Automatic emergency brake control method and device for vehicle and vehicle

Abstract

The application provides an automatic emergency braking control method and device of a vehicle and the vehicle, and relates to the active safety technical field of the vehicle. The method comprises the following steps: acquiring motion state information and environment state information of a vehicle in a driving environment, and performing fusion coding to obtain context query features; acquiring detection data of at least two types of vehicle sensors; for each sensor, obtaining a feature matrix comprising a plurality of candidate target object instance features based on the detection data, and determining a path contribution coefficient according to the context query features and the sensor type; taking the context query features as a shared query vector, performing attention threat evaluation on the instance features, adjusting the attention contribution of the instance features by using the path contribution coefficient, and obtaining threat evaluation features; fusing the threat evaluation features of all paths to obtain comprehensive threat evaluation features, and outputting an automatic emergency braking prediction signal. Through the application, the accuracy and stability of automatic emergency braking prediction in a complex environment can be improved.

Description

Technical Field

[0001] This invention relates to the field of vehicle active safety technology, and in particular to an automatic emergency braking control method, device and vehicle for a vehicle. Background Technology

[0002] Automatic Emergency Braking (AEB) is a crucial component of a vehicle's active safety system. It typically requires real-time assessment of collision risk based on surrounding objects, the vehicle's motion, and road conditions, and outputs automatic emergency braking prediction or control signals when a collision risk is identified. To improve the accuracy and robustness of environmental perception, existing AEB systems have increasingly incorporated various vehicle sensors, such as cameras, lidar, and millimeter-wave radar, to acquire information about surrounding objects through multi-sensor fusion.

[0003] Existing multi-sensor fusion methods mainly include post-fusion and early feature stitching. Post-fusion typically involves each vehicle sensor separately performing target detection, target tracking, or obstacle recognition, generating structured detection results such as bounding boxes, target categories, confidence levels, and motion states. Then, in the decision-making stage, the detection results from different sensors are weighted and fused or rule-based judgments are applied. The inventors' research found that this method heavily relies on the detection performance of each vehicle sensor individually. Furthermore, because fusion occurs at a higher semantic level, fine-grained information in the original sensor features or intermediate instance features is easily lost, making it difficult to fully utilize the complementary information between different sensors and limiting the model's ability to autonomously determine the reliability of different sensors based on the context.

[0004] Early feature stitching methods typically involved directly stitching together features extracted from different vehicle sensors and inputting them into subsequent networks to achieve fusion at the feature level. However, the inventors found that this method is mostly static weight fusion, which does not fully consider the reliability differences of the perception paths of various vehicle sensors under different environmental conditions, lacks a dynamic weight allocation mechanism, and does not have adaptive capabilities.

[0005] In summary, in multi-sensor AEB scenarios, how to achieve dynamic threat assessment based on the real-time changes in the reliability of the sensing paths of each sensor, and improve the accuracy and stability of automatic emergency braking prediction signals in complex environments, has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0006] The purpose of this invention is to provide an automatic emergency braking control method, device, and vehicle for solving the aforementioned technical problems in the prior art.

[0007] On the one hand, in order to achieve the above objectives, the present invention provides an automatic emergency braking control method, device and vehicle for a vehicle.

[0008] The automatic emergency braking control method for this vehicle includes: acquiring vehicle motion state information, environmental state information, and detection data from at least two different types of vehicle sensors in the vehicle's driving environment, wherein the environmental state information is used to characterize the environmental state affecting the reliability of the perception path corresponding to at least one vehicle sensor; fusing and encoding the vehicle motion state information and environmental state information to obtain context query features, wherein the context query features are used to characterize the automatic emergency braking decision context under the combined effect of the current vehicle motion state and the current environmental state; for each vehicle sensor: performing perception processing on the detection data of the vehicle sensor to obtain a feature matrix, determining the path contribution coefficient based on the context query features and sensor type, wherein the feature matrix includes instance features corresponding to multiple candidate target objects, using the context query features as a shared query vector, performing attention threat assessment on each instance feature to obtain threat assessment features, wherein the path contribution coefficient is used to adjust the attention contribution of the instance features in the attention threat assessment; fusing the threat assessment features of each vehicle sensor to obtain a comprehensive threat assessment feature; and outputting an automatic emergency braking prediction signal based on the comprehensive threat assessment feature.

[0009] Furthermore, at least two different types of vehicle sensors, including cameras and LiDAR, are used. The detection data includes camera image data and LiDAR point cloud data. For cameras, a path contribution coefficient is used to characterize the contribution of the camera's perception path to the automatic emergency braking threat assessment under the current vehicle motion state and the current environmental state. For LiDAR, a path contribution coefficient is used to characterize the contribution of the LiDAR's perception path to the automatic emergency braking threat assessment under the current vehicle motion state and the current environmental state. Environmental state information includes information on at least one of the following: weather state, illumination state, visibility state, precipitation state, fog / haze state, road reflection state, and sensor occlusion state. Vehicle motion state information includes at least one of the following: lateral velocity, longitudinal velocity, lateral acceleration, longitudinal acceleration, and heading angle.

[0010] Furthermore, the steps for determining the path contribution coefficient based on context query features and sensor type include: fusing the type codes corresponding to the context query features and sensor type to obtain path context features; and performing feature mapping on the path context features through a multilayer perceptron to obtain the path contribution coefficient.

[0011] Furthermore, the steps for fusing and encoding the vehicle motion state information and environmental state information to obtain context query features include: encoding the vehicle motion state information to obtain vehicle state encoding features; extracting environmental state information from the vehicle's forward-facing camera image and encoding the environmental state information to obtain environmental encoding features; concatenating the vehicle state encoding features and environmental encoding features to obtain concatenated features; and mapping the concatenated features through a fusion network to obtain context query features.

[0012] Furthermore, the step of using context query features as a shared query vector to perform attention threat assessment on each instance feature to obtain threat assessment features includes: using context query features as a shared query vector; for each instance feature: mapping the instance feature to a key and a value, obtaining an initial attention score corresponding to the instance feature based on the correlation between the shared query vector and the key, adjusting the initial attention score based on the path contribution coefficient to obtain an adjusted attention weight; and performing weighted fusion based on the adjusted attention weights corresponding to each instance feature and the values ​​corresponding to each instance feature to obtain the threat assessment features.

[0013] Furthermore, the attention threat assessment employs a multi-head cross-attention mechanism. This method further includes: for each attention head, performing linear mapping on the shared query vector, the keys corresponding to instance features, and the values, respectively, to obtain the query sub-vector, key vector, and value sub-vector corresponding to that attention head; for any instance feature under each attention head, the step of obtaining the adjusted attention weight includes: obtaining the initial intra-head attention score corresponding to the instance feature based on the correlation between the query sub-vector and the key vector, adjusting the initial intra-head attention score based on the path contribution coefficient to obtain the adjusted intra-head attention weight; the step of obtaining the threat assessment feature includes: for each attention head, performing weighted fusion based on the adjusted intra-head attention weights corresponding to each instance feature and the value sub-vectors corresponding to each instance feature to obtain the attention output corresponding to that attention head, concatenating and outputting the attention outputs corresponding to each attention head to obtain the threat assessment feature.

[0014] Furthermore, the neural network model includes a perceptual processing module for obtaining the feature matrix, a fusion encoding module for obtaining contextual query features, an attention-based threat assessment module for obtaining threat assessment features, a threat fusion module for obtaining comprehensive threat assessment features, and a prediction output module for outputting an automatic emergency braking prediction signal. The steps for training the neural network model include: acquiring training samples, which include sample detection data, sample vehicle motion state information, sample environmental state information, and automatic emergency braking labels; inputting the sample detection data, sample vehicle motion state information, and sample environmental state information into the neural network model to obtain the sample automatic emergency braking prediction signal; and calculating the loss value based on the sample automatic emergency braking prediction signal and the automatic emergency braking label, wherein the loss value is based on Focal... The network parameters of at least one of the following modules—the perception processing module, the fusion encoding module, the attention threat assessment module, the threat fusion module, and the prediction output module—are updated based on the loss value. Specifically, when the neural network model is trained end-to-end, the network parameters of the perception processing module, the fusion encoding module, the attention threat assessment module, the threat fusion module, and the prediction output module are jointly updated based on the loss value. When the neural network model is trained in a phased manner, the perception processing module is trained first, followed by the fusion encoding module, the attention threat assessment module, the threat fusion module, and the prediction output module.

[0015] Furthermore, the step of outputting an automatic emergency braking prediction signal based on the comprehensive threat assessment features includes: inputting the comprehensive threat assessment features into a multilayer perceptron to obtain a risk prediction value; normalizing the risk prediction value using a sigmoid function to obtain an automatic emergency braking trigger probability; and outputting an automatic emergency braking prediction signal based on the automatic emergency braking trigger probability.

[0016] On the other hand, in order to achieve the above objectives, the present invention provides an automatic emergency braking control device for a vehicle.

[0017] The automatic emergency braking control device of the vehicle includes: an acquisition module for acquiring vehicle motion state information, environmental state information, and detection data from at least two different types of vehicle sensors in the vehicle's driving environment, wherein the environmental state information is used to characterize the environmental state affecting the reliability of the perception path corresponding to at least one vehicle sensor; a fusion encoding module for fusing and encoding the vehicle motion state information and the environmental state information to obtain context query features, wherein the context query features are used to characterize the automatic emergency braking decision context under the combined effect of the current vehicle motion state and the current environmental state; and a perception processing module for performing perception processing on the detection data of each vehicle sensor to obtain the vehicle motion state of that path. The system comprises: a feature matrix corresponding to vehicle sensors, wherein the feature matrix includes instance features corresponding to multiple candidate target objects; an attention threat assessment module, used to determine path contribution coefficients based on the context query features and sensor type, using the context query features as a shared query vector, and performing attention threat assessment on each instance feature to obtain threat assessment features, wherein the path contribution coefficients are used to adjust the attention contribution of the instance features in the attention threat assessment; a threat fusion module, used to fuse the threat assessment features of each of the vehicle sensors to obtain comprehensive threat assessment features; and a prediction output module, used to output an automatic emergency braking prediction signal based on the comprehensive threat assessment features.

