AEB triggering method for reducing false triggering in a curve scenario

Abstract

The application relates to an AEB triggering method for reducing false triggering in a curved road scene, which comprises the following steps: constructing a lane line recognition and lane geometry modeling module, forming stable lane boundaries, a center line and a Frenet coordinate system, and completing target vehicle lane attribution determination and curved road distance compensation; identifying a cross-line target in a loss-of-control stage after a cross-line event is triggered, and generating a trajectory bundle and a three-dimensional dynamic occupancy envelope under stage label constraints to obtain a target future influence area; introducing a tree-enhanced Bayesian network into a risk inference layer to depict the dependency relationship between multiple dangerous variables, and outputting a collision point grid posterior distribution and a high-risk duration window; combining forward collision avoidance requirements and backward rear-end collision constraints to construct a braking feasible region in a braking decision layer, and introducing an impact rate constraint and a damage cost to realize AEB graded triggering and braking instruction optimization, so that the loss-of-control intrusion risk can be inhibited in advance and the false triggering probability can be reduced.

Description

Technical Field

[0001] This invention belongs to the field of traffic control systems, and specifically relates to an AEB triggering method for reducing false triggering in curve scenarios. Background Technology

[0002] In recent years, with the rapid development of intelligent driving technology, the application rate of active safety protection, especially automatic emergency braking (AEB) systems, for heavy vehicles (heavy commercial trucks or heavy-duty trucks) has been continuously increasing. However, in scenarios such as multi-lane curved roads and highways with dense traffic flow and frequent vehicle interactions, AEB still generally faces the engineering challenge of both false triggering and missed triggering. Specifically, heavy commercial trucks have typical characteristics such as heavy load, long braking distance, and poor lateral maneuverability, which means that AEB triggering strategies must not only deal with the longitudinal collision risk of targets following in the same lane, but also face the threat of intrusion into the lane by targets in adjacent lanes when crossing, straddling, inserting, or experiencing abnormal instability. Traditional AEB triggering methods based on fixed time-of-collision (TTC) thresholds or simple lane association may, on the one hand, fail to effectively distinguish between normal targets in the same lane and those in adjacent lanes, resulting in excessive braking, disrupting normal road traffic flow, and potentially causing the vehicle to be rear-ended; on the other hand, they may fail to promptly identify the instability state of vehicles in adjacent lanes, leading to insufficient reaction when other vehicles lose control and intrude into the lane, resulting in serious collision accidents.

[0003] Although some existing technologies have attempted to improve AEB performance by optimizing perception or trajectory prediction, the following technical bottlenecks still exist when dealing with the complex heavy commercial trucks mentioned above: First, existing AEB perception modules lack stability and robustness in lane geometry modeling in curved scenarios. Changes in lighting, lane marking wear, and vehicle occlusion in complex traffic environments can easily cause lane line recognition to jitter or be interrupted. If the lane geometry model is unstable, it will directly lead to inaccurate judgment of the target vehicle's lane affiliation, thus amplifying the uncertainty of AEB decision-making. For example, Chinese patent CN120071278B discloses a lane line recognition and localization method based on hybrid attention feature enhancement, which enhances feature expression by designing a calibration optimization module. However, this method lacks explicit modeling and strong constraints on lane curvature and direction consistency. In extreme road conditions with large curve curvature, occlusion, or lane line breaks, the continuity and robustness of its lane fitting still need improvement, making it difficult to meet the high reliability requirements of heavy commercial truck AEB for environmental perception. Second, previous studies have shown insufficient ability to recognize the motion and predict the trajectory of out-of-control vehicles in AEB false triggering detection objects in curved scenarios. Most existing trajectory prediction algorithms are based on idealized assumptions about the driver and vehicle, assuming the target vehicle is in a stable and controlled state. This makes it difficult to identify dynamic instability characteristics such as skidding and rollover in the early stages of a vehicle crossing a lane, and also fails to provide phased predicted trajectories and spatial influence domains. For example, Chinese patent CN116946084A discloses a method for preventing false triggering of AEB during overtaking, which identifies whether a vehicle is making a normal lane change and overtaking by judging the consistency of the steering wheel angle, thereby deciding whether to suppress AEB triggering. However, this solution can only handle normal lane changes and cannot identify uncontrolled lane crossings caused by slippery road surfaces or operational errors (in which case the steering wheel angle and the actual vehicle trajectory are often inconsistent). Furthermore, such methods fail to provide the specific intrusion range and precise collision point of the out-of-control target, resulting in a lack of proactive avoidance basis for high-risk lane-crossing targets by the AEB system. Third, existing collision risk assessment models fail to effectively decouple the dependencies between variables. In risk inference using multi-source sensor fusion, many existing methods tend to assume that the various hazardous input variables are independent of each other to simplify calculations. However, in real traffic flow, there is a significant coupling effect between variables. Ignoring this dependency between variables can cause the system to fail to accurately capture the strong correlation between a vehicle's intention and its own safety when the vehicle first shows signs of crossing the line or slightly skidding, thus missing the optimal decision-making window.

[0004] In summary, there is an urgent need for an AEB triggering method that can stably construct lane geometry in curved scenarios, accurately identify the out-of-control state of cross-line targets and predict their trajectory and impact area in stages, while explicitly characterizing variable dependencies and outputting precise collision points and high-risk time windows in risk inference, so as to effectively solve the problems of poor adaptability and insufficient safety of existing heavy commercial truck AEB systems under complex working conditions. Summary of the Invention

[0005] In view of the shortcomings and deficiencies of existing technologies, the purpose of this invention is to provide an AEB triggering method to reduce false triggering in curve scenarios. The method first constructs a lane line recognition and lane geometry modeling module that integrates geometric constraints and temporal smoothing to form stable lane boundaries and centerlines, and establishes a Frenet coordinate reference to complete target vehicle lane attribution determination and curve distance compensation. After a cross-lane event is triggered, the out-of-control stage of the cross-lane target is identified, and a trajectory bundle and a three-dimensional dynamic occupancy envelope are generated under stage label constraints to obtain the target's future influence area and spatial conflict relationship. In the risk inference layer, a tree-enhanced Bayesian network is introduced to characterize the dependency structure between multiple hazard variables, outputting the posterior distribution of the collision point grid, the collision probability sequence, and the high-risk duration window. In the braking decision layer, forward collision avoidance requirements and rear-end collision prevention constraints are combined to construct a braking feasible region, and impact rate (jerk) constraints and damage costs are introduced to achieve graded AEB triggering and braking command optimization. Simultaneously, an event-triggered rolling update mechanism is adopted to reduce continuous computing power consumption while ensuring responsiveness in critical scenarios, enhancing onboard real-time performance and deployment feasibility, thereby providing a more realistic and stable quantitative basis for AEB triggering. Finally, the minimum forward braking requirement is formed based on the earliest collision step, collision probability sequence and high-risk duration window, and the rear-end collision constraint of the most dangerous rear vehicle is introduced to obtain the braking feasible region. Within the feasible region, jerk constraints and damage costs are added to complete the joint optimization of trigger threshold and braking intensity, and output the graded trigger results and braking commands such as warning / light braking / forced braking, so as to achieve early suppression of the risk of runaway intrusion and reduce the probability of false triggering.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A method for reducing false triggering of AEB (Automatic Emergency Braking) in curve scenarios, the method includes the following steps:

[0008] Step S1. Multi-source sensing data acquisition, time alignment, and preprocessing;

[0009] Step S2. Process the road image data obtained in Step S1 and input it into the constructed lane line recognition and lane geometry modeling module to form stable lane boundaries and center lines, establish a Frenet coordinate system, and complete the lane affiliation determination and curve arc length distance compensation for radar-detected target vehicles. If the target vehicle belongs to this lane, proceed to Step S4 to perform occupancy conflict assessment and collision risk inference; if the target vehicle is located outside the lane boundary but within the buffer zone, it is classified as a candidate dangerous target in the adjacent lane, and proceed to Step S3.

[0010] Step S3. Calculate the lateral distance from the candidate hazardous target to the nearest lane boundary and determine whether it shows a tendency to cross the lane. If so, integrate appearance, motion and interaction features to identify the out-of-control stage of the cross-lane target, and generate a trajectory bundle and a three-dimensional dynamic occupancy envelope under the stage label constraint.

[0011] Step S4. Based on the current state of the vehicle, predict the occupants corresponding to the future state and trajectory, and calculate the occupancy conflict probability by combining the occupancy envelope corresponding to the predicted trajectory of the candidate dangerous target; calculate the lane intrusion probability of the cross-line target based on the trajectory bundle of the cross-line target; construct multi-input indicators, use tree-enhanced Bayesian network to learn the dependencies between multiple dangerous variables, and infer the collision point and the future high-risk duration window in a rolling manner.

[0012] Step S5. Combine forward collision avoidance requirements and rear-end collision avoidance constraints to construct a braking feasible domain. Determine the urgency level and AEB trigger level based on the predicted nearest collision time. At the same time, introduce impact rate constraints and damage cost to optimize braking commands.

