Intelligent safety protection method and system based on trajectory prediction and collision situation identification

Abstract

The application discloses an intelligent safety protection method and system based on trajectory prediction and collision situation identification. The intelligent safety protection method comprises the following steps: constructing a structured space-time heterogeneous graph in real time; outputting a time sequence enhanced feature vector of a vehicle node according to the space-time heterogeneous graph; outputting predicted trajectories of the vehicle and a target vehicle according to the time sequence enhanced feature vector; identifying a collision risk feature when it is judged that there is a collision risk according to the predicted trajectories of the vehicle and the target vehicle; predicting a collision type according to the identified collision risk feature; according to the identified collision type, solving a control instruction sequence for hierarchical triggering in real time according to a preset protection strategy; and executing a preset safety protection action according to the control instruction sequence for hierarchical triggering before a collision time. The application overcomes the defects of the prior art, such as inaccurate collision type identification and single protection strategy, and realizes early and accurate identification of a collision risk and hierarchical and differentiated protection.

Description

Technical Field

[0001] This invention relates to the field of intelligent vehicle safety technology, and in particular to an intelligent safety protection method and system based on trajectory prediction and collision situation identification. Background Technology

[0002] With the increasing intelligence of automobiles, pre-collision systems (PCS) have become a core component of vehicle safety. However, existing technologies have inherent limitations in several aspects:

[0003] First, the granularity of collision type identification is coarse. Most mainstream systems make binary decisions (whether a collision occurs) based on vehicle kinematic parameters (such as time to collision (TTC)), which cannot accurately identify the physical type of the collision (such as frontal, offset, side, and rear-end collisions). Since the dynamic processes, occupant injury mechanisms, and optimal protection strategies differ significantly among different collision types, this lack of identification capability makes it difficult for protection systems to achieve precise protection.

[0004] Secondly, existing trajectory prediction schemes, such as those based on historical trajectory fitting or vehicle motion state deduction, generally treat traffic participants as independent individuals, ignoring their strong interactive dependencies in complex scenarios. Such methods exhibit high prediction inaccuracies in interaction-intensive scenarios such as vehicle lane changes and intersection merging.

[0005] Secondly, there is a mismatch between the protection strategy and the collision situation. Due to the lack of precise identification of collision types, the protection strategies of existing systems often adopt a "one-size-fits-all" approach, that is, triggering the same or similar passive safety devices for different collision types, without being able to adaptively adjust according to the physical characteristics of the collision (such as impact direction and energy level), resulting in low protection efficiency.

[0006] Although existing research has applied graph neural networks or Transformer models to trajectory prediction, how to deeply integrate them and systematically construct a closed-loop security protection scheme that covers the entire chain from high-precision interactive perception and trajectory prediction to accurate collision type identification and the generation and execution of differentiated protection strategies remains an urgent technical problem to be solved. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings and defects of existing technologies by providing an intelligent safety protection method and system for vehicle collision type identification based on high-precision trajectory prediction and multi-dimensional collision situation recognition. By constructing an integrated technical closed loop of "perception-prediction-identification-protection", it achieves collision warning, refined collision type identification, and differentiated active protection, thereby significantly improving the occupant safety protection level of intelligent vehicles in complex traffic environments.

[0008] One objective of this invention is to provide an intelligent safety protection method based on trajectory prediction and collision situation identification, comprising the following steps:

[0009] Using the vehicle and the traffic participants within its real-time perception range as nodes, and the potential interaction relationships between nodes as directed edges, a structured spatiotemporal heterogeneous graph representing the dynamic traffic scene is constructed in real time.

[0010] Using a trained GAT-Transformer prediction model, based on the real-time spatiotemporal heterogeneous graph and node feature vectors, the temporal enhanced feature vector of the vehicle node is output. This temporal enhanced feature vector deeply integrates the vehicle node's own motion history information and its spatial interaction information with the surrounding environment. Using a trained conditional variational autoencoder, trajectory prediction is performed based on the temporal enhanced feature vector, and multiple predicted trajectories of the vehicle and the target vehicle are output, with each predicted trajectory including a confidence level.

[0011] Based on the predicted trajectories of the vehicle and the target vehicle, if a collision risk is identified, collision risk features are identified. A multi-dimensional judgment feature vector is constructed based on the identified collision risk features and input into a multilayer perceptron for processing. The collision type is determined by using the maximum probability of the probability distribution of the collision type output by the multilayer perceptron.

[0012] Based on the identified collision type and the preset protection strategy, the action time of the protection device is calculated in real time by the real-time decision engine to form a sequence of control commands for graded triggering; the preset safety protection actions are triggered and executed in a graded manner according to the control command sequence before the collision time.

[0013] Preferably, each node at time The state is determined by a multidimensional feature vector The description, encompassing kinematic states and semantic attributes, is represented as follows:

[0014]

[0015] in, These are two-dimensional position coordinates in the vehicle coordinate system; and These are velocity and acceleration vectors, respectively. For heading angle; This is a traffic participant type identifier that has been uniquely encoded.

[0016] Preferred ， A directed edge representing the potential relationship between nodes is constructed using a dynamic connection strategy based on preset rules; the preset rules include:

[0017] Spatial proximity rule: If the Euclidean distance between two nodes is less than a threshold, then a bidirectional edge is established between the two nodes;

[0018] Lane topology rules: Based on high-precision maps or visually perceived lane line information, if two nodes are in the same lane, adjacent lanes, or lanes with path conflicts, a connection edge is established.

[0019] Each directed edge carries standardized relative motion information from source node j to target node i, i.e., edge features. This serves as a basis for assessing the intensity of interaction:

[0020]

[0021] in, Representing nodes respectively Relative to node The difference between the two-dimensional coordinates, velocity, and heading angle.

[0022] Preferably, when performing trajectory prediction, the conditional variational autoencoder uses a prior network to generate a latent variable distribution based on the temporal augmentation feature vector, and samples from it. Hidden variables Each hidden variable The temporal augmentation feature vectors are input into the decoder to generate the future. The trajectory of the steps At the same time, for each generated trajectory Output a raw confidence score Based on the original confidence score The confidence level is obtained by normalizing the probability using the Softmax function. ;

[0023] = ;

[0024] , ;

[0025] in, Indicates the first time step in the future to the first The x and y coordinates of the two-dimensional position of the trajectory at the time step. It is the loop variable in the summation symbol ∑, representing the sequence from the 1st to the 2nd. The original confidence score of any one of the trajectories.