[0018] On the other hand, in order to achieve the above objectives, the present invention also provides a vehicle.

[0019] The vehicle includes at least two different types of vehicle sensors and a processor, wherein the processor executes any of the automatic emergency braking control methods for vehicles provided by the present invention.

[0020] The automatic emergency braking control method, device, and vehicle provided by this invention first acquire vehicle motion state information, environmental state information, and detection data from at least two different types of vehicle sensors in the vehicle's driving environment. The vehicle motion state information and environmental state information are then fused and encoded into contextual query features. Next, for each vehicle sensor, the detection data of that sensor is processed to obtain a feature matrix including instance features corresponding to multiple candidate target objects. A path contribution coefficient is determined based on the contextual query features and sensor type. Subsequently, the contextual query features are used as a shared query vector to perform attention threat assessment on the instance features corresponding to that vehicle sensor. During the attention threat assessment process, the path contribution coefficient is used to adjust the attention contribution of the instance features, resulting in threat assessment features corresponding to that vehicle sensor. Finally, the threat assessment features of each vehicle sensor are fused to obtain a comprehensive threat assessment feature, and an automatic emergency braking prediction signal is output based on the comprehensive threat assessment feature. This invention performs attention threat assessment at the level of candidate target instance features corresponding to various vehicle sensors, thereby preserving and utilizing fine-grained feature information of candidate targets and enhancing the information complementarity between different sensor perception paths. It fuses vehicle motion state information and environmental state information to form context query features, and determines path contribution coefficients based on these features and sensor type. These path contribution coefficients reflect the contribution of each perception path to the automatic emergency braking threat assessment under the current vehicle motion state and environmental conditions. Furthermore, during the attention threat assessment process, the path contribution coefficients are used to adjust the attention contribution of instance features, ensuring that when the reliability of a perception path corresponding to a particular vehicle sensor changes due to environmental conditions, the contribution of instance features in that perception path to the threat assessment features is adjusted accordingly. In summary, this invention enables context-driven dynamic threat assessment at the level of candidate target instance features, improving the accuracy and stability of automatic emergency braking prediction signals in complex environments. Attached Figure Description

[0021] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart of an automatic emergency braking control method for a vehicle provided in Embodiment 1 of the present invention; Figure 2 This is a block diagram of the automatic emergency braking control system for a vehicle provided in Embodiment 2 of the present invention; Figure 3 This is a block diagram of an automatic emergency braking control device for a vehicle provided in Embodiment 3 of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0023] Example 1 This invention provides an automatic emergency braking control method for vehicles. This method combines vehicle motion state information, environmental state information, and detection data from multiple different types of vehicle sensors in the driving environment to perform threat assessment on candidate targets and output an automatic emergency braking prediction signal. Specifically, Figure 1 The flowchart of the automatic emergency braking control method for a vehicle provided in Embodiment 1 of the present invention is as follows: Figure 1 As shown, the automatic emergency braking control method for a vehicle provided in this embodiment includes the following steps S101 to S106.

[0024] Step S101: Acquire vehicle motion state information, environmental state information, and detection data from at least two different types of vehicle sensors in the vehicle driving environment.

[0025] The vehicle motion state information is used to characterize the vehicle's motion state at the current moment or within the current time window. Optionally, the vehicle motion state information includes at least one of the vehicle's lateral velocity, longitudinal velocity, lateral acceleration, longitudinal acceleration, and heading angle, and may further include information such as steering wheel angle, brake pedal state, accelerator pedal state, and vehicle position information. Through the vehicle motion state information, it is possible to reflect whether the vehicle is currently traveling at high speed, accelerating rapidly, decelerating rapidly, steering, changing lanes, or in other motion states related to collision risk.

[0026] Environmental state information is used to characterize the environmental state affecting the reliability of the perception path corresponding to at least one vehicle sensor. A perception path refers to the processing path from acquiring detection data from the vehicle sensor, extracting features from the detection data, generating instance features of candidate targets, and participating in subsequent threat assessment. For example, when the vehicle sensor is a camera, the perception path for the camera includes camera image acquisition, image feature extraction, generation of candidate target instance features, and threat assessment based on these instance features; when the vehicle sensor is a LiDAR, the perception path for the LiDAR includes point cloud acquisition, point cloud feature extraction, generation of candidate target instance features, and threat assessment based on these instance features.

[0027] Optionally, environmental state information includes information on at least one of the following: weather conditions, lighting conditions, visibility conditions, precipitation conditions, fog / haze conditions, road reflection conditions, and sensor occlusion conditions. Environmental state information can be obtained through onboard camera images, onboard environmental sensors, vehicle external communication data, map data, or state data provided by the vehicle control system. For example, conditions such as strong backlight, low-light conditions at night, rain / snow, fog / haze, road glare, or sensor occlusion can all affect the perception reliability of candidate target objects for at least one vehicle sensor's corresponding perception path.

[0028] At least two different types of vehicle sensors refer to two or more types of sensors installed on a vehicle for sensing the vehicle's driving environment. Optionally, the at least two different types of vehicle sensors include cameras and LiDAR, and correspondingly, the detection data includes camera image data acquired by the camera and LiDAR point cloud data acquired by the LiDAR. Camera image data may include monocular camera images, binocular camera images, or multi-view camera images. LiDAR point cloud data may include position information, reflection intensity information, timestamp information, etc., of multiple points in the three-dimensional space surrounding the vehicle.

[0029] Step S102: The vehicle motion state information and environmental state information are fused and encoded to obtain context query features.

[0030] The context query feature is used to characterize the automatic emergency braking decision context under the combined influence of the current vehicle motion state and the current environmental state. In other words, the context query feature does not only represent the vehicle's own motion state, nor only the external environmental state, but is a comprehensive context feature obtained by fusing the vehicle's motion state and the environmental state. This context query feature is used for subsequent attention threat assessment of the instance features corresponding to each vehicle sensor, enabling the threat assessment process to simultaneously consider the impact of the current vehicle's own motion state and the current environmental state on the reliability of the perception path.

[0031] Optionally, the vehicle motion state information is numerically normalized, feature embedded, or mapped using a multilayer perceptron to obtain vehicle state coding features; the environmental state information is feature extracted and encoded to obtain environmental coding features; and the vehicle state coding features and environmental coding features are then concatenated, added, weighted fused, or mapped using a neural network to obtain context query features.

[0032] For example, in one implementation, the vehicle's numerical states, such as speed, acceleration, and heading angle, are combined to form a vehicle state vector, which is then mapped to vehicle state coding features through a linear or multilayer perceptron. An environmental feature vector is extracted from the vehicle's forward-facing camera image, and the environmental feature vector is encoded to obtain environmental coding features. Then, the vehicle state coding features and environmental coding features are concatenated, and the concatenated features are mapped through a fusion network to obtain context query features.

[0033] Through this step, the context query features can carry information such as how the current vehicle is moving and whether the current environment affects the reliability of sensor perception, providing a basis for dynamically monitoring high-threat candidate targets in different vehicle sensor paths.

[0034] Step S103: For each vehicle sensor, perform perception processing on the detection data of the vehicle sensor to obtain a feature matrix.

[0035] The feature matrix includes instance features corresponding to multiple candidate target objects. Candidate target objects refer to objects in the vehicle's driving environment that may affect the vehicle's driving safety, such as vehicles ahead, vehicles to the side, pedestrians, cyclists, static obstacles, and road debris. Instance features are high-dimensional features used to characterize a candidate target object, which may include the candidate target object's semantic information, location information, and motion information. Optionally, the feature matrix has a dimension of N×D, where N represents the number of candidate target objects obtained by the vehicle's sensors along the perception path, and D represents the feature dimension of the instance features corresponding to each candidate target object. Each row or group of features in the feature matrix can correspond to an instance feature of a candidate target object.

[0036] In this embodiment, at least two different types of vehicle sensors may include a camera and a LiDAR. Accordingly, image feature extraction is performed on the camera image data to obtain the feature matrix corresponding to the camera; point cloud feature extraction is performed on the LiDAR point cloud data to obtain the feature matrix corresponding to the LiDAR.

[0037] Camera image data includes multi-view camera images captured by multiple cameras on the vehicle. These cameras can be positioned in front of, behind, to the left, to the right, or around the vehicle to capture images of the vehicle's driving environment from different perspectives. Multi-view camera images can cover environmental areas in different directions around the vehicle, enabling the camera's perception path to obtain more complete environmental perception information. Multi-view camera images can be images captured by multiple cameras at the same time or within the same time window, or they can be multiple frames of images after time synchronization.

[0038] When extracting image features from camera image data, multi-view camera images are input into a convolutional neural network (CNN) to extract image features corresponding to each view. The CNN is used to extract features such as edges, textures, contours, local target structures, and target semantics. Furthermore, a feature pyramid network is used to fuse the image features output by the CNN at multiple scales, resulting in multi-scale image features. Multi-scale image features refer to image features obtained at different resolution levels. High-resolution image features are better suited to preserving detailed target information, while low-resolution image features are better suited to expressing high-level semantic information of the target. By extracting multi-scale image features, the recognition needs of near-field targets, small-sized targets, and far-field targets can be met.

