An unmanned vehicle-oriented pedestrian gesture adaptive understanding method

Abstract

This invention provides an adaptive pedestrian gesture understanding method for autonomous driving, including acquiring raw visual data; extracting initial features from the raw visual data and performing visual encoding; self-supervised decoupling; cross-scene consistency constraints; inferring gesture intent; and model training. This invention proposes an adaptive pedestrian gesture understanding method for autonomous driving, using scene labels as weakly supervised signals to learn a scene-independent semantic subspace and a scene-related style / intent subspace, achieving adaptive gesture understanding in different scenarios.

Description

Technical Field

[0001] This invention relates to the technical field of autonomous driving and environmental perception, and in particular to an adaptive understanding method for pedestrian gestures for unmanned driving. Background Technology

[0002] Gesture recognition technology, as an important means of human-computer interaction, has developed rapidly in recent years, and is evolving towards greater accuracy, naturalness, and intelligence. The current state of gesture recognition technology is as follows.

[0003] From static to dynamic: The system can not only recognize fixed postures, but also accurately capture and understand continuous hand movement trajectories, enabling richer interactions.

[0004] Vision-driven solutions are the mainstream: Computer vision-based solutions (using ordinary cameras or depth sensors) have become the standard, eliminating the need for users to wear additional equipment and greatly improving convenience and versatility. Key technologies include hand target detection, key point pose estimation, and dynamic tracking.

[0005] AI Deep Empowerment: Deep learning and artificial intelligence algorithms are the core of improving recognition accuracy and robustness, enabling the system to adapt to different lighting, backgrounds and individual differences.

[0006] Multimodal fusion: The future trend is to fuse visual recognition with other biometrics (such as palm prints and veins) or sensor data to achieve higher security and more accurate intent understanding.

[0007] In autonomous driving technology, pedestrian gesture recognition is a crucial element for achieving safe and natural human-vehicle interaction. Autonomous driving systems primarily rely on multi-sensor fusion and artificial intelligence algorithms to recognize pedestrian gestures. Vehicle-mounted LiDAR, high-definition cameras, and millimeter-wave radar work together to construct a 3D environment model. Cameras capture the visual information of gestures, LiDAR provides precise spatial location and motion trajectory, millimeter-wave radar can penetrate adverse weather conditions, and deep learning models analyze pedestrian postures, improving detection speed and accuracy. However, the requirements for pedestrian gestures in autonomous driving systems go beyond mere recognition, such as traffic police gesture recognition and pedestrian gesture recognition. A more important goal is for vehicles to understand non-verbal traffic intentions such as waving or pointing, thereby enabling safe responses.

[0008] Autonomous driving systems encounter various pedestrian interaction scenarios during operation, and the same gesture (such as waving) may represent completely opposite intentions in different contexts. Therefore, recognizing gestures in isolation is insufficient; they must be understood within the context of the scenario. Thus, adaptive understanding of pedestrian gestures in autonomous driving scenarios becomes crucial.

[0009] Chinese invention patent application CN110320911A discloses an unmanned vehicle control method, device, unmanned vehicle, and storage medium. The method includes: acquiring pedestrian images containing pedestrian body movements; performing posture recognition on the pedestrian body movements in the pedestrian images to obtain pedestrian posture information; determining whether the posture represented by the pedestrian posture information is a preset target posture; if the posture represented by the pedestrian posture information is the preset target posture, obtaining the target pedestrian's position information; determining a driving path based on the target pedestrian's position information, and driving according to the driving path. The drawback of this method is that it only performs a fixed classification for the gestures themselves, failing to distinguish between the essential semantics of the gestures and the scene-related intentions, and failing to model the scene invariance of semantics. Summary of the Invention

[0010] To address the aforementioned technical issues, this invention proposes an adaptive understanding method for pedestrian gestures in autonomous driving. By using scene labels as weak supervision signals, a scene-independent semantic subspace and a scene-related style / intent subspace are learned to achieve adaptive understanding of gestures in different scenarios.

[0011] This invention provides a pedestrian gesture adaptive understanding method for autonomous driving, including acquiring raw visual data and the following steps:

[0012] Step 1: Extract initial features from the raw visual data and perform visual encoding;

[0013] Step 2: Self-supervised decoupling;

[0014] Step 3: Cross-scenario consistency constraints;

[0015] Step 4: Infer the intention behind the gesture;

[0016] Step 5: Model training.