[0013] As a preferred embodiment of the present invention, the lane line recognition and lane geometry modeling module includes a lane line recognition network that outputs a pixel-level probability map based on the original road image, and a lane geometry modeling module that outputs lane boundaries, lane centerlines, and centerline curvature functions based on the pixel-level probability map. Step S2 specifically includes the following steps:

[0014] Step S201. Image Processing and Input Tensor Construction:

[0015] Grayscale images are extracted from the acquired original road images to obtain grayscale values. These grayscale values ​​are then processed to obtain horizontal and vertical gradient components, which are combined to form a gradient magnitude. This gradient magnitude is then combined with the original three-channel road image in the channel dimension to form a four-channel input tensor.

[0016] Step S202. The input tensor is processed based on the lane line recognition network. The lane line recognition network first maps the four-channel input to shallow features to obtain shallow feature maps; the encoder downsamples step by step; after downsampling, the intermediate features obtained by downsampling are added element-wise to the result after mapping by the residual branch, and the output of the l-th layer is obtained by nonlinear activation; the multi-scale fusion and attention enhancement module selects feature maps from multiple encoding layer outputs to participate in the fusion, aligns them and splices them to obtain multi-hop features; at the l-th fusion node, the features from different receptive field branches are jointly described to form a three-dimensional description map, and then the three-dimensional description map is aggregated along the branch dimension to generate the two-dimensional description quantity required for spatial attention gating, and then spatial attention weights are generated. The spatial attention weights are used to perform weighted fusion of multi-hop features; finally, decoding and recovery are performed, and after processing and fusion through two parallel coarse structure branches and detail repair branches, the feature of each pixel is mapped to an unnormalized score, and then mapped to a pixel-level probability map.

[0017] Step S203. Based on the lane geometry modeling module, perform lane line fitting and establish boundary functions:

[0018] After obtaining the pixel-level probability map, a binary lane line map is generated by thresholding. Then, connected component filtering and thinning are performed to extract the pixel sets of the left and right lane lines. The pixel sets are then mapped to the road surface plane coordinate system through inverse perspective mapping. After that, a fifth-order polynomial fitting is performed on the left and right lane line point sets to establish a boundary function with the vertical coordinate y as the independent variable, thereby outputting the lane center line and the curvature function of the center line.

[0019] Step S204. Lane attribution identification of surrounding vehicles and curve distance compensation:

[0020] First, the distance and azimuth measured by the millimeter-wave radar are converted into planar coordinates in the vehicle coordinate system. Based on the lane boundary function, it is determined whether the target vehicle belongs to the current lane. If it does, it is classified as a major hazardous target in the current lane. If it does not belong to the current lane, it is determined whether the target vehicle is located outside the lane boundary and within the buffer zone. If so, it is classified as a candidate hazardous target in the adjacent lane. Based on the identified lane centerline, the curvature correction is applied to the radar straight-line distance to obtain the actual distance between the current vehicle and the target vehicle.

[0021] Step S205. Coordinate system transformation:

[0022] Based on the lane centerline obtained in step S203, a Frenet coordinate system is established, and the target vehicle position obtained in step S204 is mapped into the Frenet coordinate system to form the longitudinal arc length coordinate and lateral offset coordinate of the target vehicle.

[0023] As a preferred embodiment of the present invention, step S3 specifically includes the following steps:

[0024] Step S301. Cross-line trigger:

[0025] Calculate the lateral distance from the candidate hazardous target to the nearest boundary. If the lateral distance in multiple frames within the detection window is greater than the boundary threshold, it is determined that there is a tendency to cross the line, and a set of cross-line targets is constructed. For each cross-line target, a set of neighborhood objects is constructed in the Frenet coordinate system, including the vehicle in front, the vehicle behind, and vulnerable traffic participants. Then, the neighborhood interaction priority score is calculated for each candidate traffic participant in the neighborhood object set of the cross-line target, and the top-k objects are selected to participate in the subsequent discrimination.

[0026] Step S302. Multimodal feature comparison and out-of-control evidence extraction:

[0027] Based on the original image of the cross-line target and the detection box, the target ROI image is cropped; the lane boundary output from step S2 is processed to generate a binary mask of the lane lines; the binary mask of the lane lines is cropped based on the detection box of the cross-line target to obtain the mask of the lane lines within the ROI; a context vector is constructed based on the positional relationship between the cross-line target and neighboring objects; the acquired data is processed using a multi-branch encoder, and then the multi-branch features are concatenated and fused using MLP to obtain the final frame-level fused features; for the extracted multi-frame data, the abrupt change intensity between adjacent frames is calculated, and the observation vector is constructed by comprehensively considering the sideslip intensity, roll amplitude, roll growth rate, longitudinal front and rear vehicle spacing and the closest distance of vulnerable traffic participants of the cross-line target;

[0028] Step S303. Phased inference of the coupled state machine:

[0029] Based on the observation vector constructed in step S302, a coupled state machine is established. ,in, Let m be the state of the target m at time t. For state geometry, This is the normal state. This indicates a normal crossover status. Indicates the start of a sideslip. Indicates a continuous sideslip state. This indicates the critical state of sideslip. This indicates that a sideslip has occurred; the loss of control process is subdivided into multiple stages, and the target state is matched to the corresponding stage at the moment of crossing the line;

[0030] Step S304. Defining the 3D envelope space based on stage labels:

[0031] Based on the target vehicle's position, geometry, heading angle, and roll angle, a proxy calculates the vehicle's three-dimensional directed envelope at time t. A three-dimensional semi-ellipsoidal reachable region is adaptively constructed based on the longitudinal, lateral, and vertical semi-axis lengths of the identified stage. Both are then processed using Minkowski summation to obtain the stage. The three-dimensional influence space of the next future moment is obtained, and the future H-step trajectory is predicted.

[0032] As a preferred embodiment of the present invention, step S4 specifically includes the following steps:

[0033] Step S401. Vehicle prediction and occupancy construction:

[0034] Based on the current state of the vehicle, construct the sequence of future states of the vehicle within the prediction window, and map it to a set of geometric occupants, thereby obtaining the future occupancy set of the vehicle;

[0035] Step S402. Construction of Multi-Input Metrics and TAN Learning:

[0036] Constructing a set of continuous variables with multiple inputs Among them, the stage hazard weight is: The probability of conflict is The lane intrusion probability is determined based on the vehicle's predicted occupant and the occupant corresponding to the nth trajectory of the cross-lane target; Based on the determination of N predicted trajectories for the cross-line target in the future; the normalized braking demand ratio is... The longitudinal TTC is determined based on the deceleration required to avoid longitudinal conflict and the upper limit of the maximum available deceleration under current attachment and load conditions; the longitudinal TTC is... The headway of the train is THW. ;

[0037] Each continuous variable is discretized into five states based on a threshold, resulting in a multi-input state vector. A tree-reinforced Naive Bayes network is used to characterize the dependency strength between input variables using conditional mutual information under class variable conditions, automatically learning the structural relationships between variables. Class variables represent the unit numbers of possible collision points at future time t+h. For random variables... Calculate conditional mutual information under class variable conditions, and construct TAN network topology based on conditional mutual information;

[0038] Step S403. Rolling Inference and High-Risk Time Window:

[0039] For each future prediction step, under the TAN network topology, the posterior probability of the collision grid cell class variable is inferred using the multi-input state vector as evidence, the grid cell number where the collision point is located is predicted, the collision probability of prediction step h is predicted, and the moment when the first collision danger event occurs is defined as the prediction step when the collision probability first exceeds the trigger threshold, thereby determining the first collision danger time and the future high-risk time period.

[0040] As a preferred embodiment of the present invention, step S5 specifically includes the following steps:

[0041] Step S501. Earliest collision step and braking feasible region:

[0042] Based on the time-series collision point set output in step S4, the prediction step where the earliest spatial conflict occurs is determined; based on the nearest distance to the collision point set and the vehicle speed, the minimum braking is constructed, the most dangerous following vehicle is selected based on the minimum required braking, the maximum allowable braking intensity in the rear is calculated, and the AEB braking feasible region is obtained.

[0043] Step S502. Graded triggering and smooth braking command:

[0044] When the high-risk duration window exceeds the set threshold, the appropriate AEB trigger level and braking intensity are selected within the braking feasible domain, taking into account the earliest collision risk advance time. The braking is made smoother by considering the impact rate constraint and the damage cost to the cargo and driver. Finally, the AEB graded trigger result and braking command are output.

[0045] As a further preferred embodiment of the present invention, the result of the intermediate features obtained by downsampling in step S202 after residual branch mapping is composed of a two-level convolution-normalization-activation mapping concatenation; when constructing the three-dimensional description map, channel average pooling is first performed on the features output by different receptive field branches to obtain the two-dimensional response map of each branch, and then the two-dimensional responses of each branch are stacked on the branch index dimension to form a three-dimensional description map; when generating spatial attention weights, spatial attention weights are obtained by weighted aggregation of neighborhood responses and Sigmoid transformation.