[0039] After obtaining multi-scale image features, multiple initial 3D anchor points are generated in the 3D space surrounding the vehicle. These initial 3D anchor points represent the initial candidate positions of the candidate target objects in the 3D space surrounding the vehicle. Optionally, the initial 3D anchor points can be generated in the vehicle coordinate system according to a preset spatial range and preset spatial interval, or they can be obtained by mapping a learnable target query vector, or generated based on the candidate target response region in the multi-scale image features. The initial 3D anchor points may include 3D position coordinates, and may further include initial parameters such as size, orientation, or velocity used to describe the candidate target objects.

[0040] The initial 3D anchor point is then projected onto the image plane corresponding to each camera view, and corresponding image features are sampled or aggregated from multi-scale image features based on the projection position. Since the same 3D spatial location may be observed by one or more camera views, image features corresponding to the initial 3D anchor point can be extracted from the multi-scale image features of one or more views based on the projection results of the initial 3D anchor point in multiple camera views. For image features extracted from the same initial 3D anchor point at different scale levels and in different views, methods such as stitching, weighted summation, attention fusion, or neural network mapping can be used to fuse them to obtain candidate target features corresponding to the initial 3D anchor point.

[0041] Furthermore, the initial 3D anchor points and their corresponding feature representations are iteratively updated based on the candidate target features to obtain camera instance features corresponding to multiple candidate target objects. Specifically, the candidate target features are input into an update network, which predicts the positional offset, size offset, orientation offset, or motion state offset of the 3D anchor points, and updates the initial 3D anchor points based on the prediction results. Simultaneously, feature mapping and fusion updates are performed on the candidate target features to obtain updated camera instance features. After one or more iterations, the camera instance features corresponding to each candidate target object can more accurately represent the information of that candidate target object. After obtaining the camera instance features corresponding to multiple candidate target objects, the multiple camera instance features are arranged in a predetermined order to obtain the feature matrix corresponding to the camera.

[0042] LiDAR point cloud data comprises three-dimensional spatial point clouds collected by a LiDAR system on a vehicle, showing the area surrounding the vehicle. The LiDAR system can be positioned on the top, front, side, or other locations of the vehicle. It emits laser beams into the surrounding environment and receives reflected signals to obtain spatial location information of targets around the vehicle. LiDAR point cloud data can include multiple point clouds, each containing three-dimensional coordinates and potentially further information such as reflection intensity, timestamp, scan line number, or echo count. LiDAR point cloud data directly reflects the distribution of targets around the vehicle in three-dimensional space, facilitating the characterization of their position, size, and spatial contours.

[0043] When extracting point cloud features from LiDAR point cloud data, the LiDAR point cloud data is first preprocessed. Optionally, the preprocessing includes at least one of coordinate system transformation, point cloud range clipping, invalid point filtering, ground point filtering, point cloud denoising, and point cloud normalization. Coordinate system transformation is used to transform the point cloud data in the LiDAR coordinate system to the vehicle coordinate system or a unified three-dimensional coordinate system; point cloud range clipping is used to retain point cloud points within a predetermined spatial range around the vehicle; invalid point filtering and point cloud denoising are used to remove points with reflection anomalies, distance anomalies, or noise, thereby improving the stability of subsequent point cloud feature extraction.

[0044] After point cloud preprocessing, the point cloud data is spatially organized to facilitate feature extraction by the neural network. Optionally, the three-dimensional space surrounding the vehicle can be divided into multiple voxel units, and point cloud points can be assigned to the corresponding voxel units to obtain a voxelized point cloud representation; alternatively, the space surrounding the vehicle can be divided into multiple columnar units along the horizontal direction to obtain a columnar point cloud representation; or the point cloud data can be projected onto a bird's-eye view plane to obtain a bird's-eye view point cloud representation. Voxelized point cloud representation can preserve the three-dimensional structural information of the target object, while columnar point cloud representation and bird's-eye view point cloud representation are beneficial to improving feature extraction efficiency and meeting the real-time requirements of vehicle-mounted automatic emergency braking scenarios.

[0045] Then, point feature encoding is performed on the point cloud points within each voxel or columnar unit to obtain local point cloud features. Point feature encoding can include encoding the 3D coordinates, reflection intensity, offset of the point cloud point relative to the center of its voxel or columnar unit, and offset of the point cloud point relative to the center of the point cloud cluster. Further, the local point cloud features are processed through a point feature encoding network, a sparse 3D convolutional network, a columnar feature extraction network, a bird's-eye view feature extraction network, or a deep neural network to obtain a point cloud feature map. The point cloud feature map is used to characterize the point cloud distribution features and target structure features at different locations in the 3D space surrounding the vehicle.

[0046] After obtaining the point cloud feature map, multiple initial 3D anchor points are generated in the 3D space surrounding the vehicle. These initial 3D anchor points represent the initial candidate positions of the target objects in the 3D space around the vehicle. Based on the spatial position of the initial 3D anchor points in the point cloud feature map, corresponding point cloud features can be sampled or aggregated. Specifically, for each initial 3D anchor point, its corresponding position in the voxel feature map, histogram feature map, or bird's-eye view feature map can be determined, and point cloud features can be extracted from that position and its neighborhood. For point cloud features extracted from the same initial 3D anchor point at different scale levels or in different spatial neighborhoods, methods such as stitching, weighted summation, pooling, attention fusion, or neural network mapping can be used to fuse them to obtain the candidate target features corresponding to that initial 3D anchor point.

[0047] Furthermore, the initial 3D anchor points and their corresponding feature representations are iteratively updated based on the candidate target features to obtain LiDAR instance features corresponding to multiple candidate target objects. Specifically, the candidate target features can be input into an update network, which predicts the positional offset, size offset, orientation offset, or motion state offset of the 3D anchor points, and updates the initial 3D anchor points based on the prediction results. Simultaneously, feature mapping and fusion updates can be performed on the candidate target features to obtain updated LiDAR instance features. After one or more iterations, the LiDAR instance features corresponding to each candidate target object can more accurately represent the information of that candidate target object. After obtaining the LiDAR instance features corresponding to multiple candidate target objects, the multiple LiDAR instance features are arranged in a predetermined order to obtain the feature matrix corresponding to the LiDAR.

[0048] The feature matrices corresponding to the aforementioned cameras and LiDAR are used as inputs for subsequent attention-based threat assessment. In this assessment, contextual query features serve as shared query vectors, and attention analysis is performed on the instance features in each feature matrix to determine the importance of each candidate target object to the automatic emergency braking decision under the current vehicle motion state and environmental conditions. Because the feature matrices retain high-dimensional instance features of candidate targets, rather than only including structured detection results such as bounding boxes, target categories, or confidence levels, they provide richer, more fine-grained target information for subsequent threat assessment.

[0049] Step S104: For each vehicle sensor, determine the path contribution coefficient based on context query features and sensor type.

[0050] The path contribution coefficient characterizes the contribution of the perception path corresponding to the vehicle's sensors on that road to the automatic emergency braking threat assessment, given the current vehicle motion state and the current environmental state. Sensor type characterizes the category of the vehicle's sensors on that road, such as camera type, LiDAR type, millimeter-wave radar type, etc.

[0051] In this embodiment, since the perception reliability of different vehicle sensors may vary under different vehicle and environmental conditions, the contribution of the perception paths corresponding to different vehicle sensors to the automatic emergency braking threat assessment may also differ. For example, in low-light or strong backlight environments, the reliability of the camera-corresponding perception path may decrease; in rain, snow, fog, or environments with strong laser scattering, the reliability of the lidar-corresponding perception path may decrease. Therefore, the corresponding path contribution coefficient for each vehicle sensor can be determined based on contextual query features and sensor type.

[0052] Optionally, a corresponding type code is set for different sensor types, and the context query features are fused with the type code corresponding to the vehicle sensor on that road to obtain the path context features. Then, the path context features are processed by a multilayer perceptron, a linear mapping network, or other trainable feature mapping network to obtain the path contribution coefficient corresponding to the vehicle sensor on that road.

[0053] For example, for camera paths, the context query features are concatenated with the camera type code to obtain camera path context features, which are then output as camera path contribution coefficients via a multilayer perceptron. For LiDAR paths, the context query features are concatenated with the LiDAR type code to obtain LiDAR path context features, which are then output as LiDAR path contribution coefficients via a multilayer perceptron. The camera path contribution coefficient and the LiDAR path contribution coefficient are used to adjust the attention contribution of camera instance features and LiDAR instance features in attention threat assessment, respectively.

[0054] The path contribution coefficient can be either a scalar or a vector. When the path contribution coefficient is a scalar, it adjusts the overall attention contribution of all instance features corresponding to the vehicle sensor on that road; when the path contribution coefficient is a vector, it adjusts the different feature dimensions or different attention heads in the instance features corresponding to the vehicle sensor on that road in a differentiated manner.

[0055] Step S105: For each vehicle sensor, use the context query features as a shared query vector to perform attention threat assessment on each instance feature to obtain threat assessment features.

[0056] In the attention threat assessment, the path contribution coefficient is used to adjust the attention contribution of instance features. The shared query vector refers to the context query feature obtained in step S102. This shared query vector can be used simultaneously for the attention threat assessment of perception paths corresponding to different vehicle sensors, enabling threat assessment of different vehicle sensor paths under the same automatic emergency braking decision context.

[0057] Specifically, for any vehicle sensor, each instance feature in the corresponding feature matrix of that vehicle sensor is mapped to a key and a value, and the context query feature is used as a shared query vector. Then, based on the correlation between the shared query vector and each key, an initial attention score is obtained for each instance feature. The initial attention score is used to characterize the degree of correlation between the instance features corresponding to each candidate target object and the current automatic emergency braking decision context under the current context query feature.

[0058] Furthermore, based on the path contribution coefficient corresponding to the vehicle sensor on that road, the initial attention score or attention weight corresponding to each instance feature is adjusted to obtain the adjusted attention weight. Then, the adjusted attention weight and the values ​​corresponding to each instance feature are weighted and fused to obtain the threat assessment feature corresponding to the vehicle sensor on that road.