[0026] Preferably, the following loss function is used during the training of the conditional variational autoencoder. :

[0027]

[0028] in, The average displacement error is used to minimize the average distance between the predicted point and the true point at the predicted time step. The final displacement error is used to minimize the distance between the predicted endpoint and the actual endpoint; For the confidence loss, a negative log-likelihood loss is used to encourage the conditional variational autoencoder to assign higher confidence scores to predictions that are closer to the true trajectory. ; KL divergence is used to normalize the posterior and prior distributions of latent variables in a conditional variational autoencoder and is the core regularization term for training a conditional variational autoencoder. , , , Weighting coefficients to balance the various losses.

[0029] Preferably, the collision risk feature characterizes a collision risk index for predicting the moment of collision, including a relative velocity vector determined based on the velocity vectors of the vehicle and the target vehicle, a collision time TTC determined based on the ratio of the Euclidean distance between the vehicle and the target vehicle to the relative velocity vector, a collision angle at the moment of collision determined based on the velocity vectors of the vehicle and the target vehicle, and a collision energy calculated based on the relative velocity vector, the mass of the vehicle, and the mass of the target vehicle; wherein, the collision angle at the moment of collision is predicted. After mapping to a continuous two-dimensional space, a vector is obtained as the final collision angle:

[0030] ;

[0031] ;

[0032] in, , These represent the velocity vectors of the vehicle and the target vehicle, respectively. Represents an angle function;

[0033] Collision energy is This is expressed as a calculation using the following formula:

[0034] ;

[0035] in, and The masses of the target vehicle and the vehicle itself are respectively. This represents the relative velocity vector.

[0036] Preferably, the multi-dimensional judgment feature vector constructed based on the identified collision risk features is represented as F, and its expression is as follows:

[0037] ;

[0038] This represents the minimum distance between the predicted trajectory pairs of this vehicle and the target vehicle;

[0039] This represents the time from the current moment to the predicted collision moment. Time difference;

[0040] This represents the current speed scalar of the vehicle.

[0041] , Indicates the type codes of this vehicle and the target vehicle;

[0042] The multidimensional judgment feature vector is input into a multilayer perceptron for processing and then outputs a probability distribution of collision types.

[0043] Then, based on the maximum probability of the probability distribution of the collision type, i.e. the confidence level of the collision type, the identification result of the collision type is determined.

[0044] Preferably, the multilayer perceptron is composed of an input layer, a first hidden layer, a second hidden layer, and an output layer connected in sequence. The first hidden layer and the second hidden layer use the ReLU function to learn the higher-order interactions between features and extract abstract features to prevent overfitting. The output layer uses the Softmax activation function to normalize the output result into a probability distribution.

[0045] Preferably, preset safety protection actions are triggered and executed in stages according to the control command sequence before the collision time, including:

[0046] The collision time (TTC) is compared with a preset threshold range, and the current warning level is determined based on the comparison result.

[0047] Based on the previous early warning level, find the pre-generated control instruction sequence for the current early warning level;

[0048] Execute preset safety protection actions according to the control command sequence of the current early warning level;

[0049] The control command sequence is calculated in real time by the real-time decision engine based on the collision type and its confidence level, collision risk characteristics, confidence level of the predicted trajectory, vehicle speed, gear position, and the position and posture of the occupants.

[0050] Another objective of this invention is to provide an intelligent safety protection system based on trajectory prediction and collision situation identification, used to implement the intelligent safety protection method based on trajectory prediction and collision situation identification, comprising:

[0051] The graphing module is used to construct a spatiotemporal heterogeneous graph representing the dynamic traffic scene in real time, using the vehicle and traffic participants within its real-time perception range as nodes and the potential interaction relationships between nodes as directed edges.

[0052] The trajectory prediction module is used to use the GAT-Transformer prediction model to output the temporal enhanced feature vector of the vehicle node based on the real-time spatiotemporal heterogeneous graph and node feature vectors. The temporal enhanced feature vector is deeply integrated with the vehicle node's own motion history information and its spatial interaction information with the surrounding environment. The conditional variational autoencoder is used to perform trajectory prediction based on the temporal enhanced feature vector, and outputs multiple predicted trajectories of the vehicle and the target vehicle respectively. Each predicted trajectory includes a confidence level.

[0053] The collision recognition module is used to identify collision risk features when the predicted trajectories of the vehicle and the target vehicle indicate a collision risk. A multi-dimensional judgment feature vector is constructed based on the identified collision risk features and input into a multilayer perceptron for processing. The collision type is determined by using the maximum probability of the probability distribution of the collision type output by the multilayer perceptron.

[0054] The protection decision module is used to calculate the action time of the protection device in real time through the real-time decision engine based on the identified collision type and the preset protection strategy, and form a sequence of control commands for graded triggering.

[0055] The protection execution module is used to trigger and execute preset safety protection actions in stages according to the control command sequence before the collision time.

[0056] This invention enables targeted active and passive safety protection based on the recognition results, forming an integrated technical closed loop of "perception-prediction-recognition-protection". It realizes a fundamental transformation from collision warning to precise collision type recognition and differentiated protection, thereby significantly improving the occupant safety protection level of intelligent vehicles in complex traffic environments. Attached Figure Description

[0057] Figure 1 This is a flowchart of an intelligent safety protection method based on trajectory prediction and collision situation identification according to an embodiment of the present invention.

[0058] Figure 2 This is a schematic diagram of an intelligent safety protection system based on trajectory prediction and collision situation identification according to an embodiment of the present invention. Detailed Implementation

[0059] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0060] In this application, a spatiotemporal heterogeneous graph is constructed to describe the interaction relationship between the vehicle and traffic participants within its perception range. A deep learning model combining Graph Attention Network (GAT) and Transformer is used to extract the spatial interaction features and temporal dependencies between the vehicle and the target vehicle, resulting in a temporal augmented feature vector. A trained conditional variational autoencoder is then used to predict the trajectory based on the temporal augmented feature vector, outputting the predicted trajectories of the vehicle and the target vehicle. Each predicted trajectory includes a confidence level, thus solving the problems of uncertainty and interaction dependency in trajectory prediction under complex interaction scenarios.

[0061] See Figure 1 As shown in the embodiments of this application, the intelligent safety protection method based on trajectory prediction and collision situation identification includes the following steps:

[0062] S1. Using the vehicle and the traffic participants within its real-time perception range as nodes, and the potential interaction relationships between nodes as directed edges, a spatiotemporal heterogeneous graph representing the dynamic traffic scene is constructed in real time.