[0017] Preferably, step 1 includes the following sub-steps:

[0018] Step 11: Perform gesture video encoding;

[0019] Step 12: Perform scene coding.

[0020] In any of the above schemes, step 11 preferably includes inputting a gesture video containing T frames, with a single frame size of T×H×W×3, and using a 3D convolutional neural network to extract spatiotemporal features.

[0021] h v =E v (Video)

[0022] Among them, h vE is the initial feature vector of the video. v It is a gesture video encoder.

[0023] In any of the above schemes, step 12 preferably includes inputting a single-frame image Scene_Image containing scene context, which is then encoded using a 2D convolutional neural network.

[0024] h s =E s (Scene_Image)

[0025] Among them, h s E is the scene feature vector. s This is a scene encoder.

[0026] In any of the above solutions, step 2 preferably includes the following sub-steps:

[0027] Step 21: Perform feature decoupling;

[0028] Step 22: Perform feature reconstruction;

[0029] Step 23: Adversarial scenario obfuscation calculation.

[0030] Preferably, in any of the above schemes, the decoupler D contains two mapping functions: a semantic mapping function f sem ( ) and style mapping function f sty ( ).

[0031] In any of the above schemes, the preferred embodiment is that the calculation formula for the feature decoupling is:

[0032] z sem =f sem (h v )=W sem h v +b sem

[0033] z sty =f sty (h v )=W sty h v +b sty

[0034] Among them, z sem and z sty These are the decoupled semantic feature vector and style feature vector, respectively, W sem and W sty b represents the weights of the feature vectors. sem and b sty This represents the bias of the eigenvectors.

[0035] In any of the above embodiments, step 22 preferably includes starting from z sem and z sty The original video features h can be reconstructed from the data. v The calculation formula is:

[0036]

[0037] The formula for calculating the reconstruction loss is:

[0038]

[0039] Where g() is the reconstruction function.

[0040] In any of the above schemes, step 23 preferably includes introducing a scene classifier Cscene to predict scene labels from the decoupled features.

[0041] In any of the above solutions, step 23 preferably includes the following sub-steps:

[0042] Step 231: Using semantic feature vector z sem Calculate the scene classification loss L adv_sem ;

[0043] Step 232: Using the style feature vector z sty Calculate the scene classification loss L sty_scene And minimize the scene classification loss L sty_scene .

[0044] In any of the above schemes, the preferred option is the scene classification loss L. adv_sem The calculation formula is:

[0045] L adv_sem =CE(Cscene(z sem ),y scene )

[0046] Among them, y scene represents the real scene label, CE() represents the cross-entropy loss, and Cscene represents the scene classifier.

[0047] In any of the above schemes, the preferred option is the scene classification loss L. sty_scene The calculation formula is:

[0048] L sty_scene =CE(Cscene(z sty ),y scene ).

[0049] In any of the above schemes, step 3 preferably includes self-supervised learning using the constraint that the semantics of the same gesture should be the same in different scenarios.

[0050] In any of the above schemes, it is preferable that the dataset contains sample pairs of the same gesture belonging to the same semantics in different scenarios, and the semantic features of the sample pair (i,j) are... and When elements are close in the feature space, the contrastive loss function is...

[0051] .

[0052] In any of the above schemes, step 4 includes combining the decoupled features with scene features after completing representation learning to perform the final intent classification.

[0053] In any of the above solutions, step 4 preferably includes the following sub-steps:

[0054] Step 41: Perform feature fusion;

[0055] Step 42: Perform intent classification.

[0056] In any of the above schemes, step 41 preferably involves concatenating the decoupled semantic features, style features, and scene features to form a comprehensive feature input.

[0057] f infer =[z sem ;z sty ;h s ]

[0058] Among them, f infer It is a feature obtained by concatenating semantic features, style features, and scene features.

[0059] In any of the above schemes, step 42 preferably includes f infer Feed into a multilayer perceptron classifier (MLP) intent The final output is the probability of intent.

[0060] p = Softmax(MLP) intent (f infer ))

[0061] Where p is the probability distribution over all possible intentions.

[0062] In any of the above solutions, step 5 preferably includes the following sub-steps:

[0063] Step 51: Prepare a dataset containing gesture videos, corresponding scene images, gesture intent labels, and scene labels;

[0064] Step 52: Train the model using the dataset.

