Generation method about human driving data trajectory generalization

Abstract

The invention provides a generation method about human driving data trajectory generalization, comprising the following steps: step 1, acquiring a driving data set, driving trajectory data comprising a state vector; step 2, constructing a space-time decoupling attention mechanism ST-DAM, and decoupling the state vector based on ST-DAM to form a potential feature; step 3, constructing a deep enhanced generation network DRGN which comprises a generator G and a discriminator; the discriminator calculates a global utility score of the prediction trajectory sequence based on the utility function; 4, performing optimization training on the DRGN based on the potential features by adopting a strategy gradient method in reinforcement learning RL; and step 5, inputting the new state vector into the trained deep enhanced generative network DRGN through a space-time decoupling attention mechanism ST-TAM, and obtaining a prediction trajectory. According to the method, the high fidelity of the generated trajectory is ensured, and meanwhile, the method has high strategy explosiveness and generalization ability under global dynamics constraint.

Description

Technical Field

[0001] This invention belongs to the field of driving trajectory prediction technology, and in particular relates to a method for generating generalized human driving data trajectories. Background Technology

[0002] Existing technologies aim to extract latent representations from human driving data to achieve trajectory generation, mainly falling into two categories, but their inherent limitations are significant: a. Variational Autoencoders (VAEs / CVAEs) based on temporal models: These methods attempt to encode and decode driving behavior in the latent space by learning the probability distribution of trajectory data. Steps (methods): Encode the driving state sequence (position, velocity, acceleration) into a latent vector z, reconstruct the trajectory sequence from z using a decoder, and constrain the distribution of the latent space using KL divergence. Drawback: The inherent smoothing constraint (caused by KL divergence) often smooths out the inherent instantaneous decision-making bursts in human driving behavior, resulting in blurred strategy edges in emergency or complex situations, making it difficult to reflect the decisiveness and nonlinear response of advanced drivers. b. Trajectory generation based on Generative Adversarial Networks (GANs): These schemes utilize the game between the generator and the discriminator, aiming to make the generated trajectory difficult to distinguish from real data in terms of local statistical characteristics. Steps (Methods): The generator G learns to map random noise to driving trajectories, and the discriminator D attempts to distinguish between real and generated trajectories. Nash equilibrium is achieved through alternating optimization. Drawbacks: GANs are highly susceptible to pattern collapse, meaning the generator can only reproduce typical trajectories that occur frequently in the dataset. This directly limits their generalization ability, especially when facing out-of-distribution (OOD) or low-frequency complex driving scenarios, where the generated trajectory sequences lack the necessary randomness and high-complexity variation.

[0003] Against this backdrop, this invention proposes a strategy that combines Deep Reinforcement Generative Networks (DRGN) with a spatiotemporally decoupled attention mechanism, aiming to resolve the fundamental contradiction between high generalization ability and explosive behavior in existing methods. Summary of the Invention

[0004] The purpose of this invention is to provide a method for generating generalized trajectories from human driving data, ensuring high fidelity while possessing high policy explosiveness and generalization ability under global dynamic constraints. The technical solution adopted is as follows:

[0005] A method for generating generalized human driving data trajectories includes the following steps:

[0006] Step 1: Obtain the driving dataset, including the driving trajectory data, which includes state vectors;

[0007] Step 2: Construct a spatiotemporal decoupled attention mechanism ST-DAM, and decouple the state vectors based on ST-DAM to form latent features.

[0008] Step 3: Construct a Deep Reinforcement Generative Network (DRGN), which includes a generator G and a discriminator; the discriminator calculates the global utility score of the predicted trajectory sequence based on the utility function.

[0009] Step 4: Use the policy gradient method in reinforcement learning (RL) to optimize and train the DRGN based on latent features;

[0010] Step 5: Input the new state vector into the trained Deep Reinforcement Generative Network (DRGN) through the spatiotemporal decoupling attention mechanism ST-TAM to obtain the predicted trajectory.

[0011] Preferably, the utility function is a weighted average of comfort constraints and safety constraints.

[0012] Preferably, the loss function L of the generator G is... G Introducing the entropy term L Entropy .

[0013] Preferably, action a t This includes longitudinal acceleration and steering angle rate.

[0014] The key innovation of this invention lies in the shift in modeling paradigm—from traditional geometric / statistical matching to spatiotemporally decoupled high-dimensional policy generation, supplemented by enhanced policy optimization.