[0059] For example, for camera paths, the context query features are used as a shared query vector to perform attention threat assessment on the features of multiple camera instances in the camera feature matrix. The attention contribution of the camera instance features is then adjusted using the camera path contribution coefficient to obtain the threat assessment features corresponding to the camera. Similarly, for LiDAR paths, the same context query features are used as a shared query vector to perform attention threat assessment on the features of multiple LiDAR instances in the LiDAR feature matrix. The attention contribution of the LiDAR instance features is then adjusted using the LiDAR path contribution coefficient to obtain the threat assessment features corresponding to the LiDAR.

[0060] In this embodiment, the attention threat assessment targets multiple instance features in the feature matrix. Each instance feature corresponds to a candidate target object, and the attention weight corresponding to each instance feature reflects the relative importance of each candidate target object to the current automatic emergency braking decision. The path contribution coefficient further adjusts the contribution of instance features in the perception path corresponding to the vehicle sensor on the threat assessment features, so that the threat assessment process can simultaneously reflect the importance of the candidate target object and the reliability differences of different perception paths under the current vehicle state and environmental state.

[0061] Optionally, before performing attention threat assessment on the feature matrices corresponding to each vehicle sensor, the instance features in the feature matrices can be embedded and encoded. Specifically, the instance features are input into an encoder, and the instance features are mapped to a preset high-dimensional feature space through a linear transformation network or a deep neural network to obtain the encoded agent features. The encoder input may also include at least one of the following: spatial position encoding of the candidate target object, 3D anchor point information, target size information, or motion state information. The spatial position encoding is used to characterize the spatial position of the candidate target object in the vehicle coordinate system or a unified 3D coordinate system, and the 3D anchor point information is used to characterize the initial candidate position of the candidate target object in the 3D space surrounding the vehicle. For camera paths, camera instance features can be fused with the corresponding 3D anchor point information, projection position information, or image scale information and then input into the encoder; for LiDAR paths, LiDAR instance features can be fused with the corresponding 3D anchor point information, point cloud spatial position encoding, or bird's-eye view position encoding and then input into the encoder. The encoded agent features are used as keys and values ​​in subsequent attention threat assessment. By introducing positional encoding or anchor point information during instance feature encoding, the agent features not only contain the semantic and motion information of candidate target objects, but also retain their positional information in the three-dimensional space surrounding the vehicle. When subsequently performing attention-based threat assessment on each instance feature using shared query vectors, both the semantic hazard and spatial proximity of candidate target objects can be considered simultaneously, which helps improve the accuracy of target threat assessment in automatic emergency braking scenarios.

[0062] Step S106: The threat assessment features of each vehicle sensor are fused to obtain a comprehensive threat assessment feature, and an automatic emergency braking prediction signal is output based on the comprehensive threat assessment feature.

[0063] The threat assessment features of each vehicle sensor reflect the degree of threat posed by candidate targets to the automatic emergency braking decision-making process under the corresponding perception path. The comprehensive threat assessment feature is obtained by fusing the threat assessment features of each vehicle sensor, and is used to comprehensively represent the collision risk status in the current vehicle driving environment.

[0064] Optionally, the threat assessment features from various vehicle sensors can be stitched together, added together, weighted summed, or mapped using a neural network to obtain comprehensive threat assessment features. For example, camera threat assessment features and lidar threat assessment features can be stitched together and mapped to comprehensive threat assessment features through a fusion network.

[0065] After obtaining the comprehensive threat assessment features, these features are input into the prediction output module to obtain the automatic emergency braking prediction signal. Optionally, the comprehensive threat assessment features are input into a multilayer perceptron to obtain a risk prediction value; the risk prediction value is normalized using a sigmoid function to obtain the automatic emergency braking trigger probability; and the automatic emergency braking prediction signal is output based on the trigger probability. The automatic emergency braking prediction signal is used to characterize whether automatic emergency braking is triggered, and can also be used to characterize the probability of triggering automatic emergency braking, the risk level, or the braking control requirement.

[0066] In the automatic emergency braking control method for vehicles provided in this embodiment, the following steps are first taken: First, the vehicle's motion state information, environmental state information, and detection data from at least two different types of vehicle sensors in the vehicle's driving environment are acquired. The vehicle's motion state information and environmental state information are then fused and encoded into context query features. Next, for each vehicle sensor, the detection data of that sensor is processed to obtain a feature matrix including instance features corresponding to multiple candidate target objects. A path contribution coefficient is determined based on the context query features and sensor type. Subsequently, the context query features are used as a shared query vector to perform attention threat assessment on the instance features corresponding to that vehicle sensor. During the attention threat assessment process, the path contribution coefficient is used to adjust the attention contribution of the instance features, resulting in threat assessment features corresponding to that vehicle sensor. Finally, the threat assessment features of each vehicle sensor are fused to obtain a comprehensive threat assessment feature, and an automatic emergency braking prediction signal is output based on the comprehensive threat assessment feature. The automatic emergency braking control method for vehicles provided in this embodiment, on the one hand, no longer relies solely on the structured detection results such as bounding boxes, categories, and confidence levels independently output by each vehicle sensor for post-fusion at the decision layer. Instead, it performs attention threat assessment at the level of candidate target object instance features corresponding to each vehicle sensor, thereby preserving and utilizing fine-grained feature information of candidate target objects and enhancing the information complementarity between different sensor perception paths. On the other hand, this method does not simply statically stitch together the features extracted from different vehicle sensors. Instead, it fuses the vehicle's motion state information and environmental state information to form contextual query features, and determines the path contribution coefficient based on the contextual query features and sensor type. This path contribution coefficient reflects the degree of contribution of each perception path to the automatic emergency braking threat assessment under the current vehicle motion state and environmental state. Furthermore, during the attention threat assessment process, the attention contribution of instance features is adjusted using the path contribution coefficient, so that when the reliability of a perception path corresponding to a certain vehicle sensor changes due to environmental state, the contribution of instance features in that perception path to the threat assessment features can be adjusted accordingly. Therefore, compared to proactive safety decision-making methods that rely solely on vehicle coding information to query overall environmental coding results, or methods that determine fusion weights only at the sensor or detection result level, this method can achieve context-driven dynamic threat assessment at the candidate target instance feature level, thereby improving the accuracy and stability of automatic emergency braking prediction signals in complex environments.

[0067] Optionally, in one embodiment, when determining the path contribution coefficient based on context query features and sensor type, the context query features and the type codes corresponding to the sensor type are fused to obtain path context features, and the path context features are then mapped using a multilayer perceptron to obtain the path contribution coefficient.

[0068] Specifically, the type code corresponding to the sensor type is used to characterize the type of vehicle sensor currently being processed, such as camera type, LiDAR type, or other vehicle sensor types. The type code can be a pre-set code or an embedding vector learned during model training. For each vehicle sensor path, the context query features are concatenated, added, weighted fused, or linearly mapped with the type code corresponding to that vehicle sensor path to obtain the path context features. The path context features include both the automatic emergency braking decision context under the combined influence of the current vehicle motion state and the environmental state, and the type information of the perception path to which the current vehicle sensor belongs. Then, the path context features are input into a multilayer perceptron, and through at least one layer of linear mapping and nonlinear activation processing, the path contribution coefficient corresponding to that vehicle sensor path is output.

[0069] The automatic emergency braking control method for vehicles provided in this embodiment fuses contextual query features with type codes corresponding to sensor types, allowing the path contribution coefficient to be jointly determined by the environmental state, the vehicle state, and the current perceived path type. Therefore, under the same environmental state, different sensor paths can yield different contribution adjustment results; and under the same sensor path, different vehicle motion states and environmental states can also cause changes in the path contribution coefficient, thereby achieving adaptive contribution adjustment based on sensor paths and improving the adaptability of subsequent attention threat assessment to complex scenarios.

[0070] Optionally, in one embodiment, when fusing and encoding the vehicle motion state information and environmental state information to obtain context query features, the vehicle motion state information is encoded to obtain vehicle state encoding features; environmental state information is extracted from the vehicle's forward-facing camera image and encoded to obtain environmental encoding features; the vehicle state encoding features and environmental encoding features are concatenated to obtain concatenated features; and the concatenated features are mapped through a fusion network to obtain context query features.

[0071] Specifically, the vehicle motion state information is first normalized and then encoded using a linear layer, multilayer perceptron, or other encoding network to obtain the vehicle state encoding features. Environmental state information can be environmental feature vectors extracted from the vehicle's forward-facing camera image. For example, the forward-facing camera image can be input into an image feature extraction network, and environmental feature vectors representing the current environmental state can be obtained through convolutional feature extraction, global pooling, or classification feature extraction. These environmental feature vectors can reflect environmental factors affecting the reliability of the perception path, such as weather, lighting, visibility, rain, fog, strong backlight, low light at night, or sensor occlusion. Then, the environmental state information, i.e., the environmental feature vector, is encoded to obtain environmental encoding features. The vehicle state encoding features and environmental encoding features are concatenated and mapped to a unified feature space through a fusion network to obtain context query features. These context query features serve as a shared query vector in subsequent attention threat assessment, used to query candidate target instance features in different sensor paths within the same automatic emergency braking decision context.

[0072] The automatic emergency braking control method for vehicles provided in this embodiment encodes the vehicle's motion state information and environmental state information separately, and then fuses them to generate contextual query features. This allows subsequent attention threat assessment to no longer rely solely on the characteristics of the target object itself, but to simultaneously consider the impact of the vehicle's current motion risk and environmental state on the reliability of the perception path. Therefore, when the vehicle is in different motion states such as high speed, acceleration, or steering, or in different environmental states such as rain, fog, low light, or strong backlight, different contextual query features can be generated. This provides a basis for dynamically focusing on candidate target objects and adjusting the contributions of different sensor paths, improving the matching degree between the automatic emergency braking prediction signal and the actual driving risk.