[0063] S2. Using the trained GAT-Transformer prediction model, based on the real-time spatiotemporal heterogeneous graph and node feature vectors, output the temporal enhanced feature vector of the vehicle node; this temporal enhanced feature vector deeply integrates the vehicle node's own motion history information and its spatial interaction information with the surrounding environment; using the trained conditional variational autoencoder, perform trajectory prediction based on the temporal enhanced feature vector, and output multiple predicted trajectories of the vehicle and the target vehicle, each predicted trajectory including a confidence level;

[0064] S3. Based on the predicted trajectories of the vehicle and the target vehicle, if a collision risk is identified, identify the collision risk features; construct a multi-dimensional judgment feature vector based on the identified collision risk features and input it into a multilayer perceptron for processing; use the maximum probability of the probability distribution of the collision type output by the multilayer perceptron to determine the collision type.

[0065] S4. Based on the identified collision type and the preset protection strategy, the action time of the protection device is calculated in real time by the real-time decision engine to form a control command sequence for graded triggering; the preset safety protection actions are triggered and executed in a graded manner according to the control command sequence before the collision time.

[0066] In some embodiments, when the vehicle is driving on the road, real-time status data of the vehicle and surrounding traffic participants are collected by fusion of onboard multi-source sensors; after time alignment, coordinate unification, data cleaning (such as removing abnormal noise) and data fusion filtering of the status data, a structured spatiotemporal heterogeneous map representing the dynamic traffic scene is constructed.

[0067] Specifically, time alignment can be achieved by combining hardware trigger signals with software timestamps to unify all sensor data to the vehicle's CAN bus time base, with time synchronization errors controlled within ±10ms.

[0068] The aforementioned coordinate unification involves unifying the coordinates of all targets identified by radar, lidar, and cameras into a vehicle coordinate system with the rear axle center of the vehicle as the origin through a calibration transformation matrix.

[0069] The data fusion and filtering described herein, for example, employs an extended Kalman filter (EKF) to fuse and track observation data from multiple sensors on the same target, thereby obtaining a smooth and continuous sequence of motion states of the vehicle and surrounding targets, including position, velocity, acceleration, and heading angle.

[0070] For example, the vehicle-mounted multi-source sensors include at least millimeter-wave radar, lidar, cameras, and a V2X communication module for acquiring information on traffic participants beyond the sensor's line of sight. For instance, a forward-looking millimeter-wave radar (such as ARSR410) with a detection range of 200 meters can accurately measure the relative distance, relative speed, and azimuth of target vehicles. A solid-state lidar (such as Livox Horizon) with a field of view of 81.7°×25.1° can provide high-precision point cloud data for target contour recognition and positioning. A forward-looking wide-angle camera (1280×720 resolution) provides traffic participant type information and lane line information through visual recognition. A V2X communication module (supporting C-V2X PC5 interface) receives real-time status information such as GPS position, speed, and heading angle of other connected vehicles within the beyond-line-of-sight range.

[0071] In constructing a spatiotemporal heterogeneity diagram =( , When constructing the data, you can extract historical time T at 0.1-second intervals. hist =3.0 seconds (30 frames in total) of historical data. Based on the historical data and perceived environmental interaction information, a spatiotemporal heterogeneous graph is constructed. Every important traffic participant (such as vehicles, motorcycles, pedestrians, cyclists) within a predetermined range (e.g., 100 meters) of the vehicle and its surroundings is defined as a node in the graph structure. and at the node Define directed edges ∈ It describes the potential interaction relationships or spatial interaction relationships between nodes, thus forming a spatiotemporal heterogeneous graph; each node represents an intelligent agent with independent movement intention, and edges are usually constructed based on physical distance thresholds or lane topology relationships (such as adjacent lanes).

[0072] For example, in the spatiotemporal heterogeneous graph representing a dynamic traffic scene, each node at time... The state is determined by a multidimensional feature vector The description, which encompasses both kinematic states and semantic attributes, is represented as follows:

[0073]

[0074] in, These are two-dimensional position coordinates in the vehicle coordinate system; and These are the corresponding velocity and acceleration vectors, respectively. For heading angle; This is a traffic participant type identifier that is uniquely encoded, such as car, truck, pedestrian, or cyclist.

[0075] For example, in this embodiment of the application, a directed edge representing the potential interaction relationship between nodes is constructed using a dynamic connection strategy based on preset rules. By using a rule-based dynamic connection strategy for edge construction, rather than full connection, computational efficiency can be further improved and key interactions can be focused on; the preset rules include:

[0076] Spatial proximity rule: If the Euclidean distance between two nodes is less than a threshold (e.g., 50m), then a bidirectional edge is established between the two nodes;

[0077] Lane topology rules: Based on lane line information from high-precision maps or visual perception, if two nodes are in the same lane, adjacent lanes, or lanes with path conflicts (such as merging areas at intersections), then a connection edge is established;

[0078] One approach is to construct spatial directed edges based on the K-nearest neighbor algorithm (K=8);

[0079] Each directed edge carries a message from the source node. Point to target node Standardized relative motion information or interrelationship information, i.e., edge features This includes relative position, relative speed, and relative heading angle, which serve as the basis for evaluating the intensity of the interaction.

[0080]

[0081] in, Representing the source node respectively Relative to the target node The differences in coordinates, velocity, and heading angle at a two-dimensional position, such as The aforementioned relative features enable the model to directly perceive the location and movement trends of its neighbors.

[0082] To better quantify the differences in the impact of various traffic participants on the target vehicle, this application implements a GAT-Transformer inference model that dynamically quantifies the influence weights of different traffic participants on the future behavior of the target node based on dynamic relation learning of a multi-head graph attention network (GAT). This avoids treating all neighboring nodes equally and performs spatial interaction encoding to quantify spatial interaction features. The GAT introduces a multi-head graph attention network, adaptively and non-linearly aggregating neighbor information through a learnable attention mechanism to process the input node features and output spatial interaction features for each node that incorporate the influence of its surrounding neighbors and strong interactive dependencies. For each node, its new features are encoded by a single-head GAT. It is all the characteristics of its neighbors Weighted sum (including itself):

[0083]

[0084] Where W is the shared linear transformation weight matrix, It is a non-linear activation function (such as ELU). It is a node The set of neighbors (usually including itself). Attention coefficient. / weight Dynamically calculate and measure neighbors through attention mechanisms. For the target Importance:

[0085]

[0086] in, It is a learnable attention vector. It is a shared linear transformation weight matrix, and ∥ indicates vector concatenation. It is a node The neighborhood set. This mechanism enables the model to explicitly learn the interaction strength based on the joint features of node pairs (such as relative position and velocity).

