N independent Experts (expert subnetworks) in parallel. Each expert subnetwork contains a two-layer MLP structure: Layer 1: 64-dimensional input is mapped to a 64-dimensional intermediate layer, equipped with a Dropout (p=0.2) operator to prevent overfitting; Layer 2: After ReLU activation, it is mapped to a single-dimensional output space. Although the expert subnetwork structure is the same, during training, due to the differentiated allocation of gating network weights, different experts will gradually "evolve" to handle specific styles (e.g., Expert 1 specializes in handling high-speed, smooth conditions, while Expert 2 specializes in handling congested, traffic-following conditions). The system then performs a weighted sum of the predicted outputs of the N experts, calculated using the following formula: Output objective: The final output is the predicted jerk j corresponding to the Pth sampling step, where P is the prediction step size parameter, which can be set according to the data sampling frequency and actual control requirements. For example, in one embodiment, the prediction step size parameter P is set to 15.

[0054] It is understood that embodiments of the present invention also include a global update of the model parameters of the physically mirror-driven style encoder, unidirectional shared network, feature modulation layer network, gating network, and multiple expert sub-networks, the steps of which are as follows: The actual jerk residual values ​​in the traffic flow sample dataset of vehicles are calculated by filtering operators; When constructing training batches, a predetermined proportion of smooth driving samples with actual acceleration residual values ​​lower than the judgment threshold are removed, while dynamic response samples with actual acceleration residual values ​​higher than the judgment threshold are retained, thus reconstructing the non-uniform training data distribution. Calculate the fundamental relative error between the model's predicted output and the true label value; Extract the actual dynamic amplitude from the true label values ​​to establish sample penalty weights, and use the sample penalty weights to weight the relative error to obtain the total loss function; The model parameters of the physically mirror-driven style encoder, one-way shared network, feature modulation layer network, gating network, and multiple expert sub-networks are updated via the full-link gradient backpropagation algorithm based on the total loss function.

[0055] For example, the following implementation steps can be used to globally update the model parameters of the physically mirror-driven style encoder, unidirectional shared network, feature modulation layer network, gating network, and multiple expert sub-networks: Step 1: Perform Savitzky-Golay (SG) filtering on the jerk signal at the tag end. Local polynomial fitting is used to remove sensor noise and extract the high signal-to-noise ratio subjective intent peak. For example, considering that over 90% of highway data consists of mundane driving conditions such as "uniform speed," the system introduces a method based on... Residual filtering operator. Operation: When constructing the training batch, actively filter out approximately 80% of the residuals. It obtains stable driving samples while fully retaining "intent-dense samples" involving high dynamic responses such as lane changes and rapid acceleration and deceleration. This non-uniform sampling forced physical mirror-driven style encoder and unidirectional shared network focus on learning the deviation between "physical benchmark and actual operation" in limited training iterations, fundamentally overcoming the "mean regression" deadlock in long-term prediction of the model; Step 2: Use Huber Loss to guide global optimization. Huber Loss is a loss function used for regression tasks that combines the advantages of mean squared error (MSE) and mean absolute error (MAE), exhibiting good robustness to outliers. First, calculate the basic Huber error. (threshold) =1.0), and then introduce sample weights. Large operations (large Jerk values) are given higher penalty weights, forcing the model to focus on scarce emergency situations; Step 2, Weighted Huber Loss Calculation Formula: To balance robustness to outliers and sensitivity to emergency conditions, the loss function... The calculation is as follows: First, calculate the predicted value. Compared with the true value The basic Huber error between : ; in Set the threshold parameter (to 1.0). Then, to enhance the model's handling of large operations (large... Learning the value, introducing sample weights. (in The final weighted loss function is: ; Step 3: The training process employs the full-link gradient backpropagation algorithm. In each training batch, long-term and short-term feature flows undergo online parameter reconstruction via the FiLM layer, and the corresponding expert subnetworks are dynamically activated by the gating network. The total loss is calculated according to the weighted Huber Loss formula, and the Adam optimizer is used in conjunction with... Weight decay is used to suppress overfitting. Early stopping is introduced: if the loss on the validation set does not decrease for 12 consecutive epochs, the model is considered to have captured stable driving style features, and training is immediately terminated to ensure generalization ability.

[0056] Understandably, the Adam optimizer refers to a gradient descent optimization algorithm widely used in deep learning that can automatically adjust the learning rate according to different parameters while maintaining fast and relatively stable convergence speed.