[0073] Optionally, in one embodiment, when using context query features as a shared query vector to perform attention threat assessment on each instance feature to obtain threat assessment features, the context query features are used as the shared query vector; for each instance feature, the instance feature is mapped to a key and a value; based on the correlation between the shared query vector and the key, an initial attention score corresponding to the instance feature is obtained; the initial attention score is adjusted according to the path contribution coefficient to obtain an adjusted attention weight; and a weighted fusion is performed based on the adjusted attention weights corresponding to each instance feature and the values ​​corresponding to each instance feature to obtain the threat assessment features.

[0074] Specifically, in this embodiment, the feature matrix corresponding to each vehicle sensor includes multiple instance features, and each instance feature corresponds to a candidate target object. After using the context query features as a shared query vector (Query), each instance feature is mapped through a key mapping network and a value mapping network to obtain the key and value corresponding to each instance feature. The correlation between the shared query vector and each key represents the degree of matching between the current automatic emergency braking decision context and each candidate target object. The higher the correlation, the more likely the candidate target object is to influence the automatic emergency braking decision in the current scenario. The correlation is a similarity, which can be obtained through dot product or scaled dot product. Optionally, the dot product or scaled dot product between the shared query vector and each key is calculated to obtain the initial attention score corresponding to each instance feature. The initial attention score characterizes the degree of matching between the current automatic emergency braking decision context and the instance features corresponding to each candidate target object in the feature space learned by the model. After obtaining the initial attention score, the path contribution coefficient is used to perform multiplicative adjustment or additive bias adjustment on the initial attention score, and then the adjusted score is normalized to obtain the adjusted attention weight. Finally, the values ​​corresponding to each instance feature are weighted and fused using the adjusted attention weights to obtain the threat assessment features corresponding to the vehicle sensors on that road.

[0075] For example, if the current vehicle speed is high and the distance ahead is short, and a certain instance feature represents "a vehicle in the same lane ahead at close range," then the key vector corresponding to this instance feature is likely to be a better match with the shared query vector, resulting in a higher initial attention score. Conversely, if an instance feature represents "a distant static target to the side and rear," its match with the current shared query vector is likely to be lower, resulting in a lower initial attention score.

[0076] The automatic emergency braking control method for vehicles provided in this embodiment determines the initial attention score of candidate target objects by sharing the correlation between the query vector and the keys corresponding to each instance feature. Then, the initial attention score is adjusted by the path contribution coefficient, so that the final attention weight is affected by the correlation between the target instance features and the current decision context, as well as the contribution of the perception path in the current environment. Therefore, the threat assessment features can not only reflect which candidate target objects are more important, but also reflect whether the sensor path should be trusted more in the current scenario, thereby avoiding the problem that static stitching or fixed fusion methods cannot dynamically adjust the sensor contribution according to the environmental state.

[0077] Optionally, in one embodiment, the attention threat assessment employs a multi-head cross-attention mechanism, whereby for each attention head, the shared query vector, the key and value corresponding to the instance features are linearly mapped to obtain the query sub-vector, key vector and value sub-vector corresponding to that attention head.

[0078] For any instance feature under each attention head, when obtaining the adjusted attention weight, the initial attention score within the head corresponding to the instance feature is obtained based on the correlation between the query sub-vector and the key sub-vector. The initial attention score within the head is adjusted based on the path contribution coefficient to obtain the adjusted attention weight within the head.

[0079] When obtaining threat assessment features, for each attention head, the attention output corresponding to the attention head is obtained by weighted fusion based on the intra-head adjusted attention weights and the value sub-vectors corresponding to each instance feature. The attention outputs corresponding to each attention head are then concatenated and output to obtain the threat assessment features.

[0080] Specifically, the multi-head cross-attention mechanism includes multiple parallel attention heads. Each attention head calculates the correlation between a shared query vector and instance features within different feature subspaces. For each attention head, a linear mapping is performed on the shared query vector, the keys corresponding to each instance feature, and the values, resulting in query sub-vectors, key vectors, and value sub-vectors. Then, within each attention head, an initial intra-head attention score is calculated based on the correlation between the query sub-vector and each key vector. The path contribution coefficient can be applied to the initial intra-head attention score of each attention head, or to the intra-head attention weights before and after normalization, to adjust the attention contribution of the perception path across different attention heads. After each attention head obtains its attention output based on the adjusted intra-head attention weights and value sub-vectors, the multiple attention outputs are concatenated and mapped to the threat assessment features corresponding to the vehicle sensors on that path through an output projection layer.

[0081] The automatic emergency braking control method for vehicles provided in this embodiment employs a multi-head cross-attention mechanism. This mechanism allows instance features from the same vehicle sensor to be queried by contextual query features in multiple feature subspaces, thereby characterizing the relationship between candidate target objects and the automatic emergency braking decision context from different perspectives. Simultaneously, the path contribution coefficient is incorporated into the attention score adjustment process of each attention head, ensuring that the target attention results obtained by different attention heads are constrained by the current perceived path contribution level. The method provided in this embodiment not only improves the expressive power of threat assessment features for the multidimensional features of candidate target objects but also maintains the consistency of path contribution adjustment across different sensors, further enhancing the accuracy and stability of threat assessment in complex driving scenarios.

[0082] Optionally, in one embodiment, the neural network model includes a perception processing module for obtaining a feature matrix, a fusion encoding module for obtaining context query features, an attention threat assessment module for obtaining threat assessment features, a threat fusion module for obtaining comprehensive threat assessment features, and a prediction output module for outputting an automatic emergency braking prediction signal. Specifically, training the neural network model includes the following steps: Acquire training samples, which include sample detection data, sample vehicle motion state information, sample environmental state information, and automatic emergency braking labels; input the sample detection data, sample vehicle motion state information, and sample environmental state information into a neural network model to obtain sample automatic emergency braking prediction signals; calculate loss values ​​based on sample automatic emergency braking prediction signals and automatic emergency braking labels, wherein the loss values ​​are calculated using the Focal Loss function; update the network parameters of at least one of the following modules: perception processing module, fusion encoding module, attention threat assessment module, threat fusion module, and prediction output module, based on the loss values.

[0083] When the neural network model is trained end-to-end, the network parameters of the perception processing module, fusion encoding module, attention threat assessment module, threat fusion module, and prediction output module are jointly updated based on the loss value. When the neural network model is trained in stages, the perception processing module is trained first, followed by the fusion encoding module, attention threat assessment module, threat fusion module, and prediction output module.

[0084] Specifically, the perception processing module is used to extract features from the detection data of each vehicle sensor to obtain the feature matrix corresponding to each vehicle sensor; the fusion encoding module is used to fuse the sample vehicle motion state information and the sample environment state information to obtain the context query features; the attention threat assessment module is used to perform attention threat assessment on the instance features in each feature matrix based on the context query features to obtain the threat assessment features corresponding to each vehicle sensor; the threat fusion module is used to fuse the threat assessment features of each path to obtain the comprehensive threat assessment features; and the prediction output module is used to output the sample automatic emergency braking prediction signal based on the comprehensive threat assessment features.

[0085] The automatic emergency braking (AEP) labels in the training samples can indicate whether AEP needs to be triggered in the corresponding sample scenario, or they can represent the probability of triggering AEP, risk level, or braking demand label. Since the number of samples requiring AEP in real driving data is usually less than the number not requiring it, the Focal Loss function is used to calculate the loss value. This reduces the dominant influence of easily classified samples on the training process, allowing the model to focus more on difficult-to-classify samples and minority class triggering samples. In end-to-end training, the parameters of each module are jointly updated through backpropagation based on the loss value. In phased training, the perception processing module is trained first to achieve stable target instance feature extraction capabilities, followed by training of subsequent fusion encoding, attention threat assessment, threat fusion, and prediction output modules.

[0086] The automatic emergency braking control method for vehicles provided in this embodiment incorporates perception processing, fusion encoding, attention threat assessment, threat fusion, and prediction output into the same neural network model training process. This allows the model to optimize the parameters of each module around the final automatic emergency braking prediction goal. End-to-end training enables the front-end perception feature extraction process and the back-end braking risk prediction process to be optimized collaboratively, avoiding inconsistencies in goals caused by independent optimization of each module. Staged training allows for optimization of decision-related modules only after the perception module has stabilized, reducing training difficulty. Furthermore, the use of the Focal Loss function to address the imbalance between automatic emergency braking triggered and non-triggered samples improves the model's learning ability for high-risk, minority-class scenarios, thereby enhancing the reliability of the automatic emergency braking prediction signal in real-world complex driving environments.

[0087] Optionally, in one embodiment, the attention threat assessment module includes at least one layer of cross-attention network; when the cross-attention network is multi-layered, the intermediate threat features output by the upper-layer cross-attention network are used as the input to the lower-layer cross-attention network. Residual connections and layer normalization processing are set between adjacent cross-attention networks.

[0088] Specifically, in each layer of the cross-attention network, the attention output of that layer is calculated using the contextual query features or the output of the previous layer as the query vector, and the instance feature encoding results corresponding to the vehicle sensor as the key and value. Then, the attention output of that layer is summed with the input features of that layer using residuals, and the summed features are then subjected to layer normalization to obtain the output features of that layer. Optionally, the number of layers L in the cross-attention network can be 1, 2, or more; when L is 1, the computational cost of the model can be reduced; when L is greater than 1, the model's ability to express the threat relationship of candidate targets can be enhanced.