[0046] Output coarse-structured branches for each channel c ;in, , For interpolation weights, , , , For feature map In the On each channel, the feature values ​​at four integer grid points in the continuous coordinate neighborhood are represented by these four values, which correspond to the top left, top right, bottom left, and bottom right adjacent sampling points during bilinear interpolation.

[0047] Detailed fixes for branch output ;in, For the fine-branch convolution kernel parameters, For the corresponding bias term, where is the convolution kernel radius, and k is the kernel size. To refine the details, the total number of channels in the input feature map of the branch is adjusted. For detail repair, the input feature map is in the first branch. Input channels, spatial location The eigenvalue at that location.

[0048] Then the two are added together element by element at the same output resolution.

[0049] As a further preferred embodiment of the present invention, the multi-branch encoder in step S302 includes an appearance branch, a motion branch, and a geometric interaction branch. The appearance branch is used to extract the tire, vehicle body, posture, and relative relationship with lane lines. The input is a mask of the ROI image and the lane lines within the ROI. The appearance input tensor is first formed by channel concatenation, and then processed by ResNet-18 to output the appearance features. The motion branch is used to extract the motion consistency / abruptness of adjacent ROI images. First, dense optical flow is estimated on adjacent ROI images, and the dense optical flow is input into a lightweight convolutional neural network to obtain motion features. The geometric interaction branch is used to perform two-layer MLP encoding on the context vector to obtain context geometric interaction features.

[0050] As a further preferred embodiment of the present invention, in step S302, the target vehicle velocity vector is estimated based on the tracking displacement, and the motion direction angle is obtained from the velocity vector; based on the vehicle heading angle... , direction angle of motion Constructing side angle proxy Rolling corner agent ;in, For the roll-off estimation module, The target region, i.e., the target ROI image; the rollover rate. , These represent the inter-frame abrupt change intensity, sideslip intensity, roll amplitude, and roll rate of the target vehicle at time t, respectively; inter-frame abrupt change intensity , , For inter-frame feature difference, This represents the fusion feature of adjacent frames.

[0051] As a further preferred embodiment of the present invention, in step S402, the random variable... The expression for calculating conditional mutual information under class variable conditions is:

[0052] ;

[0053] in, For the discrete states of a random variable, Assign values ​​to the collision point element numbers. For random variables Values , Values Class variables Values The joint probability, In order to be in conditions, and The conditional joint probability, In order to be in under conditions Conditional marginal probabilities, In order to be in under conditions The conditional marginal probability.

[0054] As a further preferred embodiment of the present invention, in step S502, the earliest collision danger advance time is first determined. Determine the level of urgency Then map to AEB mode :

[0055] ;

[0056] in, The time threshold for the classification;

[0057] ;

[0058] WARN indicates only a warning; These are light braking and forced braking, respectively; COOP indicates "maximum feasible braking + rearward coordination warning" under the braking-only framework.

[0059] The discrete impact rate is:

[0060]

[0061] in, The upper limit of the impact rate, The braking intensity at adjacent time points, represent Impact rate at any moment;

[0062] Construct cargo and driver damage cost functions ;in, , These are the weighting coefficients;

[0063] Preset braking intensity ranges for each level:

[0064] ;

[0065] in, These are the parameters for graded braking;

[0066] Determining the braking sequence under constraints yields:

[0067] ;

[0068] in, This is the lower bound of the minimum forward braking requirement. The future braking deceleration sequence output by AEB. This is the upper limit of the maximum permissible braking in the rear direction.

[0069] Advantages and beneficial effects of the present invention:

[0070] (1) This invention addresses the problems of lane fitting jitter and inaccurate lane assignment determination caused by changes in illumination, wear of lane markings, occlusion and breakage, and large curvature curves. It proposes a lane modeling method that integrates geometric constraints and temporal smoothing. The continuity of lane lines is improved by gradient enhancement, multi-scale feature fusion and attention calibration. After outputting binary results, a fifth-order polynomial fitting and Frenet coordinate system are performed, so that the AEB system can still build a smooth and stable lane geometric benchmark even in extreme conditions where lane lines are partially missing. This improves the reliability of lateral offset calculation and lane assignment determination from the source of perception and significantly suppresses the risk of false AEB triggering caused by environmental noise.

[0071] (2) Unlike traditional methods that assume rational driving and simply extrapolate cross-lane behavior, this invention constructs a multi-stage motion mode covering normal cross-lane crossing, sideslip instability and rollover criticality. By extracting evidence from multiple frames in a short window triggered by cross-lane events and comparing temporal features, it can complete anomaly identification in the early stage when the target has not yet significantly invaded the lane but has already shown signs of sideslip or rollover. The coupled state machine outputs the posterior probability and stage label of the out-of-control stage, thereby significantly improving the early warning capability and active defense timeliness of heavy commercial trucks against sudden out-of-control intrusion.

[0072] (3) In view of the problem that heavy commercial trucks with high center of gravity and long wheelbase may cause large-angle side sweep and rollover tilt in the case of loss of control, resulting in actual space occupied far exceeding the static outline and easy to cause space omission, this invention designs a phased adaptive three-dimensional dynamic occupancy envelope model, which maps the dynamic states such as side slip and roll into the lateral and vertical envelope expansion, and combines Minkowski and constructs the worst-case occupancy envelope containing position uncertainty, realizing the upgrade from traditional point-to-point distance judgment to body-to-body spatial conflict detection, effectively reducing the collision risk omission caused by ignoring the geometric displacement and rotational deformation of the vehicle body.

[0073] (4) This invention overcomes the drawback of the assumption that the input variables are mutually independent in traditional risk assessment, which leads to inference distortion. It introduces a tree-enhanced Bayesian network (TAN) based on conditional mutual information to explicitly learn the dependency structure between unstable state, intrusion probability, occupation conflict and collision risk. Combined with a rolling update mechanism, it outputs a collision probability sequence and a high-risk duration window, enabling the system to distinguish between normal overstepping and uncontrolled intrusion and other complex interaction situations. While ensuring low false alarms in high-risk scenarios, it significantly reduces the false alarm rate in complex interference scenarios and improves decision confidence.

[0074] (5) The present invention constructs a refined collision point location method based on spatial grids, discretizes potential conflict areas into grid units and infers the posterior probability distribution of collision risk for each grid, thereby achieving a cognitive upgrade from whether or not a collision will occur and where it will occur. This not only improves the interpretability of risk output, but also provides a spatial decision-making basis for selecting more accurate risk avoidance directions and control strategies.

[0075] (6) In view of the special working conditions of heavy commercial trucks with long braking distance and easy rear-end collisions, this invention proposes a forward and backward coordinated braking decision architecture, calculates the minimum braking requirement for forward collision avoidance and the maximum allowable braking intensity for rear-end collision prevention in real time, and performs multi-objective optimization within the physical feasible domain, so that the output braking command can effectively avoid danger while taking into account the safety of the rear, and reduce the risk of secondary accidents caused by emergency braking.

[0076] (7) The present invention introduces cargo constraints in the braking execution layer and adopts the optimal solution strategy to make the braking change rate as smooth as possible. Under the premise of satisfying the forward and backward safety boundaries, it minimizes the risk of cargo tipping, liquid sloshing or damage to precision instruments caused by sudden braking, achieves the dual goals of active risk avoidance and damage-free transportation, and further improves the acceptability and applicability of the system engineering implementation.

[0077] (8) The present invention adopts an event-triggered computing resource allocation and rolling inference mechanism, and performs high-frequency prediction and probability update only when key events such as cross-line intrusion occur. While ensuring the core scenario response capability, it reduces the continuous computing power occupation and enhances the vehicle-mounted real-time performance and deployment feasibility of the AEB system. Attached Figure Description

[0078] Figure 1 The flowchart of the AEB triggering method for reducing false triggering in curve scenarios provided by the present invention is shown.

[0079] Figure 2 This is a flowchart of the input data processing by the lane line recognition network of the present invention;

[0080] Figure 3 This is a flowchart of the data processing by the multi-scale fusion and attention enhancement module of the present invention;

[0081] Figure 4 This is a flowchart of the data processing in the lane geometry modeling module of the present invention. Detailed Implementation

[0082] To enable those skilled in the art to better understand the technical solutions and advantages of the present invention, the present application will be described in detail below with reference to the accompanying drawings, but this is not intended to limit the scope of protection of the present invention.

[0083] like Figures 1 to 4As shown, this embodiment provides a method for reducing false triggering of AEB in curve scenarios. The method includes the following steps:

[0084] Step S1. Data Acquisition and Preprocessing:

[0085] The system monitors and acquires vehicle data, including the vehicle's current speed, acceleration, yaw angle, weight, and real-time control commands issued by the driver. It also collects surrounding environmental information, including the speed, acceleration, steering angle, yaw angle of all vehicles within a fixed range, and images of the road ahead, for use in step S2 to identify lane lines.