[0065] In any of the above schemes, step 52 preferably includes performing self-supervised representation learning on the model to minimize the total loss function of the jointly optimized model:

[0066] L total =λ1L rec -λ2L adv_sem +λ3L sty_scene +λ4L cons

[0067] Among them, L total Let λ1, λ2, λ3, and λ4 be the total loss function, and let λ1, λ2, λ3, and λ4 be the weights for different losses.

[0068] In any of the above schemes, step 52 preferably further includes fixing the encoder and decoupler, and training the final intent inferrer MLP using data with gesture intent labels. intent .

[0069] This invention proposes an adaptive understanding method for pedestrian gestures for autonomous driving, providing a practical and innovative path for autonomous driving systems to deeply understand pedestrian gestures in the absence of abundant labeled data. Attached Figure Description

[0070] Figure 1 This is a flowchart of a preferred embodiment of the pedestrian gesture adaptive understanding method for autonomous driving according to the present invention. Detailed Implementation

[0071] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0072] Example 1

[0073] like Figure 1 As shown, an adaptive pedestrian gesture understanding method for autonomous driving, in step 1000, obtains raw visual data, and also includes the following steps:

[0074] Step 1100 involves extracting initial features from the original visual data and performing visual encoding, including the following sub-steps:

[0075] Step 1110 involves performing gesture video encoding, which includes inputting a gesture video containing T frames, with each frame having a size of T×H×W×3, and using a 3D convolutional neural network to extract spatiotemporal features.

[0076] h v =E v (Video)

[0077] Among them, h v E is the initial feature vector of the video.v It is a gesture video encoder.

[0078] Step 1120 involves scene encoding, which takes a single-frame image (Scene_Image) containing the scene context as input and encodes it using a 2D convolutional neural network.

[0079] h s =E s (Scene_Image)

[0080] Among them, h s E is the scene feature vector. s This is a scene encoder.

[0081] Execute step 1200, self-supervised decoupling. The decoupler D contains two mapping functions: semantic mapping function f. sem ( ) and style mapping function f sty ( ). Step 1200 includes the following sub-steps:

[0082] Execute step 1210 to perform feature decoupling. The calculation formula for feature decoupling is as follows:

[0083] z sem =f sem (h v )=W sem h v +b sem

[0084] z sty =f sty (h v )=W sty h v +b sty

[0085] Among them, z sem and z sty These are the decoupled semantic feature vector and style feature vector, respectively, W sem and W sty b represents the weights of the feature vectors. sem and b sty This represents the bias of the eigenvectors.

[0086] Execute step 1220 to reconstruct features, including from z sem and z sty The original video features h can be reconstructed from the data. v The calculation formula is:

[0087]

[0088] The formula for calculating the reconstruction loss is:

[0089]

[0090] Where g() is the reconstruction function.

[0091] Step 1230, adversarial scene obfuscation calculation, includes introducing a scene classifier Cscene to predict scene labels from the decoupled features, and includes the following sub-steps:

[0092] Execute step 1231, using the semantic feature vector z sem Calculate the scene classification loss L adv_sem The formula for calculating is:

[0093] L adv_sem =CE(Cscene(z sem ),y scene )

[0094] Among them, y scene represents the real scene label, CE() represents the cross-entropy loss, and Cscene represents the scene classifier.

[0095] Execute steps 1232, using the style feature vector z sty Calculate the scene classification loss L sty_scene And minimize the scene classification loss L sty_scene The calculation formula is:

[0096] L sty_scene =CE(Cscene(z sty ),y scene ).

[0097] Step 1300 is executed, which involves cross-scenario consistency constraints, including self-supervised learning using the constraint that the semantics of the same gesture should be the same in different scenarios.

[0098] The dataset contains sample pairs of the same gesture belonging to the same semantic meaning in different scenarios. The semantic features of this sample pair (i,j) are... and When elements are close in the feature space, the contrastive loss function is...

[0099] .

[0100] Step 1400 involves inferring the gesture intent, which includes combining the decoupled features with scene features after completing representation learning to perform the final intent classification. This includes the following sub-steps:

[0101] Step 1410 involves feature fusion, which includes concatenating the decoupled semantic features, style features, and scene features to form a comprehensive feature input.

[0102] f infer =[z sem ;z sty ;h s ]

[0103] Among them, f infer It is a feature obtained by concatenating semantic features, style features, and scene features.