[0015] Applications of the Spatiotemporal Decoupling Attention Mechanism (ST-DAM):

[0016] The core idea is to decouple high-dimensional driving trajectory features into spatial geometric features and temporal dynamic features, and then calculate their attention weights separately for weighted fusion, thereby achieving an accurate representation of the nonlinear and highly explosive strategies of human driving in the latent space.

[0017] Construction of Deep Reinforcement Generative Network (DRGN):

[0018] By combining the generator G with a discriminator based on the global utility function U, the traditional binary classification adversarial approach of GAN is replaced with policy optimization using reinforcement learning. This ensures that the generated trajectories are not only statistically realistic in local terms, but also have high confidence in the feasibility and security of the global policy.

[0019] Introduction of entropy weight strategy constraints:

[0020] By adding an entropy term to the generator's loss function, the randomness of the generated strategy is explicitly maintained, which solves the problems of "policy collapse" or "pattern smoothing" in traditional GAN ​​and IL (Imitation Learning) methods. This ensures that the generated generalized trajectories have high complexity and policy burstiness, and can cover atypical and complex driving behaviors at the edge of the dataset.

[0021] Utilization of N×5 Dimensional Extended Trajectory Representation: In the trajectory representation, in addition to the traditional (x,y,z), dynamic features such as instantaneous velocity v and jerk should be introduced to form an N×5 dimensional or higher-dimensional point cloud / sequence representation, which can be used to accurately capture the transient changes of human driving strategies.

[0022] Compared with the prior art, the advantages of the present invention are:

[0023] By introducing a spatiotemporal decoupling attention mechanism to optimize traditional imitation learning methods, we can achieve accurate separation and weight allocation of the spatiotemporal dynamics of driving behavior. Furthermore, by constructing a Deep Reinforcement Generative Network (DRGN) framework, we can solve the local optimality of the loss function and use a discriminator based on a global utility function to optimize the generation strategy, ensuring that the generated trajectory has high fidelity while possessing high policy explosiveness and generalization ability under global dynamic constraints. Attached Figure Description

[0024] Figure 1 A schematic diagram of the principle of a method for generating generalized human driving data trajectories. Detailed Implementation

[0025] The method for generating generalized human driving data trajectories according to the present invention will be described in more detail below with reference to the schematic diagrams, which illustrate preferred embodiments of the invention. It should be understood that those skilled in the art can modify the invention described herein while still achieving its advantageous effects. Therefore, the following description should be understood as being of general knowledge to those skilled in the art and is not intended to limit the invention.

[0026] Terminology Explanation:

[0027] Driving trajectory generalization: refers to extracting driving behavior patterns with high confidence from a limited dataset of real human driving data, which can be generalized to unseen scenarios and environmental conditions, and generating corresponding trajectory sequences.

[0028] Spatiotemporal decoupling attention mechanism: An attention model whose core is to nonlinearly decouple spatial features (such as lane keeping and lateral displacement) and temporal features (such as speed changes and transient responses to acceleration and deceleration) in the driving trajectory, and assign weights to them respectively, so as to accurately capture the nonlinear coupling characteristics of human driving strategies.

[0029] Deep Reinforcement Generative Network: This invention employs a generative model architecture that combines the generative capabilities of deep learning with the policy optimization process of reinforcement learning. It guides trajectory generation through a utility function-based discriminative mechanism (rather than the binary classification of traditional GANs), achieving a balance between high fidelity and policy feasibility. Entropy-weighted policy constraints are introduced in reinforcement learning policy optimization to maintain a certain entropy value in the generated trajectory policy space. This prevents the model from prematurely converging to deterministic policies with low generalization ability and low randomness on limited data, thus ensuring the diversity (i.e., high burstiness) of generated trajectories.

[0030] like Figure 1 As shown, a method for generating generalized human driving data trajectories includes the following steps:

[0031] Step 1: Obtain the driving dataset.

[0032] The original driving trajectory data is cleaned and normalized, and then the driving trajectory data is extracted.

[0033] Driving trajectory data includes: state vector s t and action a t .

[0034] State vector S t Includes: Driver's Intent I t Environmental State E t (Including lane and obstacle information), vehicle status C t Pedestrian status P t .