[0086] Specifically, in this embodiment, step S1 is used to complete the acquisition, time alignment, and preprocessing of multi-source sensing data, so as to provide unified, stable, and directly callable basic input data for subsequent steps. The multi-source sensing data includes at least: image frames acquired by the forward-looking camera. Target detection set output by millimeter-wave radar It includes at least target distance, azimuth and radial velocity information, as well as vehicle state parameters output by onboard state sensors. It includes at least wheel speed, vehicle speed, IMU yaw rate, longitudinal acceleration, and steering angle information; simultaneously, step S1 outputs the basic results of target detection and tracking, including target detection boxes. With target trajectory number ;in, This represents the two-dimensional detection box for target j at time t. This represents the trajectory number of target j.

[0087] Furthermore, in this embodiment, step S1 adopts a time alignment method with the camera as the main time axis. Specifically, the camera frame timestamp is used as the synchronization moment, and the image frame corresponding to that moment is recorded as the synchronization image. For radar data, the radar frame with the smallest difference between the timestamp and the synchronization moment is selected as the synchronization radar, and an expiration threshold is set to avoid using expired radar measurements. When the time difference exceeds a preset threshold, the synchronization radar is set to empty or the previous frame is used for holding. For continuous state quantities such as IMU, wheel speed, and steering angle, linear interpolation is used to align them to the synchronization moment to obtain the aligned vehicle state vector, which includes at least components such as vehicle speed, yaw rate, longitudinal acceleration, and steering angle. When there is a fixed clock offset, the clock offset can be calibrated and corrected before performing the above nearest neighbor matching and interpolation alignment. This forms an aligned data packet under a unified synchronization moment, which meets the requirement of subsequent steps to allow the joint use of multi-source information at the same moment.

[0088] Furthermore, in this embodiment, step S1 performs necessary preprocessing on the data after time alignment. For the image side, the synchronized image undergoes distortion correction, fixed-size scaling, and brightness and contrast normalization to obtain a preprocessed image, improving the recognizability of lane line textures under conditions of curve shadows, backlighting, or sudden brightness changes. For the radar side, the synchronized radar ensemble undergoes distance and speed threshold cleaning to remove outliers, and the radar measurements are structured into relative positions and approach speeds in the vehicle coordinate system for subsequent association with visual targets.

[0089] It should be noted that in this embodiment, the projection and transformation related to the lane coordinate system (Frenet) depend on the lane centerline output in step S2, and therefore are not performed in step S1. For the target side, the trajectory number and detection box of each target are output, and a short-term window buffer is established for each trajectory number to store the sequence of detection boxes and their corresponding ROI clipping results within a short time second, so that evidence from multiple frames before and after the cross-line analysis is triggered in step S3 can be directly extracted. For the vehicle state side, light filtering is applied to the vehicle speed and yaw rate, and zero-bias correction is applied to the steering angle to reduce the impact of noise on the prediction and control optimization of the vehicle's future state; the data packet output in step S1 can be considered as the window buffer.

[0090] Step S2. Lane geometry construction and target lane assignment:

[0091] Under curved conditions, construct the geometric boundary and centerline model of this lane, and based on this, classify the target vehicles ahead into dangerous targets in this lane and targets in adjacent lanes. The specific steps are as follows:

[0092] Step S201. Image Processing and Input Tensor Construction:

[0093] Based on the input image acquired by the camera in step S1 Extract the grayscale image to obtain the grayscale values;

[0094]

[0095] in, The image from the forward-looking camera at time t. , Here, represents the height and width of the image, respectively, and t represents the time on the unified time axis. For pixels grayscale value, , , pixels In the RGB three-channel pixel values, , , These are the channel weight coefficients.

[0096] The obtained grayscale values ​​are processed to obtain their gradient values; specifically, the gradient components in the horizontal and vertical directions are first obtained, expressed as follows:

[0097]

[0098]

[0099] in, , pixels The horizontal and vertical gradient components at the location, , These are the weighting coefficients of the Sobel level kernel coefficients, which are preset constants with values ​​set to [value missing]. , For pixels The grayscale value.

[0100] To form a single-channel gradient intensity representation, the gradient components are synthesized into a gradient magnitude. , .

[0101] After obtaining the gradient magnitude map, it is combined with the original three-channel image in the channel dimension to form a four-channel input:

[0102]

[0103] in, , , , These represent the pixel values ​​and gradient magnitudes of the front-view camera image in the RGB three channels at time t.

[0104] Step S202. Process the input data based on the lane line recognition network:

[0105] Four-channel input The input to the lane line recognition network is first mapped into a multi-channel "feature map". In this embodiment, to uniformly describe the calculation of each layer within the lane line recognition network, A represents the input feature map of a certain layer, and B represents the output feature map of that layer. That is, for the initial layer of the lane line recognition network... In subsequent layers, A equals the feature map output by the previous layer, meaning that the input of the current layer in the lane line recognition network comes from the calculation result of the previous layer.

[0106] The core operation of each layer of the lane line recognition network is to perform a weighted summation of the input feature map A in its local neighborhood to generate an intermediate value, and then perform normalization and nonlinear mapping to obtain a more separable feature representation.

[0107]

[0108] in, For the input feature map, This is the output feature map (intermediate value) of this layer. , These are the learnable parameters for this layer. , Let the input and output channel indices be the input and output channel indices, and the input channel index satisfy... , That is, the nuclear radius, where k is the nuclear size; The number of channels in the input feature map A (i.e., the total number of channels that the convolutional kernel needs to accumulate in the channel dimension).

[0109] Since the feature amplitude distribution may vary in different batches or scenarios, in order to maintain the stability of training and inference, each channel is normalized and learnable scaling and translation are introduced.

[0110]

[0111] in, The features are normalized. , Let be the mean and variance of channel c. and For learnable scaling and translation parameters, To prevent division by zero constant.

[0112] After normalization, nonlinear activation is introduced to enhance the expressive power of lane line morphology and improve class separability. In this embodiment, , The activation function is used to address the feature distribution shift of heavy commercial trucks under different lighting conditions, ensuring the robustness of inference.

[0113] Specifically, in this embodiment, the lane line recognition network first inputs the entrance (four-channel input). Mapping to shallow features yields a shallow feature map. :

[0114]

[0115] in, Input feature map The result after weighted summation and normalization.

[0116] To obtain a larger receptive field, the encoder performs stepwise downsampling:

[0117]

[0118] in, For the first Sub-sampling of intermediate features at the layer level This is the encoding feature of layer 0. , For the first Layer parameters, For the first layer Channel pixels The encoding features are obtained; after downsampling, in order to enhance the transferability and stability of slender structural features such as lane lines at the encoding end, this embodiment uses residual refining to enhance the intermediate features.

[0119] Specifically, the intermediate features obtained by downsampling The result obtained after residual branch mapping is added element-wise, and the output of the l-th layer is obtained by nonlinear activation:

[0120]

[0121] in, As an element-wise nonlinear activation operator, used to introduce nonlinear expressive power and suppress negative responses, a modified linear unit can be used in this embodiment. . For the reason The residual mapping is obtained after activation following two weighted summations and normalizations. It consists of a concatenated mapping of two levels of convolution, normalization, and activation; the first-level mapping is denoted as... The output of the second-level mapping is ,but:

[0122]

[0123] in, For convolution operations, , and , Learnable convolution kernels and biases for residual branches. As a normalization operator, it introduces learnable scaling and translation parameters for each channel. To enhance expressive power. Through the above residual mapping Features of the main branch The superposition of these features can enhance the continuity and noise resistance of slender structural features without destroying the semantic information of the main branch, thereby improving the lane line representation effect. Furthermore, if the main branch... With residual branch If the number of channels is inconsistent, then before adding, adjust the... Applying a 1x1 convolution for dimension alignment yields... and use Instead of the above .

[0124] At the multi-scale fusion point, instead of directly taking the single-scale encoded features, feature maps that participate in the fusion are selected from the outputs of multiple encoding layers, aligned and spliced ​​to obtain multi-hop features, and then spatial attention weights are generated to achieve selective fusion, thereby enhancing the continuity of the curve direction.

[0125] To avoid confusion between the terms "scale" and "network layer," this embodiment denotes the set of coding layer indices participating in multi-scale fusion as... in For the set of coding layers participating in the fusion, The number of scales participating in the fusion. Represents the first element in the set. One coding layer index, Therefore, the so-called "first" "Scale Feature Map" refers to the feature map of the first scale. The feature map output by the layer encoder is denoted as in For time index, The first The number of channels, height, and width of the layer-encoded feature map.