[0104] Perform step 1420 to classify intents, including classifying f infer Feed into a multilayer perceptron classifier (MLP) intent The final output is the probability of intent.

[0105] p = Softmax(MLP) intent (f infer ))

[0106] Where p is the probability distribution over all possible intentions.

[0107] Execute step 1500, model training, which includes the following sub-steps:

[0108] Perform step 1510 to prepare a scene dataset containing gesture videos, corresponding scene images, gesture intent labels, and scene labels;

[0109] Execute step 1520, train the model using the scene dataset, including self-supervised representation learning of the model, and minimize the total loss function of the jointly optimized model as follows:

[0110] L total =λ1L rec -λ2L adv_sem +λ3L sty_scene +λ4L cons

[0111] Among them, L total Let λ1, λ2, λ3, and λ4 be the total loss function, and let λ1, λ2, λ3, and λ4 be the weights for different losses.

[0112] With a fixed encoder and decoupler, the final intent inferrer MLP is trained using data with gesture intent labels. intent .

[0113] During operation, autonomous vehicles use pre-trained models to determine the intentions behind pedestrian gestures.

[0114] Example 2

[0115] Autonomous driving encompasses a wide range of scenarios, and autonomous driving systems encounter various pedestrian interaction scenarios during operation. Even with the same gesture, the intended meaning can differ significantly across different scenarios. Therefore, recognizing gestures in isolation is insufficient; it is essential to understand them within the context of the scenario. This invention aims to adaptively understand and infer the intent of pedestrian gestures in different autonomous driving scenarios. First, this invention leverages common sense: the semantics of gestures should be stable across different scenarios, while their intent and style are strongly correlated with the scenario. Based on this common sense, a cross-scenario consistency constraint is designed, using scenario labels as weakly supervised signals to guide feature separation. Through adversarial training, the model learns two subspaces: a scenario-independent semantic subspace and a scenario-dependent style / intent subspace. Finally, the decoupled semantic features, style features, and scenario features are concatenated and fed into a multilayer perceptron classifier to obtain the intent category, thereby achieving adaptive, scenario-based intent understanding of the same gesture.

[0116] The four core steps of this invention are: visual encoding, self-supervised decoupling, cross-scene consistency constraints, and scene-based gesture intent inference.

[0117] 1. Visual encoding

[0118] Extract initial features from raw visual data.

[0119] (1) Gesture video encoding

[0120] The input consists of a gesture video containing T frames, with each frame having a size of T×H×W×3. A 3D convolutional neural network is used to extract spatiotemporal features.

[0121] h v =E v (Video)

[0122] h v It is the initial feature vector of the video, E v It is a gesture video encoder.

[0123] (2) Scene coding

[0124] The input is a single-frame image Scene_Image containing the scene context, which is encoded using a 2D convolutional neural network.

[0125] h s =E s (Scene_Image)

[0126] h s It is a scene feature vector, E s This is a scene encoder.

[0127] 2. Self-supervised decoupling

[0128] Decoupler D contains two mapping functions: a semantic mapping function f sem ( ) and style mapping function f sty ( ).

[0129] To achieve effective decoupling, adversarial training is introduced, where semantic features should contain as much gesture category information as possible, but as little style information as possible; and vice versa.

[0130] Step 1: Feature Decoupling

[0131] z sem =f sem (h v )=W sem h v +b sem

[0132] z sty =f sty (h v )=W sty h v +b sty

[0133] Where z sem and z sty These are the decoupled semantic feature vector and style feature vector, respectively. sem and W sty The weights represent the eigenvectors; b sem and b sty This represents the bias of the eigenvector.

[0134] Step 2: Refactoring

[0135] To prevent information loss, it is required to start from z sem and z sty It can reconstruct the original video features hv.

[0136]

[0137] Where g() is the reconstruction function.

[0138] The reconstruction loss is:

[0139]

[0140] Step 3: Adversarial Scenario Obfuscation Calculation

[0141] We introduce a scene classifier Cscene, whose task is to predict scene labels from the decoupled features.

[0142] Semantic feature z semIt should not contain any scene information. Therefore, Cscene uses z sem When used as input, its predictive ability is as poor as possible, i.e., it cannot be correctly classified.

[0143] Style characteristics z sty Scene information should be included because style and intent are scene-dependent. Therefore, Cscene uses z... sty When used as input, its predictive power should be as strong as possible.