[0057] This invention acquires the original traffic flow features of vehicles and decomposes these features into historical trajectory time-series data based on a preset historical time window and current operating condition feature sequences based on the current moment, enabling the extraction of long-term behavioral features and instantaneous state features. Furthermore, by inputting the historical trajectory time-series data into a physically mirror-driven style encoder, a theoretical physical reference sequence is generated through the style encoder's dynamic mirror layer. The mirror residual sequence between the actual action sequence corresponding to the historical trajectory time-series data and the theoretical physical reference sequence is calculated, thus filtering out physical noise. Moreover, feature aggregation and dimensionality reduction are performed on the mirror residual sequence to extract driving style vectors representing the subjective operating habits of the vehicle driver. This quantifies the driver's personalized operations into a vector format that the model can understand, eliminating interference from single random operations by the driver. Additionally, the current operating condition feature sequence is input into a unidirectional shared network for feature compression, extracting state feature vectors representing the current instantaneous physical momentum state, which can transform the current complex... The complex driving condition features are mapped to a unified state representation. Furthermore, modulation parameters are generated based on the driving style vector, and these parameters are used to linearly modulate the state feature vector, outputting a modulated state feature vector. This allows the current instantaneous state features to adaptively adjust according to the long-term driving styles of different drivers. Additionally, the driving style vector is input into a gating network to output expert route distribution weights, and the modulated state feature vector is input into multiple parallel expert subnetworks to output candidate prediction benchmark values ​​for each expert. A hybrid expert prediction mechanism is constructed through the gating network and multiple expert subnetworks, enabling the model to assign different expert prediction paths for different driving styles. Finally, the multiple candidate prediction benchmark values ​​are weighted and aggregated according to the expert route distribution weights to obtain the predicted value of the vehicle's longitudinal kinematic state. Adaptive weighted aggregation of multiple candidate prediction benchmark values ​​using expert route weights allows the prediction results to be dynamically adjusted according to driving style and current driving conditions, improving the timeliness and accuracy of vehicle condition prediction.

[0058] Reference Figure 2 The diagram shown is an example of the overall algorithm flow of a vehicle longitudinal intent and operating condition prediction method provided in an embodiment of the present invention.

[0059] like Figure 3 The diagram shown is a model network architecture diagram of a vehicle longitudinal intent and operating condition prediction method provided in an embodiment of the present invention.

[0060] like Figure 4 The diagram shown is a functional block diagram of a vehicle longitudinal intention and operating condition prediction device provided in an embodiment of the present invention.

[0061] The vehicle longitudinal intent and operating condition prediction device 100 of the present invention can be installed in an electronic device. Depending on the functions implemented, the vehicle longitudinal intent and operating condition prediction device 100 may include a driving style vector acquisition module 101, a state feature vector extraction module 102, and a vehicle longitudinal intent and operating condition prediction module 103.

[0062] The module described in this invention can also be called a unit, which refers to a series of computer program segments that can be executed by the processor of an electronic device and can perform a fixed function, and are stored in the memory of the electronic device.

[0063] In this embodiment, the functions of each module / unit are as follows: The driving style vector acquisition module 101 is used to acquire the original traffic flow features of the vehicle and split the original traffic flow features into historical trajectory time series data based on a historical preset time window and current operating condition feature sequence based on the current moment; input the historical trajectory time series data into the style encoder driven by the physical mirror, generate a theoretical physical reference sequence through the dynamic mirror layer of the style encoder, and calculate the mirror residual sequence between the actual action sequence corresponding to the historical trajectory time series data and the theoretical physical reference sequence; perform feature aggregation and dimensionality reduction on the mirror residual sequence to extract the driving style vector representing the subjective operating habits of the vehicle driver.

[0064] The state feature vector extraction module 102 is used to input the current working condition feature sequence into a one-way shared network for feature compression and extract the state feature vector representing the current instantaneous physical momentum state.

[0065] The vehicle longitudinal intent and operating condition prediction module 103 is used to generate modulation parameters based on the driving style vector, perform linear feature modulation on the state feature vector using the modulation parameters, and output the modulated state feature vector; input the driving style vector into the gating network to output the expert routing distribution weights, input the modulated state feature vector into multiple expert sub-networks set in parallel to output the candidate prediction benchmark value of each expert; and perform weighted aggregation of multiple candidate prediction benchmark values ​​according to the expert routing distribution weights to obtain the predicted value of the vehicle's longitudinal kinematic state.

[0066] like Figure 5 The diagram shown is a structural schematic of an electronic device for implementing a method for predicting the longitudinal intent and operating conditions of a vehicle, according to an embodiment of the present invention.