[0126] To achieve point-by-point fusion of features from different coding layers at the same spatial resolution, feature maps at each scale are first aligned to a common scale. ,in, and Let these be the height and width at the common scale, respectively. Let the pixel coordinates at the common scale be... ,in Then the pixel coordinates on the common scale In the original feature map The corresponding continuous coordinates are as follows Its resolution ratio is determined by the resolution ratio. After obtaining continuous coordinates, take four integer grid points in its neighborhood to obtain:

[0127]

[0128] And define the decimal offsets in the horizontal and vertical directions as follows: ; where ⌊⋅⌋ represents the floor operation, This indicates the horizontal offset ratio of continuous coordinates relative to the integer point to the left. It represents the vertical offset ratio of a continuous coordinate relative to the integer point above it.

[0129] Therefore, the aligned first Each scale feature map is denoted as Its position ,aisle The eigenvalues ​​at a given point can be obtained using bilinear interpolation:

[0130]

[0131] in, Indicates the first After the layer-encoded feature maps are aligned to a common scale, at the location and channels The characteristic response at that point. Due to the number of channels in the output of different coding layers. They may differ. To ensure that features of different scales can be uniformly stitched together along the channel dimension, for each aligned feature map... Applying a channel linear transformation, we obtain:

[0132]

[0133] in, and All are learnable parameters. After alignment, the multi-scale features are concatenated to form a multi-hop input:

[0134]

[0135] in, For the set of scales participating in the fusion, To indicate the first At each fusion node, in spatial location The above is a multi-hop input feature vector formed by concatenating multiple aligned scale features in the channel dimension.

[0136] Next, spatial weights are generated to assign higher weights to the directional positions of continuous lane lines in the curve region. This is done by first averaging the channels to obtain a two-dimensional description:

[0137]

[0138] in, For spatial description diagram, For multi-hop input feature vectors In the passage The component value at that location, This represents the total number of channels after concatenation. Then, the local weighted sums are applied and transformed using a Sigmoid function to obtain the spatial attention weights.

[0139] Furthermore, to establish consistent spatial guidance among multi-hop / multi-scale features, this embodiment performs a joint description of features from different receptive field branches at the l-th fusion node. Let the output feature of the i-th branch be... To obtain the structural response that is only related to spatial location, the channel dimensions are first aggregated to obtain the two-dimensional response map of the nth branch. :

[0140]

[0141] in, This represents channel-average pooling. Furthermore, the two-dimensional responses of each branch are stacked on the branch index dimension to form a three-dimensional descriptive graph:

[0142]

[0143] in, This indicates that the two-dimensional response maps of each branch are stacked sequentially along the branch index dimension to form a shape of size [missing information]. The three-dimensional description map is obtained. To generate the two-dimensional description required for spatial attention gating, the three-dimensional description map is aggregated along the branch dimensions to obtain... ,in For maximum value calculation, based on the above Define spatial attention weights for:

[0144]

[0145] in, Indicates the first Layer in position Lane line response characteristics, and The learnable parameters are obtained by weighted aggregation of neighborhood responses and then mapped using a Sigmoid algorithm. This weight is used to characterize spatial location. The overall importance of lane line structure. Since the weights are driven by neighborhood consistency, when surrounding pixels belong to the same lane line structure... The weights will increase accordingly, thereby structurally strengthening continuous lane segments; when wear, occlusion, or local breakage causes unstable neighborhood responses, the weights will decrease relatively, making the network more reliant on multi-scale semantic information to complete completion and noise suppression during the fusion phase.

[0146] Finally, spatial attention weights are used to enhance multi-hop features positionally. Because As scalar weights that are only related to spatial location, they are applied to the channel dimension in a broadcast manner. The characteristics of each channel, denoted as element-wise multiplication. Then the weighted fusion can be expressed as: ,in A configurable or learnable scaling factor is used to control the magnitude of the enhancement, thereby strengthening the lane structure response while avoiding over-amplification of continuous areas.

[0147] Finally, decoding and restoration are performed to recover the semantic information at high resolution. Upsampling is also achieved using the form of a "linear combination of four neighborhood points":

[0148]

[0149] in, After upsampling (the first) (Layer) grid coordinates, , To map to the Continuous coordinates of the layered grid For the first The layer feature map size is then determined, and the continuous coordinates are decomposed into four adjacent integer grid points and interpolation weights:

[0150]

[0151] in, For the first Layer neighborhood integer coordinates, , As interpolation weights, for each channel c, the upsampled feature value is obtained by a linear combination of its four neighboring points:

[0152]

[0153] in, For the first Layer decoding features upsampled to the first The result after layer resolution, They represent the first Layer decoding feature map In the passage The eigenvalues ​​located at the four integer grid positions of the upper left, upper right, lower left, and lower right corners of the continuous coordinate neighborhood. Let... For the encoder's number Layers, same resolution, i.e. The jump-connection fusion feature.

[0154] In the decoding stage, this invention employs a strategy of recovering data step-by-step from deep to shallow layers. Assume the encoder has a total of... Layers, Indexes This represents the lowest resolution layer. Encoding process. The decoding process follows a step-by-step downsampling approach. Upsampling is performed level by level for recovery. Therefore, the decoding computation order is from deep to shallow, but the layer index itself always represents the corresponding resolution layer, and the final decoded output is located at the highest resolution layer. Define the... Fusion input of the level decoding unit The channel is structured as follows:

[0155]

[0156] in, To fuse input features, For the first Number of decoding channels per layer For the first Number of layer jump connection fusion channels.

[0157] Let the final decoded output be To repair the breakage of slender lane lines, two parallel branches are introduced: a coarse structure branch, emphasizing structural stability and strong continuity (mainly using interpolation upsampling); and a detail repair branch, emphasizing boundary details and breakage repair (mainly using learnable local weighted summation).

[0158]

[0159] in, , To make them learnable parameters, then... Upsampling to the output resolution (H, W) yields:

[0160]

[0161] in, The output of the coarse structure branch is used to directly generate detailed enhancement features using a learnable local weighted summation, resulting in:

[0162]

[0163] in, For the fine-branch convolution kernel parameters, For the corresponding bias term, Where k is the kernel radius (k is the kernel size). This occurs when the region exceeds the feature map boundary. In this embodiment, zero-padding or boundary duplication can be used for boundary processing to ensure that convolution calculations are feasible. Coarse branches emphasize continuity, while fine branches emphasize detail; both are added and merged element-wise at the same output resolution.

[0164]

[0165] The features of each pixel are mapped to unnormalized scores, which are then mapped to the final probability; where the scores are:

[0166]

[0167] in, , For classification layer parameters, for The number of channels is then used to map the scores to probabilities:

[0168]

[0169] in, It is an exponential function. Output resolution grid coordinates.

[0170] Step S203. Based on the lane geometry modeling module, perform lane line fitting and establish boundary functions:

[0171] After obtaining the pixel-level probability map, a binary lane line map is generated by thresholding.

[0172]

[0173] in, As a binarization threshold, to obtain a continuous and fitable set of lane boundary points, the binary mask is... Connectivity component filtering and refinement are performed to extract the pixel sets of the left and right lane lines, denoted as follows: Extracting the set of lane line pixels from the binary image: Considering the need for subsequent lane geometry calculations and curve distance corrections within the vehicle plane coordinate system, the pixel set is mapped to the road surface plane coordinate system using inverse perspective mapping (IPM). Let the mapping from pixel to road surface plane be... Then we have:

[0174]

[0175] in, , These are the coordinates of a point in the ground coordinate system.

[0176] In the road surface coordinate system, to obtain an analytical lane boundary function, a fifth-order polynomial fit is performed on the left and right lane line point sets respectively to establish a boundary function with the longitudinal coordinate y as the independent variable: in These are the fitting coefficients, which can be obtained by least squares solution;

[0177] Finally, output the lane centerline:

[0178]

[0179] The first and second derivatives can be directly calculated from the centerline function: Based on this, the curvature function of the centerline is given: .

[0180] Step S204. Lane attribution identification of surrounding vehicles and curve distance compensation:

[0181] In this embodiment, for target vehicle measurements transmitted back by millimeter-wave radar, it is first determined whether the target belongs to the current lane, thus classifying the target into primary hazardous targets and candidate hazardous targets. Further considering the inconsistency between straight-line distance and actual distance along the lane in curved road sections, the radar straight-line distance is corrected for curvature based on the identified lane centerline to obtain the actual distance between the current vehicle and the target vehicle. After this step is completed, all distance measurements involved in subsequent steps use the corrected curved distance as a unified distance caliber.

[0182] First, the distance and azimuth measured by the millimeter-wave radar are converted into planar coordinates in the vehicle coordinate system:

[0183]

[0184] in, Number the target vehicle. For the current moment, The straight-line distance between the target vehicle and this vehicle as measured by the radar. The millimeter-wave radar azimuth angle of the target vehicle j relative to the longitudinal forward axis of the vehicle body;

[0185] Based on the lane boundary function, the left and right boundary lateral coordinates are taken at the longitudinal position of the target to determine the target vehicle. Whether a lane belongs to this lane is indicated by the lane ownership indicator. and define indicator functions The expression means: the value is 1 if the condition inside the parentheses is true, and 0 otherwise. Therefore:

[0186]

[0187] That is when At that time, the target vehicle was identified. Located within this lane; when At that time, the target vehicle was identified. Located outside this lane. A lateral buffer threshold is introduced to indicate whether a target vehicle belongs to the current lane, and to capture candidate targets that may intrude from adjacent lanes. If the target vehicle is located outside the lane boundary but within the buffer zone, it is classified as a candidate hazardous target.