[0087] M independent attention heads (e.g., M=8) are used, each learning interaction patterns in a different feature subspace. The output features of each attention head are then concatenated or averaged to obtain the final spatial interaction encoding features. By stacking Layer (such as) =2) GAT, the model is able to capture multi-hop indirect interactions and interactions over longer distances.

[0088] Because the movements of traffic participants are strongly temporally correlated, this model uses GAT for spatial encoding and then Transformer for temporal dependency modeling. The encoder captures the historical evolution of the state of each node (traffic participant) and extracts the movement patterns or information of each node.

[0089] The input sequence for Transform is, for each node, its position within the historical time window [tT] hist The spatial features within [+1,t] after GAT encoding, arranged in chronological order, form a sequence vector as shown below:

[0090] ;

[0091] After receiving the sequence vector, first in the sequence Learnable positional codes are added to inject temporal sequence information; then the sequence is input into a... Layer (such as) =3) Transformer encoder. Each encoder layer contains two sublayers: Multi-Head Self-Attention (MHA) and Feedforward Network (FFN). Multi-Head Self-Attention (MHA) enables the model to pay attention to all time steps in the sequence in parallel, automatically identifying key historical moments for predicting the future (e.g., moments of sudden deceleration); the Feedforward Network (FFN) performs a non-linear transformation on the representation at each position. After L t After layer encoding, the model outputs the final temporal augmentation feature vector for each node. This time-series enhanced feature vector By deeply integrating the node's own motion history patterns and its complex spatial interactions with the surrounding environment, a unified contextual representation for predicting its future behavior is formed.

[0092] For example, the Transformer encoder has a 3-layer structure with a hidden layer dimension of 256, a feedforward network dimension of 512, and 8 self-attention heads. It processes the spatial features of nodes at multiple historical moments (e.g., 30) in chronological order and outputs a feature vector for each node that is rich in spatiotemporal context information.

[0093] In this application, multimodal trajectory prediction generates multiple possible future trajectories or predicted trajectories to reflect the inherent uncertainties in traffic scenarios (such as whether a vehicle can go straight, change lanes, or slow down). In order to reflect the uncertainty of driving intentions, this application uses a probabilistic generation framework or a probabilistic generation model of generative adversarial network (GAN) as a decoder to output multiple possible future trajectories and their occurrence probabilities.

[0094] In a preferred embodiment, the model preferably uses a conditional variational autoencoder (CVAE) as its decoder to conditionally decode the final output feature of the Transformer encoder. The conditional variational autoencoder (CVAE) includes an encoder... ( | ， Prior networks ( | ) and decoder ( | , ).

[0095] Training phase: The conditional variational encoder will generate the true future trajectory and conditional features Mapping to latent variables The decoder learns from the latent distribution and conditional features the posterior distribution. Reconstructing the trajectory of the future.

[0096] For example, in the embodiments of this application, during the inference or prediction stage, the conditional variational autoencoder uses a prior network to perform trajectory prediction based on temporal augmentation feature vectors. (Conditional features) Generate latent variables Distribution, from which to sample One (e.g.) =5) Latent variables Each hidden variable With temporal enhancement feature vectors Input the decoder together to generate the future Multiple trajectories in one step (e.g., 3.0 seconds, 30 points in total) At the same time, for each trajectory Output a raw confidence score Based on the original confidence score The confidence level is obtained by normalizing the probability using the Softmax function. This confidence level This reflects the model's assessment of the likelihood of the trajectory assumption;

[0097] = ,..., ；

[0098] , ;

[0099] in, Indicates the first time step in the future to the first The x and y coordinates of the two-dimensional position of the trajectory at the time step. The decoder is the first The raw confidence scores of the trajectory outputs. It is the loop variable in the summation symbol ∑, representing the sequence from the 1st to the 2nd. The original confidence score of any one of the trajectories.

[0100] For example, in this embodiment of the application, the following loss function is used during the training of the conditional variational autoencoder. :

[0101]

[0102] in, The average displacement error is used to minimize the average distance between the predicted point and the true point at the predicted time step. The final displacement error is used to minimize the distance between the predicted endpoint and the actual endpoint; For the confidence loss, a negative log-likelihood loss is used to encourage the conditional variational autoencoder to assign higher confidence scores to predictions that are closer to the true trajectory. ; KL divergence is used to normalize the posterior and prior distributions of latent variables in a conditional variational autoencoder and is the core regularization term for training a conditional variational autoencoder. , , , To balance the weighting coefficients of various losses, the weights of the loss function can be set to 0.5, 0.3, 0.1, and 0.1 respectively, in order to balance trajectory accuracy and confidence calibration.

[0103] In this embodiment, the trajectory prediction module systematically improves the accuracy and reliability of trajectory prediction in complex, interaction-intensive scenarios by constructing a structured spatiotemporal heterogeneous graph, using attention-based dynamic interactive coding, powerful temporal pattern extraction, and probabilistic multimodal decoding. It outputs multiple sets of "trajectory-confidence scores," providing core data support for subsequent accurate collision risk situation identification and classification.

[0104] In this embodiment, the probabilistic future trajectory output by the trajectory prediction model is transformed into a physical understanding and early classification of potential collision events. It is no longer satisfied with the binary judgment of "whether a collision occurs". Instead, it calculates deep risk indicators, constructs discriminative features, and uses machine learning models to achieve early and accurate discrimination of collision types, providing a direct basis for the generation of differentiated protection strategies.

[0105] After predicting and outputting future multimodal predicted trajectories, collision detection is performed based on these trajectories and the distances between the vehicle and the target vehicle's multiple multimodal predicted trajectories. This determines whether a collision or intersection will occur between the two vehicles in the future. Then, at the most probable collision point—the point where the two predicted trajectories are closest in the spatiotemporal domain—a set of core indicators characterizing the physical nature of the collision (collision risk features) is calculated to prepare for the next step of active protection. 。

[0106] For example, in the embodiments of this application, the collision risk feature characterizes the collision risk index for predicting the collision time, including the relative velocity vector determined based on the velocity vectors of the vehicle and the target vehicle, the collision time TTC determined based on the ratio of the Euclidean distance between the vehicle and the target vehicle to the relative velocity vector, the collision angle when the two vehicles collide, determined based on the velocity vectors of the vehicle and the target vehicle, and the collision energy calculated based on the relative velocity vector and the mass of the vehicle and the mass of the target vehicle.