[0067] The electronic device may include a processor 10, a memory 11, a communication bus 12, and a communication interface 13. It may also include a computer program stored in the memory 11 and executable on the processor 10, such as a method program for predicting vehicle longitudinal intent and operating conditions.

[0068] In some embodiments, the processor 10 may be composed of integrated circuits, such as a single packaged integrated circuit or multiple integrated circuits with the same or different functions, including combinations of one or more central processing units (CPUs), microprocessors, digital processing chips, graphics processors, and various control chips. The processor 10 is the control unit of the electronic device, connecting various components of the entire electronic device through various interfaces and lines. It executes programs or modules stored in the memory 11 (e.g., executing a vehicle longitudinal intent and operating condition prediction method program), and calls data stored in the memory 11 to perform various functions of the electronic device and process data.

[0069] The memory 11 includes at least one type of readable storage medium, including flash memory, portable hard drive, multimedia card, card-type memory (e.g., SD or DX memory), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 11 can be an internal storage unit of an electronic device, such as a portable hard drive. In other embodiments, the memory 11 can be an external storage device of the electronic device, such as a plug-in portable hard drive, smart media card (SMC), secure digital (SD) card, flash card, etc. Furthermore, the memory 11 can include both internal and external storage units of the electronic device. The memory 11 can be used not only to store application software and various types of data installed on the electronic device, such as the code of a vehicle longitudinal intent and operating condition prediction method program, but also to temporarily store data that has been output or will be output.

[0070] The communication bus 12 can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This bus can be divided into an address bus, a data bus, a control bus, etc. The bus is configured to enable communication between the memory 11 and at least one processor 10, etc.

[0071] The communication interface 13 is used for communication between the aforementioned electronic device and other devices, including a network interface and a user interface. Optionally, the network interface may include a wired interface and / or a wireless interface (such as a Wi-Fi interface, Bluetooth interface, etc.), typically used to establish communication connections between the electronic device and other electronic devices. The user interface may be a display, an input unit (such as a keyboard), or optionally, a standard wired or wireless interface. Optionally, in some embodiments, the display may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen, etc. The display may also be appropriately referred to as a screen or display unit, used to display information processed in the electronic device and to display a visual user interface.

[0072] Figure 5 Only electronic devices with components are shown; it will be understood by those skilled in the art that... Figure 5 The structure shown does not constitute a limitation on the electronic device and may include fewer or more components than shown, or combine certain components, or have different component arrangements.

[0073] For example, although not shown, the electronic device may also include a power supply (such as a battery) to power the various components. Preferably, the power supply can be logically connected to the at least one processor 10 through a power management device, thereby enabling functions such as charging management, discharging management, and power consumption management. The power supply may also include one or more DC or AC power supplies, recharging devices, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components. The electronic device may also include various sensors, Bluetooth modules, Wi-Fi modules, etc., which will not be described in detail here.

[0074] It should be understood that the embodiments described are for illustrative purposes only and are not limited to this structure in the scope of the patent application.

[0075] The memory 11 in the electronic device stores a vehicle longitudinal intention and operating condition prediction method program, which is a combination of multiple instructions. When run in the processor 10, it can achieve the following: Obtain the original traffic flow characteristics of vehicles, and split the original traffic flow characteristics into historical trajectory time series data based on historical preset time windows and current operating condition characteristic sequence based on the current moment; Historical trajectory time series data is input into a style encoder driven by a physical mirror. The theoretical physical reference sequence is generated through the dynamic mirror layer of the style encoder, and the mirror residual sequence between the actual action sequence corresponding to the historical trajectory time series data and the theoretical physical reference sequence is calculated. Feature aggregation and dimensionality reduction are performed on the mirror residual sequence to extract driving style vectors that represent the subjective operating habits of the vehicle driver. The current operating condition feature sequence is input into a one-way shared network for feature compression, and a state feature vector representing the current instantaneous physical momentum state is extracted. Modulation parameters are generated based on driving style vectors, and linear feature modulation is performed on the state feature vector using the modulation parameters to output the modulated state feature vector. The driving style vector is input into the gating network to output the expert routing distribution weights, and the modulation state feature vector is input into multiple expert sub-networks set in parallel to output the candidate prediction benchmark value of each expert. The predicted value of the vehicle's longitudinal kinematic state is obtained by weighting and aggregating multiple candidate prediction benchmark values ​​according to the expert route distribution weights.