[0188]

[0189] This results in two sets of objectives: ,in This is the set of the main hazardous targets in this lane. This is the set of candidate hazardous targets in adjacent lanes. For the set... The target in the sequence already meets the lane ownership criteria, therefore cross-lane screening is no longer performed. Instead, it is directly input as the forward primary hazard target into step S4 to calculate longitudinal TTC, headway THW, normalized braking demand ratio, and occupancy conflict risk with the predicted occupant of the vehicle. This information then participates in AEB trigger level determination and braking command optimization in step S5. For the set... The target in the process does not directly trigger AEB. Instead, it is first input into step S3 for cross-lane event triggering, continuous consistency verification, loss of control stage identification and trajectory bundle prediction. Only when it forms a stable cross-lane target and shows a high probability of lane intrusion, lane occupancy conflict or collision risk in step S4 will it participate in the AEB braking decision in step S5 as a cross-lane dangerous target.

[0190] Furthermore, since the distance provided by radar is a straight-line distance, while the vehicle is actually traveling along the lane, AEB should use the arc length along the centerline in curved scenarios. Therefore, the projection parameters of the target point on the centerline are first calculated, and then the arc length is calculated.

[0191] target point Projected onto the lane centerline Find the projection parameters that minimize the Euclidean distance between the two objects. Taking the derivative of the squared distance function and setting it to zero, we obtain the optimal condition that is satisfied:

[0192]

[0193] in, The projection parameters are obtained analytically from the polynomial fitted in step S203. Since this equation is nonlinear, Newton's iteration is used to solve it, yielding the projection parameters. After obtaining the projection parameters, the arc length of the lane centerline is used as the curve correction distance between the vehicle and the target vehicle. Let the reference parameter for the vehicle on the centerline be... Then the curve distance of the target vehicle j is defined as: This involves precise integral compensation for the future trajectory length of the vehicle. This ensures that the distance parameters input to the AEB system are physically equivalent to the actual remaining travel distance, thereby eliminating the nonlinear interference of road curvature on hazard assessment.

[0194] To meet real-time requirements, the integral is discretized and approximated using the trapezoidal rule:

[0195]

[0196] in, ;in, To integrate the interval The number of subintervals after uniform division, which is the number of discrete segments in the trapezoidal method, is taken as a positive integer. To indicate the first The vertical coordinates corresponding to each discrete sampling point are located from... arrive On the integration interval, To represent the vertical coordinate The distance from the step in the direction, that is, the distance of each small trapezoid in... Width on the axis.

[0197] Step S205. Coordinate system transformation:

[0198] To ensure that subsequent identification of the target vehicle's cross-lane behavior, judgment of intrusion trends, and calculation of collision risks have a unified coordinate representation consistent with road geometry, a Frenet coordinate system is established based on the lane centerline obtained in step S203, and the target vehicle position obtained in step S204 is mapped into the Frenet coordinate system to form the longitudinal arc length coordinate and lateral offset coordinate of the target vehicle.

[0199] The lane centerline is parameterized as follows: Differentiating it with respect to the parameter y, we obtain the tangential direction of the centerline at y and normalize it:

[0200]

[0201] Further construct a unit normal vector orthogonal to the tangent vector:

[0202]

[0203] For target vehicle j, its position in the lower plane of the vehicle coordinate system is: Its projection parameters on the centerline are obtained from step S204. The corresponding projection points are:

[0204]

[0205] The displacement vector of the target point relative to the projection point is: Therefore, the lateral offset of the target vehicle in the Frenet coordinate system is defined as the projection of the displacement onto the normal direction:

[0206]

[0207] This refers to the lateral deviation of the target vehicle relative to the lane centerline; simultaneously, the Frenet longitudinal coordinate is defined using the arc length along the centerline. Since step S204 has already uniformly corrected the actual longitudinal distance between the target vehicle and the current vehicle to the centerline arc length... Therefore, the longitudinal arc length coordinate of the target vehicle is defined as ;

[0208] Therefore, the coordinates of the target vehicle in the Frenet coordinate system are represented as follows: .

[0209] Step S3. Cross-line behavior recognition and classification:

[0210] Step S301. Cross-line triggering and window capture:

[0211] For any adjacent lane candidate hazardous target Let its center point in the vehicle coordinate system be denoted as . The center point of the candidate dangerous target Projected onto the lane centerline output by step S2 The vertical arc length coordinate and horizontal offset coordinate in the Frenet coordinate system are obtained as follows: ;in, Candidate hazardous targets in adjacent lanes At any moment The vehicle center point plane coordinate vector, These are the lateral and longitudinal coordinates of the center point in the vehicle coordinate system, respectively. To the center line of the lane The minimum distance projection operator, wherein the lane centerline is obtained by fitting in step S2, and is expressed with arc length parameter. Represented as a curve denoted by the arc length parameter s. ,Right now:

[0212]

[0213]

[0214]

[0215] in, The nearest projection point of the candidate hazardous target's center point onto the lane centerline. Let be the unit normal vector of the lane centerline at the projected location. Represents the vertical arc length coordinate. This indicates the lateral offset relative to the lane centerline. The unit normal vector of the centerline. The centerline is obtained by lane identification and fitting in step S2.

[0216] The lateral boundary functions of the left and right boundaries of this lane in the Frenet coordinate system are as follows: and ,but and These represent the longitudinal projection positions of the candidate hazardous targets. At this location, the lateral coordinates corresponding to the left and right boundaries of this lane are defined. Based on this, the signed lateral distance from the candidate hazardous target to the nearest boundary is defined. for:

[0217]

[0218] Constructing the cross-line indicator:

[0219]

[0220] in, Indicate candidate target Does the line cross at time t? For out-of-bounds threshold, For indicator functions, i.e. when Define the cross-line event trigger time for this candidate target: . Candidate targets The moment of the cross-line event.

[0221] To reduce false triggering, a "continuous consistency" constraint is introduced:

[0222]

[0223] in, To ensure stable cross-line determination, To verify the window length, To minimize the number of triggers (i.e., at least several frames of triggers are required for a valid line crossing), a set of line-crossing targets is constructed for all candidate targets. ,in The set of targets where stable cross-line events occur.

[0224] Define the cross-line analysis window ,in For the goal The cross-line front and back windows, The window width is half. For each cross-line target, a neighborhood object set is constructed within the Frenet system, including the preceding vehicle, the following vehicle, and vulnerable traffic participants (VRUs): The neighborhood object set It is not a static collection, but rather a cross-line analysis window. The time t within the frame is obtained by updating it frame by frame.

[0225] That is, at every moment Below, the system is based on cross-line targets. Frenet longitudinal arc length coordinates Longitudinal arc length coordinates of surrounding traffic participants Real-time updates and The selection results ensure that the timing of the determination of preceding and following vehicles is consistent with the timing of the cross-lane event; among which, Indicates the nearest vehicle ahead of the target crossing the line. This indicates the nearest vehicle behind the target vehicle on the cross-line. The set of VRUs in the neighborhood (pedestrians / non-motorized vehicles, etc., obtained by detection in step S1).

[0226] For each stable cross-line target, a set of neighborhood objects is constructed in the Frenet coordinate system, including the vehicle in front, the vehicle behind, and vulnerable traffic participants.

[0227] Taking the determination of preceding and following vehicles as an example, the principle of minimum value is used to determine the situation within the same lane or adjacent lanes:

[0228]

[0229]

[0230] in, Indexing candidate traffic participants Candidate traffic participants The vertical arc length coordinate.

[0231] For cross-line targets neighborhood object set Candidate traffic participants Calculate neighborhood interaction priority score The candidate objects are then sorted in descending order based on their scores to select the top few objects most relevant to the cross-line analysis for subsequent judgment. The priority score is defined as:

[0232]

[0233] in, For cross-line hazard priority scores, , The target Vertical arc length and horizontal offset in Frenet coordinate system These are the coordinates of the candidate object 𝑘. , Here, is the weighting coefficient; is a small constant to prevent division by zero. This is the lateral scale parameter.

[0234] Step S302. Multimodal feature comparison and out-of-control evidence extraction:

[0235] For cross-line targets Each frame within the window Extract the target region and overlay a lane semantic mask:

[0236]

[0237] in, The original image at time t, For cross-line targets The detection box is obtained by acquiring and tracking the target state in step S1 (detection - association - tracking). The outer margin is a fixed ratio (e.g., 10%–20%) or a fixed pixel value of the detection box width and height to ensure that pose cues such as tires and vehicle side profiles are included. Indicates the detection box Expand outwards to obtain the expanded target area. This indicates that the target ROI is obtained by cropping.