[0035] s t =(C t P t E t I t )

[0036]

[0037] Among them, a long,t -Longitudinal acceleration, - Steering angular rate, t - current time; T - transpose.

[0038] Step 2: Construct a spatiotemporal decoupled attention mechanism ST-DAM, and decouple the state vectors based on ST-DAM to form latent features.

[0039] ST-DAM: Spatio-Temporal Decoupled Attention Mechanism.

[0040] ST-DAM includes: Spatial Attention Module (SAM) and Temporal Attention Module (TAM).

[0041] The temporal attention module (TAM) and the spatial attention module (SAM) are configured in parallel.

[0042] The state vector s t The inputs are fed into the temporal attention module (TAM) and the spatial attention module (SAM) respectively, and the temporal enhancement feature F is obtained through the temporal attention module (TAM). T Spatial augmented features F are obtained through the Spatial Attention Module (SAM). S .

[0043] Among them, S-Spatial Attention Module (SAM) and T-Temporal Attention Module (TAM) are defined.

[0044] The two attention weights α are combined through a nonlinear fusion layer. S and α T Applied to the original features, generating decoupled high-dimensional latent features.

[0045]

[0046] Here, b is the bias term, a flexible correction value added to the formula to ensure the most accurate result.

[0047] Specifically:

[0048] Spatial Attention Module (SAM): Focuses on capturing geometric and topological constraints such as lateral displacement and lane keeping accuracy.

[0049] The Time Attention Module (TAM) focuses on capturing transient dynamic characteristics such as velocity, acceleration, and jerk speed.

[0050] Among them, lateral displacement, lane keeping accuracy, speed, acceleration, and jerk speed all belong to the vehicle's state C. t .

[0051] Step 3: Construct a Deep Reinforcement Generative Network (DRGN); the DRGN consists of a generator G and a discriminator D. U Generator G and discriminator D U Constructing a generative adversarial network;

[0052] Generator G (based on Transformer encoder-decoder architecture):

[0053] The loss function L of generator G G Introducing the entropy term L Entropy .

[0054] L G=L GaN +λ Entropy L Entropy

[0055] L GAN - Combat losses;

[0056] L Entropy - Loss due to entropy weighting strategy constraints;

[0057] λ Entropy - Weight.

[0058] L GAN Guide G to minimize D U The probability of an inefficient (low-score) trajectory.

[0059] L Entropy Explicitly encourage G to maintain a high degree of randomness / exploration in the action space to prevent premature policy convergence and ensure that the policy burstiness of generated trajectories remains high.

[0060] The input of traditional G is (Features) and random noise.

[0061] This invention should introduce an additional intent / randomness injection module into G, as a component in the generator, specifically designed to transform random noise z into a random factor (α) that drives the temporal attention mechanism. S and α T This ensures that randomness is policy-driven rather than a simple geometric perturbation.

[0062] Discriminator D U :

[0063] D U It not only determines the truth value of τ, but more importantly, it calculates its global utility score on preset safety, comfort, and efficiency indicators.

[0064] The global utility score U of the predicted trajectory τ is calculated based on the utility function U(τ).

[0065] U=U(τ=ω1·U Safety (τ)+ω2·U Comfort (τ)

[0066] U Safety - Minimum safe distance or collision risk.

[0067] U Comfort - Squared constraints based on acceleration and jerk (Jerk) to limit unnecessary smoothing.

[0068] U Efficiency -Track-based end-state error;

[0069] ω1, ω2, ω1 - weighting coefficients.

[0070] Step 4: Use the policy gradient method in reinforcement learning (RL) to optimize and train the generator G based on latent features.

[0071] Training: Encoder reception The decoder is responsible for determining the current time. The autoregressive output of the next time-step state vector s t+1 and action a t+1 (Longitudinal acceleration, steering angular rate), thus forming the predicted trajectory sequence t+1.

[0072] τ t+1 ={s t+1 a t+1}

[0073] Similarly, τ t+1 ={s t+1，at+1}

[0074] That is, t = 0 ~ T′.

[0075] Step 5: Input the new state vector s0 into the trained deep reinforcement generative network DRGN through the spatiotemporal decoupling attention mechanism ST-TAM to generate the predicted trajectory sequence and finally obtain the predicted trajectory.

[0076] In step 5, it is necessary to collect the action a0 corresponding to the state vector s0.