[0144] To achieve this, do the following:

[0145] (1) Through semantic feature z sem Calculate the scene classification loss L adv_sem .

[0146] L adv_sem =CE(Cscene(z sem ),y scene )

[0147] Where y scene These are the real-world scene labels, and CE() is the cross-entropy loss. During training, z sem It will increase L adv_sem The direction of the update.

[0148] (2) Through style feature z sty Calculate the scene classification loss L sty_scene And minimize it, encouraging style features to learn scene-related information.

[0149] L sty_scene =CE(Cscene(z sty ),y scene )

[0150] Through this adversarial setting, the model is forced to adapt to style features z. sty The system learns context-related information, thereby ensuring the semantic feature z. sem Scene invariance.

[0151] 3. Cross-scenario consistency constraints

[0152] Self-supervised learning is performed using the constraint that "the semantics of the same gesture should be the same in different scenarios".

[0153] The dataset contains sample pairs where the same gesture in different scenarios belongs to the same semantic meaning. For such a sample pair (i,j), their semantic features... and The distances in the feature space should be very short. The contrastive loss function is as follows:

[0154]

[0155] This loss term enhances the cross-scene stability of semantic features.

[0156] 4. Inference of gesture intention

[0157] After completing representation learning, the decoupled features are combined with scene features to perform the final intent classification.

[0158] (1) Feature fusion

[0159] The decoupled semantic features, style features, and scene features are concatenated to form a comprehensive feature input.

[0160] f infer =[z sem ;z sty ;h s ]

[0161] (2) Intention Classification

[0162] f infer Feed into a multilayer perceptron classifier (MLP) intent The final output is the probability of the intent.

[0163] p = Softmax(MLP) intent (f infer ))

[0164] Where p is the probability distribution over all possible intentions.

[0165] 5. Method Implementation (Model Training)

[0166] (1) Dataset preparation

[0167] Prepare a dataset containing gesture videos, corresponding scene images, gesture intent labels, and scene labels.

[0168] (2) Training process

[0169] Phase 1: Self-supervised representation learning

[0170] Minimize the following joint optimization loss function:

[0171] L total =λ1L rec -λ2L adv_sem +λ3L sty_scene +λ4L cons

[0172] Phase Two: Intent Classification

[0173] With a fixed encoder and decoupler, the final intent inferrer MLP is trained using data with gesture intent labels. intent .

[0174] This invention proposes an explicit modeling of the scene invariance of gesture semantics. The semantics of gestures should be stable across different scenes, while their intent and style are strongly correlated with the scene. By modeling the prior of "scene invariance of gesture semantics" as a cross-scene consistency constraint, the learned representations become more essential and general.

[0175] This invention proposes a scene-based self-supervised decoupling mechanism. Utilizing readily available scene labels, the scene serves as a supervisory signal, driving the model to learn complex internal representations, including a scene-independent semantic subspace and a scene-related style / intent subspace. Feature decoupling based on visual data is achieved through adversarial training and consistency constraints.

[0176] This invention can be more easily extended to different countries and cultural backgrounds, simply by collecting local video and scene data.

[0177] This method provides a practical and innovative path for autonomous driving systems to gain a deep understanding of pedestrian gestures, even in the absence of abundant labeled data.

[0178] Example 3

[0179] The same gesture can have different meanings in different contexts.

[0180] Autonomous driving systems encounter various pedestrian interaction scenarios during operation, and the same gesture (such as waving) may represent completely opposite intentions in different scenarios. Therefore, recognizing gestures in isolation is insufficient; it is essential to understand them within the context of the scenario.

[0181] Scenario A (Uncontrolled intersection): Pedestrians may make gestures such as "waving to signal vehicles to go first" or "reaching out to indicate that they want to cross first".

[0182] Scenario B (pedestrian crossing): Pedestrians give a thumbs-up to express their gratitude or wave to indicate that vehicles can cross.

[0183] Scenario C (parking lot / residential area): Pedestrians may make complex gestures to "direct the reversing".

[0184] Scenario D: At an intersection, the semantic meaning of "waving" + the style of "facing vehicles" + the setting of "intersection" = "give way to vehicles". Meanwhile, on the roadside, the semantic meaning of "waving" + the style of "facing into the distance" + the setting of "roadside" = "call a taxi".