[0076] Specifically, the specific implementation method of the processor 10 for the above instructions can be referred to the description of the relevant steps in the corresponding embodiment of the accompanying drawings, and will not be repeated here.

[0077] Furthermore, if the modules / units integrated into the electronic device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. The computer-readable storage medium can be volatile or non-volatile. For example, the computer-readable medium may include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, or a read-only memory (ROM).

[0078] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor of an electronic device, can perform the following: Obtain the original traffic flow characteristics of vehicles, and split the original traffic flow characteristics into historical trajectory time series data based on historical preset time windows and current operating condition characteristic sequence based on the current moment; Historical trajectory time series data is input into a style encoder driven by a physical mirror. The theoretical physical reference sequence is generated through the dynamic mirror layer of the style encoder, and the mirror residual sequence between the actual action sequence corresponding to the historical trajectory time series data and the theoretical physical reference sequence is calculated. Feature aggregation and dimensionality reduction are performed on the mirror residual sequence to extract driving style vectors that represent the subjective operating habits of the vehicle driver. The current operating condition feature sequence is input into a one-way shared network for feature compression, and a state feature vector representing the current instantaneous physical momentum state is extracted. Modulation parameters are generated based on driving style vectors, and linear feature modulation is performed on the state feature vector using the modulation parameters to output the modulated state feature vector. The driving style vector is input into the gating network to output the expert routing distribution weights, and the modulation state feature vector is input into multiple expert sub-networks set in parallel to output the candidate prediction benchmark value of each expert. The predicted value of the vehicle's longitudinal kinematic state is obtained by weighting and aggregating multiple candidate prediction benchmark values ​​according to the expert route distribution weights.

[0079] In the several embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.

[0080] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0081] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0082] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0083] Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be embraced within the invention. No appended diagram markings in the claims should be construed as limiting the scope of the claims.

[0084] The blockchain referred to in this invention is a novel application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and encryption algorithms. Essentially, a blockchain is a decentralized database, a chain of data blocks linked together using cryptographic methods. Each data block contains information about a batch of network transactions, used to verify the validity of the information (anti-counterfeiting) and generate the next block. A blockchain can include an underlying blockchain platform, a platform product service layer, and an application service layer.

[0085] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence (AI) refers to the theories, methods, technologies, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0086] Furthermore, it is clear that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units or devices recited in a system claim may also be implemented by a single unit or device through software or hardware. The terms "first," "second," etc., are used to indicate names and do not indicate any specific order.

[0087] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A vehicle longitudinal intention and working condition prediction method, characterized in that, The method includes: Obtain the original traffic flow characteristics of vehicles, and split the original traffic flow characteristics into historical trajectory time series data based on historical preset time windows and current operating condition characteristic sequence based on the current moment; Historical trajectory time series data is input into a style encoder driven by a physical mirror. The theoretical physical reference sequence is generated through the dynamic mirror layer of the style encoder, and the mirror residual sequence between the actual action sequence corresponding to the historical trajectory time series data and the theoretical physical reference sequence is calculated. Feature aggregation and dimensionality reduction are performed on the mirror residual sequence to extract driving style vectors that represent the subjective operating habits of the vehicle driver. The current operating condition feature sequence is input into a one-way shared network for feature compression, and a state feature vector representing the current instantaneous physical momentum state is extracted. Modulation parameters are generated based on driving style vectors, and linear feature modulation is performed on the state feature vector using the modulation parameters to output the modulated state feature vector. The driving style vector is input into the gating network to output the expert routing distribution weights, and the modulation state feature vector is input into multiple expert sub-networks set in parallel to output the candidate prediction benchmark value of each expert. The predicted value of the vehicle's longitudinal kinematic state is obtained by weighting and aggregating multiple candidate prediction benchmark values ​​according to the expert route distribution weights.

2. The vehicle longitudinal intent and operating condition prediction method as described in claim 1, characterized in that, The process of generating a theoretical physical reference sequence through the dynamic mirror layer of the style encoder and calculating the mirror residual sequence between the actual action sequence corresponding to the historical trajectory time series data and the theoretical physical reference sequence includes: Extract the actual jerk sequence from historical trajectory time series data using indexing operators as the actual action sequence; By inputting historical trajectory time series data into a linear neuron containing a physical boundary truncation operator, a theoretical physical benchmark sequence that conforms to the consensus boundary of dynamic physics and is within a preset value range is obtained. The actual action sequence and the theoretical physical reference sequence are input into the subtraction node to calculate the mirror residual sequence.