[0089] The automatic emergency braking control method for vehicles provided in this embodiment utilizes a stacked multi-layer cross-attention network. This allows the model to repeatedly leverage contextual query features to perform threat analysis on instance features of candidate target objects, thereby progressively strengthening target features relevant to the current automatic emergency braking decision. Furthermore, residual connections preserve input information from the previous layer, reducing the risk of information decay in deep networks. Layer normalization improves training stability and convergence speed, thus enhancing the expressive power of the attention threat assessment module while maintaining model trainability and improving the stability of automatic emergency braking prediction signals in complex traffic scenarios.

[0090] Optionally, in one embodiment, when outputting the automatic emergency braking prediction signal based on the comprehensive threat assessment features, the comprehensive threat assessment features are input into a multilayer perceptron to obtain a risk prediction value; the risk prediction value is normalized by a Sigmoid function (i.e., a logic function) to obtain the automatic emergency braking trigger probability; and the automatic emergency braking prediction signal is output based on the automatic emergency braking trigger probability.

[0091] Specifically, the comprehensive threat assessment feature integrates threat assessment features from various vehicle sensors, representing the overall risk state related to automatic emergency braking decisions in the current vehicle driving environment. After inputting the comprehensive threat assessment feature into a multilayer perceptron, the multilayer perceptron outputs a risk prediction value through one or more fully connected mappings and nonlinear activation processing. The risk prediction value can be an unnormalized risk score. Then, the risk prediction value is mapped to a preset probability range using a sigmoid function to obtain the automatic emergency braking trigger probability. The automatic emergency braking prediction signal can be generated based on the automatic emergency braking trigger probability. For example, when the automatic emergency braking trigger probability is greater than or equal to a preset trigger threshold, a prediction signal indicating that automatic emergency braking will be triggered is output; when the automatic emergency braking trigger probability is less than the preset trigger threshold, a prediction signal indicating that automatic emergency braking will not be triggered is output. Optionally, the automatic emergency braking prediction signal can also include the automatic emergency braking trigger probability itself, allowing the vehicle control system to combine it with other control strategies for braking decisions.

[0092] The automatic emergency braking control method for vehicles provided in this embodiment uses a multi-layer perceptron to perform nonlinear mapping on comprehensive threat assessment features, converting threat information fused from multiple sensor paths into risk prediction values. Then, the automatic emergency braking trigger probability is obtained through a sigmoid function, giving the output results a clear probabilistic meaning and comparability. Finally, the vehicle control system can perform threshold judgments, risk classifications, or joint judgments with other control conditions based on the trigger probability, thereby improving the availability and decision stability of the automatic emergency braking prediction signal.

[0093] Optionally, in one embodiment, the neural network model employs a lightweight encoder, a standard attention module, and a small multilayer perceptron to meet the real-time requirements of the vehicle-mounted automatic emergency braking system. Specifically, the perception processing module outputs instance features corresponding to a preset number of candidate target objects, the fusion encoding module encodes the vehicle's motion state information and environmental state information into contextual query features of a preset dimension, the attention threat assessment module employs a preset number of attention heads and a cross-attention network with a preset number of layers, and the prediction output module uses a small multilayer perceptron to output an automatic emergency braking prediction signal.

[0094] In one specific implementation, the computational cost of the neural network model can be controlled by limiting the number of candidate target objects, feature dimensions, number of attention heads, number of cross-attention layers, and number of multilayer perceptron layers. For example, the number of candidate target objects can be set to a fixed number, the instance features corresponding to each candidate target object can be mapped to a preset dimension, and one or two layers of cross-attention networks can be used for threat assessment. For candidate target objects with low confidence or far from the vehicle, they can be screened before entering the attention threat assessment to reduce subsequent computational cost. By forming a concise and efficient end-to-end inference structure through a lightweight encoder, a standard attention module, and a small multilayer perceptron, the model can reduce computational overhead while retaining the instance features of candidate target objects and dynamic attention threat assessment capabilities, making it suitable for deployment on in-vehicle computing platforms and meeting the real-time response requirements of automatic emergency braking systems.

[0095] Example 2 Embodiment 2 of the present invention provides an automatic emergency braking control method for a vehicle, wherein the two vehicle sensors are a camera and a lidar. Figure 2 This is a block diagram of the automatic emergency braking control system for a vehicle provided in Embodiment 2 of the present invention, as shown below. Figure 2 As shown, the automatic emergency braking control method for vehicles provided by this invention, based on this system, specifically includes the following steps: S1. Camera perception module extracts agent features. The camera perception module is responsible for extracting and fusing features from the camera to generate refined 3D object detection results. This module receives multi-view camera images as input, extracts multi-scale image features through convolutional neural networks and feature pyramid networks, generates initial 3D anchor points and instance features, and finally outputs agent features. The agent's feature dimension is N× , where N represents the number of targets detected (e.g., 300). The feature dimension (e.g., 128) represents each target, and these features include semantic information, positional information, or motion information of the candidate target object.

[0096] S2. Camera Feature Embedding and Fusion The agent features, vehicle motion state information, and environmental visual features output by the camera perception module are embedded, encoded, and fused. The specific process is as follows: 1. Input Features (1) Camera agent feature matrix : From the camera perception module, with a dimension of N× , where N is the number of detected targets. This matrix contains semantic, positional, and motion information for all potential collision targets.

[0097] (2) Vehicle state vector The dimension is 1×5, which includes the lateral and longitudinal velocity components of the vehicle; the lateral and longitudinal acceleration components; and the heading angle θ.

[0098] (3) Environment state vector Visual features extracted from the forward-facing camera image are processed by a feature extraction network and global pooling to obtain a feature vector of dimension 1×512, representing the semantic information of the current driving environment.

[0099] 2. Feature Encoding and Representation Learning To unify heterogeneous inputs into a common feature space, an encoder network is used for feature transformation: (1) Agent feature encoding: The agent feature matrix is ​​encoded by an encoder. Mapping to a high-dimensional feature space: = ( ),in ∈ N is the number of detected targets. Encoder It can be a linear transformation or a deep neural network. Depending on the instance feature fusion mode, the input may include position encoding and anchor information, etc.

[0100] (2) Vehicle status coding: via encoder The vehicle state vector Mapping to a high-dimensional feature space: ' = ( ),in '∈ Encoder Learn the semantic representation of the vehicle's motion state and capture the impact of speed, acceleration, and attitude on collision risk.

[0101] (3) Environment state coding: via encoder Environmental state vector Mapping to a high-dimensional feature space: ' = ( ),in '∈ Encoder Extracting environmental semantic information, such as road type, weather conditions, and lighting conditions, affects the assessment of collision risk.

[0102] (4) Context-aware feature fusion To comprehensively consider both vehicle and environmental states, a feature fusion strategy is employed. Vehicle and environmental features are concatenated and then processed through a fusion network. The concatenated features are mapped to a high-dimensional feature space to generate a context feature vector. = ( ),in ∈ This context feature vector serves as a query vector for subsequent attention calculations, integrating information about the vehicle's motion state and the current driving environment.

[0103] (5) Generation of camera path contribution coefficient To dynamically adjust the contribution of the camera path to threat assessment based on the current vehicle motion state and the current environmental state, and according to the context feature vector... The sensor type code corresponding to the camera determines the camera path contribution coefficient. The camera path contribution coefficient Attention contributions used to adjust camera agent features during subsequent camera agent threat assessment.

[0104] S3. Threat Assessment Based on Cross-Attention Driven by Camera Agents A multi-head cross attention mechanism is employed to enable contextual features to dynamically focus on relevant agents, achieving intelligent threat assessment. For the camera path, the contextual feature vector is... As a query vector, camera agent features Cross-attention is performed using the keys and values, and the camera path contribution coefficient is introduced during the attention calculation process. This is to adjust the attention contribution of camera agent features. The specific steps are as follows: 1. Attention Calculation =

[0105] in, ∈ , where represents the weighted fused feature vector, and H represents the number of heads in the multi-head attention process. Specifically, after calculating the correlation between the query vector and the key to obtain the initial attention score, the camera path contribution coefficient is used to adjust the initial attention score or the attention weights obtained by normalizing the initial attention score. This ensures that the reliability and contribution of the camera path in the current vehicle motion state and the current environment state are reflected in the attention calculation process. Through the above attention process, the model can focus on important candidate target features within the camera path.

[0106] 2. Multi-head integration For H attention heads (e.g., H=8), the outputs of each head are concatenated and then projected: =

[0107] Where: Concat(·) means concatenating H vectors into one vector, and the dimension of the concatenated vector is 1×128. To output the projection matrix (weight matrix), the concatenated features are projected back into the high-dimensional feature space.

[0108] The multi-head mechanism enables the model to understand the relationships between features from multiple perspectives (different representation subspaces). Each head focuses on different feature dimensions, and finally, the information from all heads is fused to improve the model's expressive power.

[0109] 3. Deep Attention Networks To enhance the model's expressive power, L layers of cross-attention are stacked (L≥1, typically L=1 or 2), with residual connections and layer normalization used between each layer:

[0110] in LayerNorm(·) represents the layer normalization operation, which is used to improve training stability and accelerate model convergence.

[0111] S4. The radar perception module extracts features from the intelligent agent. The radar perception module is responsible for extracting and fusing features from radar point clouds to generate refined 3D target detection results. This module receives radar point cloud data as input, extracts features through convolutional neural networks and feature pyramid networks, generates initial 3D anchor points and instance features, and finally outputs agent features. The agent's feature dimension is N× , where N represents the number of targets detected (e.g., 300). This represents the feature dimension of each target (e.g., 128), which includes semantic information, location information, motion information, etc. of the target.