[0107] To avoid problems caused by angular periodicity, i.e., to eliminate the interference of angular periodicity (such as 359° being close to 1°) on the machine learning model, the collision angle is not used directly. It is not the collision angle itself, but rather in the feature vector. Simultaneous use A continuous representation of the collision direction, specifically, the predicted collision angle when the two vehicles collide. Mapping to a continuous two-dimensional space, we get: the collision angle is defined as the velocity vectors of the two vehicles at the point of collision. and The angle between them is a decisive geometric feature for distinguishing collision types (such as frontal, rear-end, and side collisions);

[0108] ;

[0109] ;

[0110] in, , These represent the velocity vectors of the vehicle and the target vehicle, respectively. A continuous representation function of the collision direction representing the collision angle;

[0111] Collision energy is This is expressed as a calculation using the following formula:

[0112] ;

[0113] in, and The masses of the target vehicle and the vehicle itself are respectively. Represents the relative velocity vector. Mass This parameter can be preset based on vehicle type. It is directly related to the severity of the collision, and the vehicle mass can be obtained from a preset database based on its type code (e.g., sedan, SUV, truck). The formula reflects the relative kinetic energy under equivalent collision mass and is directly related to vehicle deformation and the impact experienced by the occupants.

[0114] Wherein, the relative velocity vector ,in, , To predict the velocity vectors of the vehicle and the target vehicle at the moment of collision, which directly reflect the severity of the collision; the magnitude of this vector... It directly reflects the intensity of the collision (kinetic energy exchange rate), and its direction itself also implies the geometric relationship of the collision, which is the basis for subsequent calculations;

[0115] The collision time mentioned ,in To predict the relative distance between the vehicle and the target vehicle at the predicted point of impact, Time-to-Care (TTC) is the gold standard for measuring collision urgency. A shorter TTC means a smaller reaction time window, requiring a higher level of warning or protection.

[0116] The calculated raw physical indicators need to undergo feature engineering to transform them into high-dimensional features suitable for machine learning model discrimination, and finally be mapped to specific collision types through a classifier. For example, in this embodiment, the construction of a multi-dimensional judgment feature vector based on the identified collision risk features is a comprehensive multi-dimensional discriminative feature vector F;

[0117] ;

[0118] This represents the minimum distance between the predicted trajectory pairs of the current vehicle and the target vehicle, serving as an auxiliary confidence level for collision risk.

[0119] This represents the time from the current moment to the predicted collision moment. The time difference provides time-dimensional information;

[0120] This represents the current speed scalar of the vehicle.

[0121] , The type encoding of this vehicle and the target vehicle is represented (simplified to a one-dimensional scalar or embedded vector), and prior knowledge of vehicle attributes (such as the difference between a truck and a car collision) is introduced.

[0122] After constructing a multidimensional judgment feature vector, it is input into a multilayer perceptron for processing, and the probability distribution of the collision type is output. Then, based on the maximum probability of the probability distribution of the collision type, the collision type recognition result is determined, thereby determining the collision type, such as frontal collision, side collision, or rear-end collision.

[0123] For example, in the embodiments of this application, the multilayer perceptron (MLP) is composed of an input layer, a first hidden layer, a second hidden layer, and an output layer connected in sequence. The first hidden layer and the second hidden layer use the ReLU function to learn the higher-order interactions between features and extract abstract features to prevent overfitting. The output layer uses the Softmax activation function to normalize the output result into a probability distribution.

[0124] Specifically, a typical three-layer multilayer perceptron (MLP) design is as follows:

[0125] Input Layer: Contains 10 nodes, aligned with the dimension of the feature vector F.

[0126] Hidden Layer 1: Consists of 32 nodes, uses the ReLU (Rectified Linear Unit) activation function to introduce non-linearity, and learns higher-order interactions between features.

[0127] Hidden Layer 2: Includes 16 nodes, also using ReLU activation, to further refine abstract features and prevent overfitting.

[0128] Output Layer: Consists of 6 nodes, corresponding to the 6 collision types to be identified. The number of nodes can be adjusted based on the number of collision types. The Softmax activation function is used to normalize the output to a probability distribution. =[ , ,..., ], Indicates belonging to the first The probability of collision class, .

[0129] The Multilayer Perceptron (MLP) is trained using a large amount of labeled historical collision data or high-fidelity simulation data for supervised training. Each training sample contains a feature vector F and its corresponding real collision type label (one-hot encoded). The loss function uses cross-entropy loss as the optimization objective.

[0130]

[0131] in, It is the first of the real labels Bit (0 or 1) It is the model's prediction of the first Class probability. Minimizing this loss forces the probability distribution predicted by the model to approximate the true distribution.

[0132] After training, during the inference and decision-making phase in the real-time system, the constructed discriminative feature vector F is input into the pre-trained multilayer perceptron (MLP) to obtain the probability distribution; the final collision type recognition result is the category with the highest probability. Meanwhile, the maximum probability value is used as the confidence level of the identification result and is used for subsequent decision fusion or reliability judgment.

[0133] In this application, after obtaining the collision risk characteristics, a feature vector is constructed and input into a multilayer perceptron to accurately identify various collision types such as frontal collisions, side collisions, and rear-end collisions, thereby achieving high-precision trajectory prediction and enabling early and accurate identification of collision types; finally, targeted active and passive safety protection strategies are triggered based on the identification results.

[0134] In this application, after identifying the collision type, a decision is made from "whether there is a collision" to "what kind of collision and how severe the collision is". Then, based on this decision, a safety protection action corresponding to different collision risks is initiated according to a preset safety protection strategy to ensure the optimal allocation of protection resources.

[0135] This application's active protection strategy, based on high-precision trajectory prediction, refined collision type identification, and multi-dimensional risk quantification results, achieves a fundamental upgrade from a binary response of "whether there is a collision" to differentiated and adaptive protection based on "what kind of collision and what severity." By constructing a control system with multi-input fusion, computable decision-making, and hierarchical triggering, it ensures that the most suitable safety devices are activated at the optimal timing and intensity before and during a collision, maximizing occupant protection effectiveness and optimizing system resource utilization.

[0136] For example, in this application, the protection decision generation is based on multi-source information fusion and quantitative calculation. The generation of the protection decision does not rely on a single threshold, but is obtained by a real-time decision engine after fusing, calculating and comprehensively arbitrating the following multi-dimensional input information:

[0137] The decision input vector of a real-time decision engine includes:

[0138] 1. Collision types and their confidence levels (i.e., the corresponding complete probability distributions) );

[0139] 2. Collision risk indicators: Time to Collision (TTC) (used to assess the urgency level and is the core basis for triggering the sequence of events), estimated collision energy. (Used to quantify the severity of a collision, it is a key parameter for adjusting the strength of protection), Collision angle (Geometric features used to accurately distinguish collisions).