3. The vehicle longitudinal intent and operating condition prediction method as described in claim 1, characterized in that, The process of feature aggregation and dimensionality reduction of the mirror residual sequence to extract a driving style vector representing the subjective operating habits of the vehicle driver includes: The enhanced feature matrix is ​​obtained by concatenating the mirror residual sequence with the original historical trajectory time series data; The enhanced feature matrix is ​​input into a bidirectional long short-term memory network that is regularized by a layer normalization function; The evolution trend features of driving behavior are extracted using the forward network of the bidirectional long short-term memory network, and the backward consistency features of driving behavior are extracted using the backward network of the bidirectional long short-term memory network. The evolution trend features and the inverse consistency features are fused to obtain the initial fused features. The initial fused features are then input into a mapping network consisting of fully connected layers, nonlinear activation functions, and layer normalization functions to perform low-dimensional mapping and obtain the driving style vector.

4. The vehicle longitudinal intent and operating condition prediction method as described in claim 1, characterized in that, The process of generating modulation parameters based on driving style vectors, using these modulation parameters to perform linear feature modulation on the state feature vector, and outputting a modulated state feature vector includes: The driving style vector is input into two parallel fully connected layer mapping networks to generate scaling factors and offset factors with the same dimensions as the state feature vector. After performing element-wise multiplication of the state feature vector with the scaling factor, and then adding it with the offset factor, the spatial distribution modulation parameters of the state feature vector are reconstructed, and the modulation state feature vector is obtained.

5. The vehicle longitudinal intent and operating condition prediction method as described in claim 1, characterized in that, The step of weighting and aggregating multiple candidate prediction benchmark values ​​according to the expert routing distribution weights to obtain the predicted value of the vehicle's longitudinal kinematic state includes: The gating network includes a multilayer perceptron and a normalization operator. The gating network maps the driving style vector to a probability distribution matrix of the activation of each expert subnetwork, which serves as the expert routing distribution weight. Each expert subnetwork independently performs dimensionality reduction mapping on the modulation state feature vector and outputs a single-dimensional candidate prediction benchmark value. The candidate prediction baseline value output by each expert subnetwork is multiplied by the corresponding expert route distribution weight, and the predicted jerk is output as the predicted value of the vehicle's longitudinal kinematic state.

6. The vehicle longitudinal intent and operating condition prediction method as described in claim 1, characterized in that, Before obtaining the original traffic flow characteristics of vehicles, the method further includes: Acquire raw data including the vehicle's longitudinal velocity, longitudinal acceleration, actual jerk, and relative distance and relative velocity representing the relationship between the vehicle and the lead vehicle; The original data is input into a local polynomial least squares fitting filter. After smoothing and denoising the high-frequency noise of the third derivative signal in the local polynomial least squares fitting filter, the original traffic flow characteristics are obtained.

7. A vehicle longitudinal intention and operating condition prediction device, characterized in that, The apparatus is used to implement the vehicle longitudinal intent and operating condition prediction method as described in any one of claims 1 to 6, the apparatus comprising: The driving style vector acquisition module is used to acquire the original traffic flow features of vehicles and split the original traffic flow features into historical trajectory time series data based on a historical preset time window and current operating condition feature sequence based on the current moment. The historical trajectory time series data is input into the style encoder driven by physical mirror. The theoretical physical reference sequence is generated through the dynamic mirror layer of the style encoder, and the mirror residual sequence between the actual action sequence corresponding to the historical trajectory time series data and the theoretical physical reference sequence is calculated. The mirror residual sequence is subjected to feature aggregation and dimensionality reduction to extract the driving style vector representing the subjective operating habits of the vehicle driver. The state feature vector extraction module is used to input the current working condition feature sequence into a one-way shared network for feature compression and extract the state feature vector representing the current instantaneous physical momentum state. The vehicle longitudinal intent and operating condition prediction module is used to generate modulation parameters based on the driving style vector, perform linear feature modulation on the state feature vector using the modulation parameters, and output the modulated state feature vector. The driving style vector is input into the gating network to output the expert routing distribution weights, and the modulated state feature vector is input into multiple expert sub-networks set in parallel to output the candidate prediction benchmark value of each expert. The multiple candidate prediction benchmark values ​​are weighted and aggregated according to the expert routing distribution weights to obtain the predicted value of the vehicle's longitudinal kinematic state.

8. An electronic device, characterized in that, The electronic device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the vehicle longitudinal intent and operating condition prediction method as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the vehicle longitudinal intent and operating condition prediction method as described in any one of claims 1 to 6.

Images