[0112] S5. Radar Feature Embedding and Fusion The agent features output by the radar sensing module After the same processing as in S2, we obtain the feature vectors mapped to the high-dimensional feature space: = ( ),in ∈ N is the number of detected targets. Encoder It can be a linear transformation or a deep neural network. Depending on the instance feature fusion mode, the input may include position encoding, anchor information, etc.

[0113] Furthermore, in order to dynamically adjust the contribution of the radar path to threat assessment based on the current vehicle motion state and the current environmental state, the context feature vector is used... The radar path contribution coefficient is determined by the sensor type code corresponding to the radar. Specifically, the context feature vector The radar path context features are obtained by concatenating or fusing the sensor type codes corresponding to the radar, and then the radar path contribution coefficient is obtained by performing feature mapping on the radar path context features through a multilayer perceptron. The radar path contribution coefficient Attention contribution used to adjust radar agent features during subsequent radar agent threat assessment.

[0114] S6. Threat Assessment of Radar Agents Based on Cross-Attention Based on the same principle as S3, for radar paths, a radar agent threat assessment driven by cross-attention is calculated, and the specific process is as follows: 1. Attention Calculation =

[0115] in, ∈ , where represents the weighted fused feature vector, and H represents the number of heads in the multi-head attention. After calculating the correlation between the query vector and the key to obtain the initial attention score, the radar path contribution coefficient is used. The initial attention score or the attention weight obtained by normalizing the initial attention score is adjusted so that the reliability and contribution of the radar path in the current vehicle motion state and the current environment state can be reflected in the attention calculation process.

[0116] 2. Multi-head integration For H attention heads (e.g., H=8), the outputs of each head are concatenated and then projected: =

[0117] Concat(·) means concatenating H vectors into a single vector, with the concatenated vector having a dimension of 1×128. To output the projection matrix (weight matrix), the concatenated features are projected back into the high-dimensional feature space.

[0118] 3. Deep Attention Networks

[0119] in LayerNorm(·) represents the layer normalization operation, which is used to improve training stability and accelerate model convergence.

[0120] S7. Decision Output and Risk Quantification The outputs of the camera attention module and the radar attention module are fused and then input into a multilayer perceptron (MLP) for final decision-making. The output of the camera attention module already includes the adjustment result of the camera path contribution coefficient on the feature attention contribution of the camera agent, and the output of the radar attention module already includes the adjustment result of the radar path contribution coefficient on the feature attention contribution of the radar agent. Therefore, the fused output... It can simultaneously reflect the differentiated contributions of camera path and radar path under the current vehicle motion state and the current environmental state:

[0121]

[0122] Where: Sigmoid(·) is the Sigmoid activation function, Sigmoid(x) = 1 / (1+exp(-x)), which maps any real number to the interval [0, 1]; The value range is [0, 1], which represents the probability of AEB being triggered. The closer it is to 1, the higher the collision risk, and the more likely AEB needs to be triggered.

[0123] The entire system adopts an end-to-end training approach, and the parameters of the perception module and the decision module are jointly optimized through the backpropagation algorithm.

[0124] 1. Loss Function Design For AEB trigger signal prediction, Focal Loss is used to address the positive-negative sample imbalance problem. Focal Loss is defined as:

[0125] in: For predicting probabilities; The balancing parameter is used to balance the weights of positive and negative samples (where... (Or set adaptively according to the sample ratio); γ is the focusing parameter, used to control the weight difference between easy and difficult samples (usually γ=2).

[0126] 2. Training Strategies Phased training or end-to-end training can be employed. In phased training, the perception module is trained first, and then its parameters are fixed before training the decision module. In end-to-end training, the entire system is jointly optimized, but the training speed of each module can be balanced by setting different learning rates.

[0127] This embodiment uses environmental semantic information (such as weather, lighting, and visibility) as the core basis for adjusting the weights of multi-sensor fusion. By extracting environmental features from the forward-facing camera image and encoding them as context vectors to participate in cross-attention calculation, the system can automatically reduce the attention weight (i.e., confidence) of the output in scenarios where radar performance is degraded, such as rainy or foggy weather, and instead increase the attention to the camera path; conversely, in low-light environments, it increases the importance of the radar path. This dynamic and intelligent fusion mechanism significantly improves the robustness and reliability of the system in extreme environments.

[0128] From raw sensor data to the final AEB trigger signal, the entire system adopts an end-to-end training approach, avoiding the information abstraction and loss caused by traditional modular designs. The high-dimensional agent features output by the perception module are fully preserved and directly fed into the decision network, ensuring that the decision layer can access rich underlying details.

[0129] A multi-head cross-attention mechanism is introduced to enable contextual features (including vehicle state and environment state) to dynamically focus on the most threatening target. Furthermore, camera path contribution coefficients and radar path contribution coefficients are generated based on contextual features and sensor type, respectively. During the cross-attention calculation of camera and radar paths, the attention contribution of each agent's features is adjusted using the corresponding path contribution coefficients. This mechanism inherently supports cross-modal attention allocation. The agent features of the camera and radar each enter independent attention branches, their target attention is controlled by a shared context query vector, and their path contribution is adjusted by the corresponding path contribution coefficients, thereby achieving differentiated trust allocation based on vehicle state and environment state. The system's fusion strategy is learned autonomously by a deep neural network, eliminating the need for manual setting of fusion rules or threshold switching logic under different weather conditions. The system can be trained on a large amount of diverse scenario data, automatically learning "when to trust whom" and demonstrating excellent generalization ability.

[0130] The system architecture is simple and efficient, employing a lightweight encoder, a standard attention module, and a small multilayer perceptron, enabling real-time inference on an onboard computing platform and meeting the millisecond-level response requirements of the AEB system.

[0131] This system implements an automatic emergency braking (AEB) control method for vehicles. The perception module outputs agent feature matrices for both camera and radar paths (instead of traditional target detection results), preserving the original feature information. The decision-making part processes the two types of features separately using a dual-channel cross-attention structure, and introduces vehicle state information and environmental state information to generate a context query vector. This dynamically adjusts the attention weights of each sensor path, enabling adaptive adjustment of camera and radar confidence levels based on vehicle and environmental states. This architecture unifies perception and decision-making within an end-to-end framework, achieving a direct mapping from multimodal input to AEB trigger probability.

[0132] Example 3 Corresponding to Embodiment 1 above, Embodiment 3 of the present invention provides an automatic emergency braking control device for a vehicle. The technical features and corresponding technical effects can be referred to Embodiment 1 above, and will not be repeated in this embodiment. Figure 3 This is a block diagram of the automatic emergency braking control device for a vehicle provided in Embodiment 3 of the present invention, as shown below. Figure 3 As shown, the device includes: an acquisition module 201, a fusion encoding module 202, a perception processing module 203, an attention threat assessment module 204, a threat fusion module 205, and a prediction output module 206.

[0133] The acquisition module 201 is used to acquire vehicle motion state information, environmental state information, and detection data from at least two different types of vehicle sensors in the vehicle driving environment. The environmental state information is used to characterize the environmental state affecting the reliability of the perception path corresponding to at least one vehicle sensor. The fusion encoding module 202 is used to fuse and encode the vehicle motion state information and the environmental state information to obtain context query features. The context query features are used to characterize the automatic emergency braking decision context under the combined effect of the current vehicle motion state and the current environmental state. The perception processing module 203 is used to perform perception processing on the detection data from each vehicle sensor to obtain the corresponding data from that vehicle sensor. The feature matrix includes instance features corresponding to multiple candidate target objects; the attention threat assessment module 204 is used to determine the path contribution coefficient based on the context query features and sensor type, and uses the context query features as a shared query vector to perform attention threat assessment on each instance feature to obtain threat assessment features, wherein the path contribution coefficient is used to adjust the attention contribution of the instance features in the attention threat assessment; the threat fusion module 205 is used to fuse the threat assessment features of each of the vehicle sensors to obtain a comprehensive threat assessment feature; and the prediction output module 206 is used to output an automatic emergency braking prediction signal based on the comprehensive threat assessment feature.

[0134] Optionally, in one embodiment, the step of determining the path contribution coefficient based on the context query features and sensor type includes: fusing the context query features and the type encoding corresponding to the sensor type to obtain path context features; and performing feature mapping on the path context features through a multilayer perceptron to obtain the path contribution coefficient.

[0135] Optionally, in one embodiment, the fusion encoding module includes: a first encoding unit for encoding the vehicle motion state information to obtain vehicle state encoding features; a second encoding unit for extracting environmental state information from the vehicle's forward-facing camera image and encoding the environmental state information to obtain environmental encoding features; a stitching unit for stitching the vehicle state encoding features and the environmental encoding features to obtain stitched features; and a mapping unit for mapping the stitched features through a fusion network to obtain the context query features.

[0136] Optionally, in one embodiment, the step of using the context query features as a shared query vector to perform attention threat assessment on each instance feature to obtain a threat assessment feature includes: using the context query features as a shared query vector; for each instance feature: mapping the instance feature to a key and a value; obtaining an initial attention score corresponding to the instance feature based on the correlation between the shared query vector and the key; adjusting the initial attention score based on the path contribution coefficient to obtain an adjusted attention weight; and performing weighted fusion based on the adjusted attention weights corresponding to each instance feature and the values ​​corresponding to each instance feature to obtain the threat assessment feature.

[0137] Optionally, in one embodiment, the attention threat assessment employs a multi-head cross-attention mechanism, and the device further includes: a mapping module, configured to perform linear mapping on the shared query vector, the key and value corresponding to the instance feature for each attention head, respectively, to obtain the query sub-vector, key vector and value sub-vector corresponding to the attention head; the step of obtaining the adjusted attention weight for any instance feature under each attention head includes: obtaining the initial intra-head attention score corresponding to the instance feature based on the correlation between the query sub-vector and the key vector, adjusting the initial intra-head attention score based on the path contribution coefficient to obtain the adjusted intra-head attention weight; the step of obtaining the threat assessment feature includes: for each attention head, performing weighted fusion based on the adjusted intra-head attention weight and the value sub-vector corresponding to each instance feature, respectively, to obtain the attention output corresponding to the attention head, concatenating and outputting the attention outputs corresponding to each attention head to obtain the threat assessment feature.