[0140] 3. Vehicle status information: including vehicle speed, gear, and occupant seating posture and position obtained through in-vehicle sensors.

[0141] The global confidence score is formed by combining the confidence score of trajectory prediction and the confidence score of collision type. This forms an assessment of the overall reliability of the current situation assessment, serving as a reliability valve for executing irreversible actions.

[0142] Based on the above inputs, the real-time decision engine uses an embedded parametric calculation model and predefined strategy mapping rules to calculate specific control commands for each actuator in real time. For example, it calculates the timing of airbag deployment. It is composed of functions = - ( , The result is obtained through dynamic calculation, where This refers to the moment when the trigger command for the airbag (or pretensioner, etc.) is issued. The predicted time of the physical collision. Indicates the magnitude of the relative velocity during the collision. For the collision angle, The time advance is a dynamically calculated, non-fixed delay based on relative velocity and collision angle optimization.

[0143] A parameterized, configurable, and differentiated protection strategy library can be built into the vehicle system. This library maps identified collision types to a series of precisely controllable execution instructions, enabling differentiated strategy execution based on collision type. For example, differentiated strategies include:

[0144] 1) For frontal / offset collisions (protection targets are to suppress forward displacement and reduce chest and head injuries):

[0145] Seat belts: Implement two levels of pretensioning: Level 1 quickly eliminates webbing slack; Level 2... The calculated target force value is used for adaptive pretensioning to secure the occupant to the seat.

[0146] Intelligent Seat: Automatically adjusts to an impact-resistant posture (the rear of the seat cushion is raised to prevent submersion, and the backrest is moderately straightened to optimize the load path).

[0147] Front airbags: Deployment timing depends on the relative speed between the vehicle and the target vehicle. Dynamic calculations ensure optimal catching; inflation level is determined based on... Single-stage / dual-stage adaptive selection; in offset collisions, asymmetrical adjustment of deployment parameters for airbags on the impact side and the non-impact side.

[0148] 2) For side impacts (the protection target is to cope with extremely short intrusion time and protect the chest, abdomen and head):

[0149] Side airbags / curtains: Based on the predicted collision point (such as the A-pillar, B-pillar, and C-pillar areas), the inflation process is initiated in advance, predictively inflating and optimizing the deployment shape. In high-energy scenarios, a high-pressure maintenance mode is triggered to extend protection.

[0150] Active seat side wings: Instantaneous expansion provides early lateral support and reduces lateral acceleration of the torso.

[0151] Vehicle attitude coordination: In conjunction with the chassis control system, the height of the suspension on the impact side is slightly adjusted at the moment of impact or instantaneously to optimize the connection of the body structure and enable the body structure to meet the collision in the best posture.

[0152] 3) Regarding rear-end collisions (the core protection objective is to prevent excessive extension of the cervical spine, i.e., whiplash injury):

[0153] Active headrest: Before the collision energy is transferred to the occupant, the headrest motor drives the headrest to move forward and upward to fill the gap between the head and the headrest, thus providing continuous support for the head.

[0154] Energy-absorbing seat mechanism: Triggers the energy-absorbing device of the seat back or slide rail to achieve controllable buffered backward movement, such as releasing the energy-absorbing mechanism of the seat back or controlling the entire seat system to move backward and sink in a buffered manner, smoothing the deceleration curve.

[0155] 4) For oblique / compound collisions (the core protection objective is to cope with complex loads and achieve all-dimensional protection):

[0156] The collision direction vector is decomposed into basic components such as frontal and lateral components; the final protection command is generated by weighted fusion of the basic strategy commands, with the weights determined by the sine and cosine values ​​of the collision direction angle, thereby achieving adaptive protection against complex collision geometry; for example, the left front oblique collision will fuse part of the frontal strategy (seatbelt pretensioning, front airbag preparation) and part of the left side strategy (left side curtain airbag deployment), with the weights determined by the sine and cosine values ​​of the collision direction angle.

[0157] For example, in this embodiment of the application, preset safety protection actions are triggered and executed in stages according to a control command sequence before the collision time, including:

[0158] The collision time (TTC) is compared with a preset threshold range, and the current warning level is determined based on the comparison result.

[0159] Based on the previous early warning level, find the pre-generated control instruction sequence for the current early warning level;

[0160] Based on the current warning level control command sequence, execute preset safety protection actions.

[0161] In this application, by making decisions for protection, a sequence of coordinated instructions with precise timestamps is output, which can ensure that devices such as seat belts, airbags, seats, and headrests work together in an optimized sequence to form a protective force; before the collision occurs, safety protection actions are executed to achieve maximum safety protection for occupants.

[0162] Furthermore, in this application, the decision-making process incorporates conflict arbitration logic, which, when faced with multiple potential threats, bases the decision on "risk = severity ( We use a comprehensive measure of “urgency (TTC)” to make global decisions and ensure that the most urgent and necessary protective actions are taken.

[0163] To achieve the optimal balance between safety and comfort, this application implements a strict tiered triggering system based on the Time to Collision (TTC), forming a progressive "warning-preparation-protection" chain. Specific time windows and execution permissions are assigned to each stage. For example, based on the calculated Time to Collision (TTC), tiered warnings and triggering are implemented according to time thresholds in the following manner:

[0164] Level 1 (TTC≥2.0s): The system enters the warning and preparation state, and issues a red warning or sound alert through the head-up display (HUD). It performs wake-up operations on the safety protection device control system, such as waking up and self-checking the relevant safety controllers; the pre-inflation system establishes the basic pressure, but does not perform physical actions of the protection devices.

[0165] Level 2 (1.0s≤TTC<2.0s): A yellow warning is issued via the head-up display (HUD), the warning signal is upgraded, and pre-action is performed. A smooth, reversible preparatory action is performed to put the safety protection device in a ready state, such as pre-tightening the seat belt by a certain distance (e.g., 5cm) to eliminate slack; the active headrest moves to the ready position; the seat begins to slowly adjust to the preset protective posture; the airbag module capacitor is charged to a ready state, preparing for final triggering; if the risk is eliminated, the system automatically resets.

[0166] Level 3 (TTC<1.0s): An audible warning is issued via the head-up display (HUD) to determine that a collision is unavoidable. The final protective devices with all optimized parameters are triggered synchronously within milliseconds (such as full pretensioning of seat belts and airbag deployment as set). At this time, the precise matching final protection plan is fully executed, and other vehicle systems (such as fuel cut-off and door unlocking) are linked to prepare for the post-collision stage.