[0185] To better understand this invention, specific embodiments have been described in detail above, but these are not intended to limit the invention. Any simple modifications made to the above embodiments based on the technical essence of this invention still fall within the scope of this invention. Each embodiment in this specification focuses on its differences from other embodiments; similar or identical parts between embodiments can be referred to mutually. For system embodiments, since they basically correspond to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

Claims

1. A pedestrian gesture adaptive understanding method for autonomous driving, comprising acquiring raw visual data, characterized in that, It also includes the following steps: Step 1: Extract initial features from the raw visual data and perform visual encoding; Step 2: Self-supervised decoupling; Step 3: Cross-scenario consistency constraints; Step 4: Infer the intention behind the gesture; Step 5: Model training.

2. The pedestrian gesture adaptive understanding method for autonomous driving as described in claim 1, characterized in that, Step 1 includes the following sub-steps: Step 11: Perform gesture video encoding, including input containing... T A frame-by-frame gesture video, with a single frame size of [size missing]. T × H × W ×3, using a 3D convolutional neural network to extract spatiotemporal features. < i>h v = E v (Video) in, h v This is the initial feature vector of the video. E v For gesture video encoders; Step 12: Perform scene encoding, which involves encoding a single-frame image Scene_Image containing the scene context using a 2D convolutional neural network. h s = E s (Scene_Image) in, h s For scene feature vectors, E s This is a scene encoder.

3. The pedestrian gesture adaptive understanding method for autonomous driving as described in claim 2, characterized in that, Step 2 includes the following sub-steps: Step 21: Perform feature decoupling; Step 22: Perform feature reconstruction; Step 23: Adversarial scenario obfuscation calculation.

4. The pedestrian gesture adaptive understanding method for autonomous driving as described in claim 3, characterized in that, Decoupling D It contains two mapping functions: semantic mapping function. f sem ( ) and style mapping function f sty ( ).

5. The pedestrian gesture adaptive understanding method for autonomous driving as described in claim 4, characterized in that, The calculation formula for the feature decoupling is as follows: z sem = f sem ( h v )=W sem h v + b sem With sty = f sty ( h v )=W sty h v + b sty in, z sem and z sty These are the decoupled semantic feature vector and style feature vector, respectively, W sem and W sty The weights of the feature vectors, b sem and b sty This represents the bias of the eigenvectors.

6. The pedestrian gesture adaptive understanding method for autonomous driving as described in claim 5, characterized in that, Step 22 includes from z sem and z sty The original video features can be reconstructed from the data. h v The calculation formula is: , The formula for calculating the reconstruction loss is: , in, g ( ) represents the refactoring function.

7. The pedestrian gesture adaptive understanding method for autonomous driving as described in claim 6, characterized in that, Step 23 includes introducing a scene classifier. Cscene Predicting scene labels from the decoupled features includes the following sub-steps: Step 231: Using semantic feature vector z sem Calculate scene classification loss L adv_sem The calculation formula is: L adv_sem =CE( Cscene (z sem ), y scene ) in, y scene The labels are real-world scene labels, and CE() represents the cross-entropy loss. Cscene For scene classifiers; Step 232: Using the style feature vector z sty Calculate scene classification loss L sty_scene And minimize the scene classification loss. L sty_scene The calculation formula is: L sty_scene =CE( Cscene (z sty ), y scene )。 8. The pedestrian gesture adaptive understanding method for autonomous driving as described in claim 7, characterized in that, Step 3 includes self-supervised learning using constraints, wherein the constraints are that the semantics of the same gesture are the same in different scenarios.

9. The pedestrian gesture adaptive understanding method for autonomous driving as described in claim 8, characterized in that, The dataset contains sample pairs of the same gesture belonging to the same semantics in different scenarios. i , j The semantic features of () are similar to their distance in the feature space, and the contrastive loss function is: .

10. The pedestrian gesture adaptive understanding method for autonomous driving as described in claim 9, characterized in that, Step 4 includes combining the decoupled features with scene features after completing representation learning to perform the final intent classification, including the following sub-steps: Step 41: Perform feature fusion, which involves concatenating the decoupled semantic features, style features, and scene features to form a comprehensive feature input. f infer =[of sem ;With sty ; h s ] in, f infer The feature is a concatenation of semantic features, style features, and scene features; Step 42: Perform intent classification, including... f infer Feed into a multilayer perceptron classifier (MLP) intent The final output is the probability of intent. p=Softmax(MLP intent ( f infer )) Where p is the probability distribution over all possible intentions.

Images