[0077] The predicted trajectory sequence is a generalized driving trajectory with high fidelity, policy coherence, and the ability to reflect the explosive nature of human instantaneous decision-making.

[0078] In step 5, the encoder receives The decoder is responsible for determining the state vector s at the current time step. t The autoregressive output of the next time-step state vector s t+1 and action a t+1 (Longitudinal acceleration, steering angular rate), thus forming the predicted trajectory sequence τ t+1 .

[0079] τ t+1 ={s t+1 a t+1}

[0080] Where, τ t The predicted trajectory sequence corresponding to -t.

[0081] t = 0 ~ T′.

[0082] Predicted trajectory τ=(τ 0，τ1 …, τt , τ t+1 ,…τ T′ )

[0083] τ0={s0,a0}

[0084] Loop mechanism:

[0085] 1. Input Output τ1 = {s1, a1}

[0086] 2. The predicted s1 is used as a new input and passed through ST-DAM again.

[0087] 3. G predicts τ2 based on the updated state.

[0088] 4. Repeat the above process until the prediction time horizon T′, and combine all predicted trajectory sequences to form τ.

[0089] Case Study:

[0090] This embodiment aims to address the challenge of acquiring instantaneous, multimodal pedestrian avoidance strategies for autonomous driving systems in unstructured, low-visibility environments. We focus on how to utilize the spatiotemporal decoupling attention mechanism (ST-DAM) and entropy-weighted policy constraints (L) of this invention. Entropy This generates an avoidance trajectory τ with high strategic burst potential and global dynamic feasibility.

[0091] To capture the highly dynamic process of pedestrian avoidance, a state vector s was constructed that incorporates vehicle dynamics, environmental geometry, and target interaction. t and action vector a t :

[0092]

[0093] Where ψ is the heading angle, d lane P(Intent) is the lateral distance between the vehicle's current position and the edge of its lane (or curb), and P(Intent) is the probability of the driver's intention (such as acceleration, braking, or detouring) estimated by RNN / LSTM.

[0094] x t y t : Refers to the lateral and longitudinal position coordinates of the vehicle at time t.

[0095] v t : refers to the instantaneous speed of the vehicle at time t.

[0096] Action vector a t :

[0097]

[0098] Among them, along,t For longitudinal acceleration, The turning angular rate and the turning angular rate directly determine the instantaneous dynamics of the trajectory.

[0099] The ST-DAM module receives s in the generator G. t Then, it is first reconstructed in high dimension to prevent geometric constraints from inhibiting instantaneous dynamics.

[0100] Decoupling characteristic matrix F Input :

[0101]

[0102] Where Input refers to all S received by generator G. t sequence;

[0103] MLP: A basic neural network architecture

[0104] T obs The observation time step refers to the row number in the matrix and represents the length of time for backtracking in the model.

[0105] d feat Feature dimension: In a matrix, it refers to the number of columns; in a model, it represents the depth of understanding or richness of description of information at a certain moment.

[0106] Strategy decoupling computation:

[0107] (Time-explosive characteristics) t Weight α T Focus on s t v in t , and v p,t High-frequency differential information. This ensures that the strategy receives maximum attention weight when pedestrians suddenly accelerate or decelerate, guaranteeing the explosiveness of instantaneous decision-making.

[0108] F T =Attention(F Input W T )

[0109] Spatial coherence characteristic F S Weight α S Focus on x t ,y t and d lane Equal geometric integral information. This ensures that the avoidance trajectory has minimum safe distance constraints and lane topology consistency in lateral displacement.

[0110] F S =Aeeention(F Input WS )

[0111] Among them W T W S This refers to the learnable weight matrix under each attention module.

[0112] Nonlinear Reconstruction Strategy Final strategy representation It serves as the input to the Transformer encoder of the generator G.

[0113]

[0114] Meaning: It is a spatiotemporally heterogeneous and weighted policy latent space representation whose structure is optimized to simultaneously maintain instantaneous burstiness (high Jerk tolerance) and global geometric rationality (high d). lane constraint).

[0115] Regarding U(τ):

[0116] U(τ)=R safety (τ)+R comfort (τ)

[0117] R safety (τ)=ω1·U Safety (τ)

[0118] R comfort (τ)=ω2·U Comfort (τ)

[0119] Safety constraint R safety (τ): The penalty based on the predicted collision risk in the time domain is the primary indicator to ensure the feasibility of the trajectory strategy. We adopt the minimum safe distance (d) based on the predicted trajectory. min Exponential penalty:

[0120]

[0121] Where l(·) is the indicator function;

[0122] When d min <d th d is 1 if it is 1, otherwise it is 0. th This is the minimum safety threshold.