[0167] See Figure 2As shown in the figure, this application embodiment also provides an intelligent safety protection system based on trajectory prediction and collision situation identification. This system is the hardware and software entity for implementing the above-mentioned intelligent safety protection method based on trajectory prediction and collision situation identification, and its modular architecture and functions are as follows:

[0168] The graphing module is used to construct a spatiotemporal heterogeneous graph representing the dynamic traffic scene in real time, using the vehicle and traffic participants within its real-time perception range as nodes and the potential interaction relationships between nodes as directed edges.

[0169] The trajectory prediction module is used to output the temporal enhanced feature vector of the vehicle node based on the real-time spatiotemporal heterogeneous graph and node feature vectors using a trained GAT-Transformer prediction model. This temporal enhanced feature vector deeply integrates the vehicle node's own motion history information and its spatial interaction information with the surrounding environment. The trained conditional variational autoencoder is used to perform trajectory prediction based on the temporal enhanced feature vector, and outputs multiple predicted trajectories of the vehicle and the target vehicle respectively. Each predicted trajectory includes a confidence level.

[0170] The collision recognition module is used to identify collision risk characteristics when there is a collision risk based on the predicted trajectories of the vehicle and the target vehicle; and to identify the collision type based on the collision risk characteristics.

[0171] The protection decision module is used to calculate the action time of the protection device in real time through the real-time decision engine based on the identified collision type and the preset protection strategy, and form a sequence of control commands for graded triggering.

[0172] The protection execution module is used to trigger and execute preset safety protection actions in stages according to the control command sequence before the collision time.

[0173] The trajectory prediction module's core hardware is a high-performance in-vehicle computing platform (such as NVIDIA DRIVE AGXOrin), and its software deploys and runs the GAT-Transformer trajectory prediction model, including a data processing unit (data formatting), a GAT-Transformer model (online inference), and a multimodal trajectory generation unit (trajectory decoding and output). The collision recognition module's hardware is integrated into the aforementioned in-vehicle computing platform, and its software includes a risk calculation unit (an algorithm for calculating collision risk), an MLP classifier (loading trained model files for online classification), and a confidence calibration unit (weighted averaging of recognition results from multiple trajectories to improve robustness). The protection decision module has a dedicated ECU (electronic control unit) connected to the vehicle's main CAN network in its hardware, and a built-in strategy matching unit (which queries a predefined protection strategy library) and a hierarchical triggering unit (which generates a specific control command sequence based on the TTC level) in its software. The protection execution module has its hardware connected to the vehicle's existing safety system ECUs (such as airbag control units, seat belt pretensioners, active suspension controllers, and HUD controllers) via a CAN / FD bus, and its software is responsible for translating the instructions from the decision module into standard CAN messages that each actuator can recognize, and ensuring the priority and timeliness of the control instructions.

[0174] To verify the effectiveness of the invention, a closed-track vehicle test was conducted at the test track (C-NCAP standard track) by a third-party testing agency.

[0175] Test scenario:

[0176] Designed and executed 120 valid tests, covering: 100% frontal overlap rigid wall collision (reference); 40% frontal offset deformable wall collision; side moving trolley collision (left / right); vehicle rear-end collision with stationary target vehicle; and oblique collision at intersection.

[0177] Compare to baseline:

[0178] A traditional collision warning system based on TTC and MRC (vehicle movement radius) used in a mainstream L2-level assisted driving vehicle on the market was selected as the comparison baseline.

[0179] Experimental results:

[0180] 1. Collision type recognition accuracy: The overall accuracy of the method of this invention reaches 96.3%, significantly higher than the baseline system's 63.6%, representing an improvement of 32.7 percentage points. It is particularly advantageous in distinguishing between frontal collisions and frontal offset collisions, as well as in recognizing oblique collisions.

[0181] 2. Average warning time: Based on the predicted trajectory, this invention can issue the first warning on average 2.8 seconds before a physical collision occurs, which is 0.8 seconds earlier than the baseline system (average 2.0 seconds), allowing the system more preparation time.

[0182] 3. System false alarm rate: During a week-long urban road test (approximately 1000 kilometers), the system of this invention was able to better distinguish between real conflicts and normal overtaking and lane changing scenarios, with a false alarm rate of 2.1%, which is far lower than the baseline system's 9.5%, effectively improving system credibility and user experience.

[0183] Compared with the prior art, the present invention has the following beneficial effects:

[0184] This invention offers high collision recognition accuracy: by fusing trajectory prediction and machine learning classification, it can accurately identify multiple collision types. Its protection strategies are highly targeted: providing personalized protection strategies based on the characteristics of different collision types. The system is highly adaptable: able to adapt to complex and ever-changing traffic environments. It provides ample warning time: through accurate trajectory prediction, it gains valuable pre-action time for the safety system.

[0185] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects. The scope of the present invention is defined by the appended claims rather than the foregoing description, and therefore all changes falling within the meaning and scope of the equivalents of the claims are intended to be included within the present invention.

[0186] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. An intelligent safety protection method based on trajectory prediction and collision situation identification, characterized in that, Include: Using the vehicle and the traffic participants within its real-time perception range as nodes, and the potential interaction relationships between nodes as directed edges, a structured spatiotemporal heterogeneous graph representing the dynamic traffic scene is constructed in real time. Using the GAT-Transformer prediction model, based on the real-time spatiotemporal heterogeneous graph and node feature vectors, the temporal enhanced feature vector of the vehicle node is output. The temporal enhancement feature vector is deeply integrated with the vehicle's own motion history information and its spatial interaction information with the surrounding environment; the conditional variational autoencoder is used to predict the trajectory based on the temporal enhancement feature vector, and outputs multiple predicted trajectories of the vehicle and the target vehicle, each predicted trajectory including a confidence level; Based on the predicted trajectories of the vehicle and the target vehicle, if a collision risk is identified, collision risk features are identified. A multi-dimensional judgment feature vector is constructed based on the identified collision risk features and input into a multilayer perceptron for processing. The collision type is determined by using the maximum probability of the probability distribution of the collision type output by the multilayer perceptron. Based on the identified collision type and the preset protection strategy, the action time of the protection device is calculated in real time by the real-time decision engine to form a sequence of control commands for graded triggering; the preset safety protection actions are triggered and executed in a graded manner according to the control command sequence before the collision time. A directed edge representing the potential interaction relationship between nodes is constructed using a dynamic connection strategy based on preset rules; The preset rules include: Spatial proximity rule: If the Euclidean distance between two nodes is less than a threshold, then a bidirectional edge is established between the two nodes; Lane topology rules: Based on high-precision maps or visually perceived lane line information, if two nodes are in the same lane, adjacent lanes, or lanes with path conflicts, a connection edge is established. Each directed edge carries a node from the source node. Point to target node Standardized relative motion information, i.e., edge features This serves as a basis for assessing the intensity of interaction: ； in, Representing the source node respectively Relative to the target node The difference in two-dimensional coordinates, velocity, and heading angle; The conditional variational autoencoder, when performing trajectory prediction, uses a prior network to generate a latent variable distribution based on the temporal augmentation feature vector, and samples from it. Hidden variables Each hidden variable The temporal augmentation feature vectors are input into the decoder to generate the future. The trajectory of the steps At the same time, for each trajectory Output a raw confidence score Based on the original confidence score The confidence level is obtained by normalizing the probability using the Softmax function. ; = ； ， ； in, Indicates the first time step in the future to the first The x and y coordinates of the two-dimensional position of the trajectory at the time step. It is the loop variable in the summation symbol ∑, representing the sequence from the 1st to the 2nd. The original confidence score of any one of the trajectories.