[0138] Optionally, in one embodiment, the neural network model includes the perception processing module, the fusion encoding module, the attention threat assessment module, the threat fusion module, and the prediction output module, wherein the step of training the neural network model includes: Acquire training samples, wherein the training samples include sample detection data, sample vehicle motion state information, sample environmental state information, and automatic emergency braking tags; The sample detection data, the sample vehicle motion state information, and the sample environmental state information are input into the neural network model to obtain the sample automatic emergency braking prediction signal; The loss value is calculated based on the sample automatic emergency braking prediction signal and the automatic emergency braking label, wherein the loss value is calculated using the Focal Loss function; The network parameters of at least one of the following modules—the perception processing module, the fusion encoding module, the attention threat assessment module, the threat fusion module, and the prediction output module—are updated based on the loss value. When the neural network model is trained end-to-end, the network parameters of the perception processing module, the fusion encoding module, the attention threat assessment module, the threat fusion module, and the prediction output module are jointly updated based on the loss value. When the neural network model is trained in a phased manner, the perception processing module is trained first, and then the fusion encoding module, the attention threat assessment module, the threat fusion module, and the prediction output module are trained.

[0139] Optionally, in one embodiment, the step of outputting an automatic emergency braking prediction signal based on the comprehensive threat assessment features includes: inputting the comprehensive threat assessment features into a multilayer perceptron to obtain a risk prediction value; normalizing the risk prediction value using a sigmoid function to obtain an automatic emergency braking trigger probability; and outputting the automatic emergency braking prediction signal based on the automatic emergency braking trigger probability.

[0140] Example 4 Embodiment 4 of the present invention provides a vehicle, which includes at least two different types of vehicle sensors and a processor. The processor is an automatic emergency braking control method for any of the above-mentioned vehicles. For detailed technical features and corresponding technical effects, please refer to the above embodiments, and will not be repeated here.

[0141] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0142] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0143] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method.

[0144] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. An automatic emergency braking control method for a vehicle, characterized in that, include: The vehicle acquires vehicle motion state information, environmental state information, and detection data from at least two different types of vehicle sensors in the vehicle driving environment. The environmental state information is used to characterize the environmental state that affects the reliability of the perception path corresponding to at least one vehicle sensor. The vehicle motion state information and the environmental state information are fused and encoded to obtain context query features, wherein the context query features are used to characterize the automatic emergency braking decision context under the combined effect of the current vehicle motion state and the current environmental state. For each vehicle sensor: the detection data of the vehicle sensor on that route is processed to obtain a feature matrix. The path contribution coefficient is determined based on the context query features and the sensor type. The feature matrix includes instance features corresponding to multiple candidate target objects. The context query features are used as a shared query vector. Attention threat assessment is performed on each instance feature to obtain threat assessment features. In the attention threat assessment, the path contribution coefficient is used to adjust the attention contribution of the instance features. The threat assessment features of the various vehicle sensors are fused to obtain a comprehensive threat assessment feature; and An automatic emergency braking prediction signal is output based on the comprehensive threat assessment characteristics.

2. The automatic emergency braking control method for a vehicle according to claim 1, characterized in that, The at least two different types of vehicle sensors include a camera and a lidar, and the detection data includes camera image data collected by the camera and lidar point cloud data collected by the lidar. For the camera, the path contribution coefficient is used to characterize the degree of contribution of the perception path corresponding to the camera to the automatic emergency braking threat assessment under the current vehicle motion state and the current environmental state. For the lidar, the path contribution coefficient is used to characterize the degree of contribution of the perception path corresponding to the lidar to the automatic emergency braking threat assessment under the current vehicle motion state and the current environmental state. The environmental status information includes information on at least one of the following: weather status, lighting status, visibility status, precipitation status, haze status, road reflection status, and sensor obstruction status. The vehicle motion state information includes at least one of the vehicle's lateral velocity, longitudinal velocity, lateral acceleration, longitudinal acceleration, and heading angle.

3. The automatic emergency braking control method for a vehicle according to claim 1, characterized in that, The steps for determining the path contribution coefficient based on the context query features and sensor type include: The context query features and the type codes corresponding to the sensor types are fused to obtain path context features; The path contribution coefficient is obtained by performing feature mapping on the path context features using a multilayer perceptron.

4. The automatic emergency braking control method for a vehicle according to claim 1, characterized in that, The steps of fusing and encoding the vehicle motion state information and the environmental state information to obtain context query features include: The vehicle motion state information is encoded to obtain the vehicle state encoding features; Environmental state information is extracted from images taken from the vehicle's forward-facing camera, and this environmental state information is encoded to obtain environmental coding features. The vehicle state coding features and the environment coding features are concatenated to obtain the concatenated features; and The context query features are obtained by mapping the concatenated features through a fusion network.

5. The automatic emergency braking control method for a vehicle according to claim 1, characterized in that, The step of using the context query features as a shared query vector and performing attention threat assessment on each instance feature to obtain threat assessment features includes: Use the context query features as a shared query vector; For each instance feature: the instance feature is mapped to a key and a value; based on the correlation between the shared query vector and the key, an initial attention score corresponding to the instance feature is obtained; the initial attention score is adjusted based on the path contribution coefficient to obtain the adjusted attention weight. The threat assessment features are obtained by weighting and fusing the adjusted attention weights and values ​​corresponding to each instance feature.

6. The automatic emergency braking control method for a vehicle according to claim 5, characterized in that, The attention threat assessment adopts a multi-head cross-attention mechanism. The method further includes: for each attention head, performing linear mapping on the shared query vector, the key and value corresponding to the instance feature, respectively, to obtain the query sub-vector, key vector and value sub-vector corresponding to the attention head; The step of obtaining the adjusted attention weight for any instance feature under each attention head includes: obtaining the initial intra-head attention score corresponding to the instance feature based on the correlation between the query sub-vector and the key sub-vector; adjusting the initial intra-head attention score based on the path contribution coefficient to obtain the adjusted intra-head attention weight. The steps for obtaining the threat assessment features include: for each attention head, performing weighted fusion based on the intra-head adjusted attention weights and the value sub-vectors corresponding to each instance feature to obtain the attention output corresponding to the attention head; concatenating and outputting the attention outputs corresponding to each attention head to obtain the threat assessment features.

7. The automatic emergency braking control method for a vehicle according to claim 1, characterized in that, The neural network model includes a perceptual processing module for obtaining the feature matrix, a fusion encoding module for obtaining the context query features, an attention-based threat assessment module for obtaining the threat assessment features, a threat fusion module for obtaining the comprehensive threat assessment features, and a prediction output module for outputting an automatic emergency braking prediction signal. The steps for training the neural network model include: Acquire training samples, wherein the training samples include sample detection data, sample vehicle motion state information, sample environmental state information, and automatic emergency braking tags; The sample detection data, the sample vehicle motion state information, and the sample environmental state information are input into the neural network model to obtain the sample automatic emergency braking prediction signal; The loss value is calculated based on the sample automatic emergency braking prediction signal and the automatic emergency braking label, wherein the loss value is calculated using the Focal Loss function; The network parameters of at least one of the following modules—the perception processing module, the fusion encoding module, the attention threat assessment module, the threat fusion module, and the prediction output module—are updated based on the loss value. When the neural network model is trained end-to-end, the network parameters of the perception processing module, the fusion encoding module, the attention threat assessment module, the threat fusion module, and the prediction output module are jointly updated based on the loss value. When the neural network model is trained in a phased manner, the perception processing module is trained first, and then the fusion encoding module, the attention threat assessment module, the threat fusion module, and the prediction output module are trained.

8. The automatic emergency braking control method for a vehicle according to claim 1, characterized in that, The steps for outputting an automatic emergency braking prediction signal based on the comprehensive threat assessment characteristics include: The comprehensive threat assessment features are input into a multilayer perceptron to obtain risk prediction values; The risk prediction value is normalized using the Sigmoid function to obtain the probability of automatic emergency braking triggering; and The automatic emergency braking prediction signal is output based on the automatic emergency braking trigger probability.

9. An automatic emergency braking control device for a vehicle, characterized in that, include: The acquisition module is used to acquire vehicle motion state information, environmental state information, and detection data from at least two different types of vehicle sensors in the vehicle driving environment. The environmental state information is used to characterize the environmental state that affects the reliability of the perception path corresponding to at least one vehicle sensor. The fusion encoding module is used to fuse and encode the vehicle motion state information and the environmental state information to obtain context query features, wherein the context query features are used to characterize the automatic emergency braking decision context under the combined effect of the current vehicle motion state and the current environmental state. The perception processing module is used to process the detection data of each vehicle sensor to obtain the feature matrix corresponding to the vehicle sensor. The feature matrix includes instance features corresponding to multiple candidate target objects. The attention threat assessment module is used to determine the path contribution coefficient based on the context query features and sensor type, use the context query features as a shared query vector, and perform attention threat assessment on each instance feature to obtain threat assessment features. In the attention threat assessment, the path contribution coefficient is used to adjust the attention contribution of the instance feature. The threat fusion module is used to fuse the threat assessment features from the various vehicle sensors to obtain a comprehensive threat assessment feature; and The prediction output module is used to output an automatic emergency braking prediction signal based on the comprehensive threat assessment characteristics.

10. A vehicle, characterized in that, include: At least two different types of vehicle sensors and processors, wherein the processors execute the automatic emergency braking control method for the vehicle as described in any one of claims 1 to 8.

Images