[0123] This optimization ensures strong security constraints on the strategy.

[0124] Comfort / strategy stability constraint term R comfort (τ): This term directly and accurately models the explosiveness of human driving behavior. We employ a squared penalty on the jerk intensity, but with weight λ. comfIt is dynamically adjusted during training to tolerate reasonable bursts of speed:

[0125] R comfort (τ)=-(||J x,t || 2 +||J y,t || 2 )·λ comf

[0126] in

[0127] a is the vehicle acceleration, and J is the acceleration of the vehicle. x,t J refers to the lateral abruptness of a vehicle. y,t This refers to the longitudinal speed of the vehicle.

[0128] For example, high lateral abruptness indicates that the vehicle is turning the steering wheel sharply, while high lateral abruptness indicates that the vehicle is braking suddenly or accelerating sharply.

[0129] In traditional LQR control, this is intended to be minimized, resulting in a smooth trajectory.

[0130] And the D of the present invention U The evaluator will tolerate Jerk bursts (U(τ)) that fit the scene in human trajectories, and will feed this feature back to the generator G through reinforcement learning.

[0131] The optimization objective of generator G is to maximize expected utility and introduce negative policy entropy β·H(π). θ () is used as a regularization term to explicitly enhance the randomness and complexity of the strategy.

[0132]

[0133] Counteracting utility loss The expectation of the trajectory is used to guide G to generate a trajectory with high global utility.

[0134] Where: π represents the driving strategy, which is the probability distribution of which strategy the system decides to adopt in various situations. θ represents all the weights and biases within the generator G neural network, which are the parameters that determine the driving level and style at the current moment.

[0135] Geometric Fidelity Loss Examples of methods include Chamfer Distance or Earth Mover's Distance, which ensure that the generated trajectory is consistent with the local geometry of the real data.

[0136] Entropy weighting strategy constraint -β·H(π) θ This is the core guarantee for the explosiveness of the strategy. β > 0 is the entropy weighting coefficient.

[0137]

[0138] Maximize H(π) θ This is equivalent to the incentive strategy π. θ The action space is more uniform and distributed. In pedestrian avoidance scenarios, this means the model will not only learn one "average" avoidance method (such as moderate deceleration), but will be able to simultaneously learn and generate decisive stopping and waiting (low v high Jerk burst) and rapid detour (high v instantaneous). Multimodal and highly randomized strategies such as bursts are used to ensure that the generated trajectories have extremely high generalization ability and strategy bursting degree when facing unseen scenarios.

[0139] The above are merely preferred embodiments of the present invention and do not constitute any limitation on the present invention. Any equivalent substitutions or modifications made by those skilled in the art to the technical solutions and content disclosed in the present invention without departing from the scope of the present invention shall be deemed to have remained within the protection scope of the present invention.

Claims

1. A method for generating generalized human driving data trajectories, characterized in that, Includes the following steps: Step 1: Obtain the driving dataset, including the driving trajectory data, which includes state vectors; Step 2: Construct a spatiotemporal decoupled attention mechanism ST-DAM, and decouple the state vectors based on ST-DAM to form latent features. Step 3: Construct a Deep Reinforcement Generative Network (DRGN), which includes a generator G and a discriminator; the discriminator calculates the global utility score of the predicted trajectory sequence based on the utility function. Step 4: Use the policy gradient method in reinforcement learning (RL) to optimize and train the DRGN based on latent features; Step 5: Input the new state vector into the trained Deep Reinforcement Generative Network (DRGN) through the spatiotemporal decoupling attention mechanism ST-TAM to obtain the predicted trajectory.

2. The method for generating generalized human driving data trajectories according to claim 1, characterized in that, The utility function is a weighted average of comfort and safety constraints.

3. The method for generating generalized human driving data trajectories according to claim 1, characterized in that, The loss function L of generator G G Introducing the entropy term L Entropy .

4. The method for generating generalized human driving data trajectories according to claim 1, characterized in that, Action a t This includes longitudinal acceleration and steering angle rate.

Images