2. The intelligent safety protection method based on trajectory prediction and collision situation identification according to claim 1, characterized in that, The state of each node at time t is represented by a multidimensional feature vector. The description, encompassing kinematic states and semantic attributes, is represented as follows: ； in, These are two-dimensional position coordinates in the vehicle coordinate system; and These are velocity and acceleration vectors, respectively. For heading angle; This is a traffic participant type identifier that has been uniquely encoded.

3. The intelligent safety protection method based on trajectory prediction and collision situation identification according to claim 1, characterized in that, The conditional variational autoencoder is trained using the following loss function. : ； in, The average displacement error is used to minimize the average distance between the predicted point and the true point at the predicted time step. The final displacement error is used to minimize the distance between the predicted endpoint and the actual endpoint; For the confidence loss, a negative log-likelihood loss is used to encourage the conditional variational autoencoder to assign higher confidence scores to predictions that are closer to the true trajectory. ; KL divergence is used to normalize the posterior and prior distributions of latent variables in a conditional variational autoencoder and is the core regularization term for training a conditional variational autoencoder. , , , Weighting coefficients to balance the various losses.

4. The intelligent safety protection method based on trajectory prediction and collision situation identification according to claim 1, characterized in that, The collision risk characteristics characterize the collision risk index for predicting the moment of collision, including the relative velocity vector determined based on the velocity vectors of the vehicle and the target vehicle, the collision time TTC determined based on the ratio of the Euclidean distance between the vehicle and the target vehicle to the relative velocity vector, the collision angle at the moment of collision determined based on the velocity vectors of the vehicle and the target vehicle, and the collision energy calculated based on the relative velocity vector and the mass of the vehicle and the target vehicle; wherein, the collision angle at the moment of collision is predicted. After mapping to a continuous two-dimensional space, a vector is obtained as the final collision angle: ； ; in, , These represent the velocity vectors of the vehicle and the target vehicle, respectively. Represents an angle function; Collision energy is This is expressed as a calculation using the following formula: ； in, and The masses of the target vehicle and the vehicle itself are respectively. This represents the relative velocity vector.

5. The intelligent safety protection method based on trajectory prediction and collision situation identification according to claim 4, characterized in that, The multidimensional judgment feature vector F, constructed based on the identified collision risk features, is expressed as follows: ； This represents the minimum distance between the predicted trajectory pairs of this vehicle and the target vehicle; This represents the time from the current moment to the predicted collision moment. Time difference; This represents the current speed scalar of the vehicle. , Indicates the type codes of this vehicle and the target vehicle; The multidimensional judgment feature vector is input into the multilayer perceptron for processing and outputs the probability distribution of collision types. Then, based on the maximum probability of the probability distribution of collision types, i.e. the confidence level of the collision type, the recognition result of the collision type is determined.

6. The intelligent safety protection method based on trajectory prediction and collision situation identification according to claim 1, characterized in that, The multilayer perceptron consists of an input layer, a first hidden layer, a second hidden layer, and an output layer connected in sequence. The first hidden layer and the second hidden layer use the ReLU function to learn the higher-order interactions between features and extract abstract features to prevent overfitting. The output layer uses the Softmax activation function to normalize the output result into a probability distribution.

7. The intelligent safety protection method based on trajectory prediction and collision situation identification according to claim 6, characterized in that... Prior to the collision time, preset safety protection actions are triggered and executed in stages according to the control command sequence, including: The collision time (TTC) is compared with a preset threshold range, and the current warning level is determined based on the comparison result. Based on the previous early warning level, find the pre-generated control instruction sequence for the current early warning level; Execute preset safety protection actions according to the control command sequence of the current early warning level; The control command sequence is calculated in real time by the real-time decision engine based on the collision type and its confidence level, collision risk characteristics, confidence level of the predicted trajectory, vehicle speed, gear position, and the position and posture of the occupants.

8. An intelligent safety protection system based on trajectory prediction and collision situation identification, characterized in that, The intelligent safety protection method based on trajectory prediction and collision situation identification as described in any one of claims 1-7 includes: The graphing module is used to construct a spatiotemporal heterogeneous graph representing the dynamic traffic scene in real time, using the vehicle and traffic participants within its real-time perception range as nodes and the potential interaction relationships between nodes as directed edges. The trajectory prediction module uses the GAT-Transformer prediction model to output the temporal enhanced feature vector of the vehicle node based on the real-time spatiotemporal heterogeneous graph and node feature vectors. This temporal enhanced feature vector deeply integrates the vehicle node's own motion history information and its spatial interaction information with the surrounding environment. The conditional variational autoencoder performs trajectory prediction based on the temporal enhanced feature vector and outputs multiple predicted trajectories of the vehicle and the target vehicle. Each predicted trajectory includes a confidence level. The collision recognition module is used to identify collision risk features when the predicted trajectories of the vehicle and the target vehicle indicate a collision risk. A multi-dimensional judgment feature vector is constructed based on the identified collision risk features and input into a multilayer perceptron for processing. The collision type is determined by using the maximum probability of the probability distribution of the collision type output by the multilayer perceptron. The protection decision module is used to calculate the action time of the protection device in real time through the real-time decision engine based on the identified collision type and the preset protection strategy, and form a sequence of control commands for graded triggering. The protection execution module is used to trigger and execute preset safety protection actions in stages according to the control command sequence before the collision time.

Images