[0238]

[0239] in, For the binary mask of lane lines within the ROI, The lane lines and boundary geometry output in step S2; This means projecting geometry into a pixel-space mask; specifically, it means projecting... Lane boundaries or centerlines are represented as polylines or sampling points. A binary mask is generated by rasterizing the projected lane lines in the entire image coordinate system. This binary mask is then applied to the region. The ROI mask is obtained by inner trimming; thus, the ROI mask is obtained. Its pixel value reflects the positional constraints of lane lines / boundaries within the ROI, and is used to enhance the association features between cross-line targets and lane structures.

[0240] Construct a fused feature vector for each frame:

[0241]

[0242] in, Let D be the fused feature of frame t, and D be the feature dimension. For context vectors, This is a multi-branch encoder, which includes an appearance branch, a motion branch, and a geometric interaction branch. The appearance branch is responsible for extracting static textures and shape cues such as tires, vehicle body, pose, and relative relationship with lane lines. The input is a ROI image and a lane mask. Within the branch, channel stitching is used to form the appearance input tensor. The appearance branch model uses ResNet-18, and the expression is: ,in Describes appearance characteristics.

[0243] The motion branch is responsible for extracting the motion consistency / abruptness of adjacent ROIs, preferably using optical flow as an explicit motion representation. First, dense optical flow is estimated on adjacent ROIs. ,in For optical flow estimation operators, dense optical flow is input into lightweight CNN encoding to obtain motion features. ,in, (⋅) represents the motion branch coding network, preferably constructed from a lightweight convolutional neural network, used to encode the dense optical flow between adjacent frame ROIs. Mapped to motion feature vectors To characterize cross-line targets Dynamic clues such as translation, rotation, local deformation, and sudden changes in motion at time t.

[0244] The geometric interaction branch involves performing two-layer MLP encoding on the context vector. .in, (⋅) is a geometric interactive branch coding network, preferably composed of a two-layer, multi-layer sensing mechanism; For the goal The contextual geometric interaction features at time t are used to characterize the local interaction relationships between the vehicle in front, the vehicle behind, and vulnerable traffic participants.

[0245] The three-branch features are concatenated and then fused using MLP to obtain the final frame-level fused features. , among them Characterizing MLP models, The network is a frame-level feature fusion network, preferably composed of multi-layer perceptrons, used for appearance feature fusion. Motion characteristics Geometric interaction features Perform joint mapping and output cross-line targets. The final frame-level fusion features at time t .

[0246] The context vector is defined as the geometric interaction quantity of the neighborhood objects described in step S301:

[0247]

[0248] in, This represents the longitudinal distance between the target across the line and the vehicle in front of it in the Frenet coordinate system. The longitudinal distance between the following vehicle and the target crossing the line. This represents the minimum Euclidean distance in the Frenet coordinate system between the cross-line target and its nearest vulnerable traffic participant. This represents the change in distance between the cross-line target and the nearest vulnerable traffic participant at adjacent time points, used to characterize whether the VRU is rapidly approaching or moving away from the cross-line target.

[0249] For the extracted multi-frame images, construct the feature difference between adjacent frames and the abrupt change intensity:

[0250]

[0251]

[0252] in, For inter-frame feature difference, The intensity of inter-frame mutation. It is a 2-norm. The larger the value, the more drastic the change before and after crossing a lane; normal lane changes are usually smooth and continuous, while sideslips or rollovers are often accompanied by significant abrupt changes in local posture or tire area.

[0253] Estimate the target vehicle velocity vector based on tracking displacement:

[0254]

[0255] in, For tracking speed estimation, The position of the target in the lane coordinate system. This is the inter-frame time interval. The multi-target tracking output from step S1 (which can fuse camera, millimeter-wave radar or V2X); Obtained from timestamps or a fixed frame rate.

[0256] The direction angle of motion is obtained from the velocity vector:

[0257]

[0258] in, To represent the target vehicle's motion direction angle calculated from the velocity components, Let be the arctangent function, then we have , Indicates the target vehicle The tracking velocity vector at time t lies in the adopted planar coordinate system. The components in the y-axis and y-axis directions.

[0259] Define vehicle body heading angle Construct side-slip angle proxy quantity :

[0260]

[0261] in, This means normalizing the angle difference to... .

[0262] Define roll angle proxy for:

[0263]

[0264] in, This is the roll-off estimation module.

[0265] Define the rolling growth rate :

[0266]

[0267] It should be noted that, All are cross-line target vehicles The amount of surrogate evidence for sideslip / rollover is a combination of mutation, sideslip, and rollover evidence, and the observation vector is defined as:

[0268]

[0269] in, Let be the observation vector at time t. Each component represents the target vehicle. The inter-frame abrupt change intensity, sideslip intensity, roll amplitude, and roll rate at time t. It is used to describe longitudinal distance between vehicles and the closest distance for vulnerable road users, and to characterize interaction constraints and available avoidance space.

[0270] In this embodiment, the observation vector is used as the observation evidence input to the coupled state machine in step S303, for constructing the observation likelihood. And based on this, the posterior phase is calculated recursively. This allows for continuous phased assessment and risk level updates of the target vehicle's sideslip and rollover evolution stages. Unlike relying on a single indicator (such as lateral offset or roll angle threshold), By jointly characterizing mutations, sideslip, roll, and interactive context, the stability of stage identification can be maintained even in the presence of occlusion, measurement noise, or lane geometry errors, and a consistent quantitative basis can be provided for subsequent stage transition constraints and hazard triggering.

[0271] Step S303. Phased inference of the coupled state machine:

[0272] The observation vector sequence constructed based on step S302 A coupled state machine is established that has a forward-backward correlation and a transformable relationship between sideslip and rollover. The runaway process is subdivided into multiple stages, and the target state is matched to the corresponding stage at the cross-line moment.

[0273] The coupled state machine serves to establish a hierarchical filtering mechanism based on hazard priority. Unlike traditional methods that rely solely on a single threshold to determine hazard, this invention decouples the vehicle's motion process into two paths: 'normal maneuvering' and 'loss of control evolution', using a state transition matrix. Specifically, this state machine utilizes the physical inclusion relationship between rollover and skidding to force the system to prioritize assessing high-risk states (rollover) at each moment, followed by medium-risk states (skidding), and only after eliminating the aforementioned risks does it use the cross-line indication quantity (…). Determine whether it is a normal lane change. This addresses the technical challenge of misinterpreting lateral motion components generated by normal vehicle cornering or lane changes in curves as sideslip instability.

[0274] Based on finitely coupled state machines:

[0275]

[0276] in, Let m be the state of the target m at time t. For state geometry, This is the normal state. This indicates a normal crossover status. Indicates the start of a sideslip. Indicates a continuous sideslip state. This indicates the critical state of sideslip. This indicates that a sideslip has occurred.

[0277] Define the probability of side-sliding gate control:

[0278]

[0279] in, For the Sigmoid function, , , These are the weighting coefficients. represents the sideslip confidence level. It indicates the probability of the target vehicle R being subjected to sideslip gating at time R, and is used to measure the strength of the probability of whether a sideslip state is triggered.

[0280] Define the probability of side-tipping and pushing gate control:

[0281]

[0282] in, These are the weighting coefficients. This represents the total gating probability of entering the rollover critical stage. The rollover propagation refers to the evolutionary trend of the target vehicle from the controllable sideslip stage to the rollover critical stage, and is used for control stage transition and risk level recursive update.

[0283] Define the probability of continuous gating of sideslip:

[0284]

[0285] in, For the probability of sideslip continuation, and These are the mean values ​​within a sliding window of length L, i.e. , It is used to measure the probability intensity of whether sideslip is a continuous rather than a momentary shudder.

[0286] Define the recovery gating probability:

[0287]

[0288] in, , All are recovery thresholds. , , This is a parameter. This parameter represents the recovery gating probability from a sideslip state to a stable state, and is used to measure the probability strength of whether the vehicle has returned to a controllable state.

[0289] Define the gating probability of rollover:

[0290]

[0291] in, The roll angle threshold, The rolling growth rate threshold, , , This is a parameter. This parameter is used to measure whether the target vehicle has approached or reached the conditions for a rollover.

[0292] Define the probability transition matrix:

[0293]

[0294] In this matrix, rows represent the current state, columns represent the state at the next time step, and the last row represents the absorption state.

[0295] Each element is constructed according to the gating probability as follows:

[0296] 1. From The transfer:

[0297]

[0298]

[0299] 2. From The transfer:

[0300]

[0301]

[0302] It should be noted that the multiplication term introduced in the above transfer formula... This indicates a combined condition where both rollover risk and sideslip risk are at a low level. Its physical meaning is: effective rollover risk is triggered only when the target vehicle crosses the lane. Only when there is no obvious tendency to roll over or slide, and no obvious sideslip characteristics, will the system allow the state to remain with a high probability. Conversely, when the probability of any risk gate increases, This will rapidly decrease, thereby suppressing the self-loop of the "normal crossing" state and diverting the probability mass to or This design enables a strict distinction between normal lane changes and unstable lane changes (side slip / rollover) under high-risk evolutionary conditions. While ensuring a sensitive response to real instability processes, it significantly reduces the probability of false triggering and false braking caused by fluctuations in single pieces of evidence during routine lane changes / crossings of heavy commercial trucks.

[0303] 3. From The transfer:

[0304]

[0305]

[0306] 4. From The transfer:

[0307]

[0308]

[0309]

[0310] 5. From The transfer:

[0311]

[0312]

[0313] To suppress jumps caused by single-frame noise, a recursive method combining prior advancement and observation fusion is used to calculate the posterior probability, and then the hysteresis threshold is used to determine the entry state.

[0314] It should be noted that, in the critical state of rollover... In the starting transfer design, the present invention sets This means that the system is not allowed to operate without an intermediate buffer phase. Return directly to normal state The reason is that the rollover threshold corresponds to a high-risk range in the evolution of vehicle roll energy and attitude, exhibiting significant dynamic hysteresis characteristics: even if instantaneous observational evidence shows a decline, the vehicle roll angle and roll rate increase may still rise again in a short period of time. If allowed... to Direct jumps will cause the state to fluctuate frequently between dangerous and normal, leading to frequent switching and miscontrol of the AEB strategy.

[0315] Therefore, this invention employs a forced buffering hysteresis mechanism: when in At that time, the system only allows buffer states related to sideslip or roll (such as...) (etc.) and combined with side-slide continuous gating With recovery gating The common criteria are gradually withdrawn from the high-risk range, thereby ensuring the continuity and stability of state transitions and improving the robust identification capability of real rollover risks.

[0316] Prior advancement:

[0317]

[0318] in, This is a one-step prediction of the prior.

[0319] Observation fusion:

[0320]

[0321] in, The observation likelihood of state i is expressed as follows: If and only if the consecutive L frames are satisfied Then it is determined that state i has been entered. When exiting the judgment, among which , All parameters are preset. To represent state The corresponding observed mean vector, i.e. .

[0322] Finally, define the cross-stage output labels: ,also Represents cross-line target At the moment of cross-line event triggering The posterior probability of being in state i. That is:

[0323]

[0324] This fusion mechanism can effectively overcome the measurement noise of a single sensor under changes in lighting conditions on curves or road bumps, and uses historical prior information to smooth out instantaneous interference, thereby ensuring that the final output stage posterior probability has both sensitivity to sudden loss of control and robustness to sensor noise, and can solve the problem of high-frequency state jitter that is easy to occur in traditional logic judgment.

[0325] Step S304. Defining the 3D envelope space based on stage labels:

[0326] A three-dimensional envelope influence space of an out-of-control vehicle in the future time domain is constructed. This influence space is used to characterize the spatial domain that the vehicle may occupy and its uncertainty expansion, and the differences in spatial morphology between the skidding stage and the rollover stage are explicitly encoded.

[0327] The geometric dimensions of the target vehicle are represented as ,in Let be the length, width, and height of the target vehicle, respectively. Define the three-dimensional directed envelope of the vehicle at time t as . ,in It is a three-dimensional collider. For the target location, For heading angle, For the rolling corner proxy, This indicates the directed envelope box construction operator.

[0328] Construct a three-dimensional semi-ellipsoidal reachable region for the h-th predicted time t+h:

[0329]

[0330] in, For the stage Calculate the three-dimensional radius below. The coordinates are in the lane coordinate system. The longitudinal, lateral, and vertical half-axis lengths are phase-dependent. This design enables the system to calculate the potential expansion space of an out-of-control vehicle due to abrupt attitude changes before a physical collision occurs, providing a geometric benchmark with uncertainty boundaries for subsequent collision probability calculations.

[0331] Adaptive construction of the semi-axis based on the identified stage:

[0332]

[0333] in, For the target speed, For the long walk, Based on the standard expansion, The expansion coefficient varies with the stage. This represents the average value of the sliding window's side-slip strength. To reflect the vertical expansion caused by the risk of rollover, This indicates the length of the prediction time domain, reflecting the expansion of model uncertainty that accumulates as the prediction step size increases.

[0334] The influence space of the defined phase conditions is:

[0335]

[0336] in, For the stage The three-dimensional influence space of the next future moment. Minkowski sum is used to express the geometric expansion of vehicle entity occupancy and uncertainty. The nominal occupancy set of the target vehicle at time t+h (which can be obtained by placing the vehicle's rectangular / polygonal shape in the predicted attitude in the ground coordinate system) is used to characterize the vehicle's geometric dimensions.

[0337] It should be noted that the three-dimensional semi-ellipsoid affects space. The uncertainty envelope used to characterize the possible position of the target vehicle m in the prediction time domain has the following geometric meaning: while the target vehicle is moving along the nominal trajectory, considering perception error, motion model error, and spatial diffusion caused by instability evolution, the uncertainty envelope of the target vehicle at time [time value missing] is [value missing]. The potential occupied region is approximated by a computable set of quadratic forms. In the formula... , , These correspond to the longitudinal, lateral, and vertical semi-axis scales, respectively, and are used to control the intensity of risk expansion in different directions, thereby providing differentiable and recursive geometric constraints for subsequent collision prediction and energy dissipation corridor planning.

[0338] Define the confidence set , Given the confidence level, the output confidence influence space is: .

[0339] Define the predicted state vector ;in, Let the target be the vertical and horizontal coordinates in Frenet. For speed, For heading angle, For the side slip angle proxy, These are the roll angle and its growth rate, respectively.

[0340] Based on the stage posterior definition of the stage conditional dynamic equation obtained in step S303:

[0341]

[0342] in, Let z be the conditional dynamics evolution function corresponding to stage z, used to describe the target vehicle's evolution from its current state in the current stage z. Predict the state for the next time step Evolutionary relationships; For the stage The dynamic parameter vector; It represents a random perturbation term. It has parameter space. , For the stage The parameter domain, The road surface adhesion coefficient, The side-off attenuation coefficient is... This is the roll damping coefficient. This is the roll stiffness coefficient. The longitudinal acceleration coefficient, This is the equivalent steering input.

[0343] when hour, ;in, The lane tangential direction angle, For longitudinal acceleration, For heading angle, For speed, This represents the lateral offset coordinates of the target vehicle relative to the lane centerline in the Frenet coordinate system. The rate of change of the lateral offset coordinate;

[0344] when hour, ;in, The attenuation coefficient of the side plate. For lateral excitation gain, For lateral acceleration agents, Let the rate of change of the longitudinal arc length coordinate of the target vehicle in the Frenet coordinate system be denoted as . The rate of change of the longitudinal arc length coordinate of the target vehicle in the Frenet coordinate system.

[0345] when hour,

[0346]

[0347] in, For the roll angular velocity, The rolling torque is given by:

[0348]

[0349]

[0350] in, These are roll damping, roll stiffness, and roll inertia, respectively. Let be the first derivative of the target vehicle's roll angular velocity, i.e., the roll angular acceleration. For the roll angle proxy value, The equivalent wheelbase of the target vehicle. The equivalent height of the target vehicle's center of gravity relative to the roll rotation reference axis.

[0351] N sets of parameter samples are generated using Latin hypercube sampling: ;in, For Latin hypercube sampling operators, The number of samples is given. For each set of parameter samples, the future H-step trajectory is obtained using the fourth-order Runge-Kutta method or equivalent discrete integral:

[0352]

[0353] in, For the nth predicted trajectory, For the fourth-order Rungekuta integral operator, To predict the number of steps, The step size.

[0354] It should be noted that numerical integration is used to achieve nonlinear dynamic prediction for the instability phase. Since the kinematics / dynamics of the target vehicle exhibit strong nonlinear characteristics during the critical phase of sideslip or rollover, using low-order discretization methods can easily lead to accumulated errors and prediction bias, thus affecting the reliability of subsequent collision determination. Therefore, this invention preferably uses a fourth-order Runge-Kutta (RK4) approach for the phase-conditional dynamic equations. Stepwise integration is performed to improve the numerical stability and accuracy of multi-step predictions.

[0355] Furthermore, considering the stage parameters Uncertainties exist (e.g., road surface adhesion, lateral damping, roll damping, etc. fluctuate with operating conditions), and this invention addresses these uncertainties in the parameter domain. N sets of parameter samples are generated using Latin hypercube sampling. And perform integration on each group of samples to obtain N predicted trajectories. The resulting trajectory bundle covers multiple possible evolution modes of the target vehicle in space and time, providing sample support for subsequent occupancy set construction and collision risk assessment based on statistical characteristics.

[0356] Find the center (mean) of the predicted trajectory set at each time step:

[0357]

[0358] in, The predicted center location at future time t+h. For trajectory At the position of t+h in the future, The statistical center of the trajectory bundle is used to generate the nominal occupancy set. , For the target vehicle The set of predicted trajectory samples (track bundle index set).

[0359] Step S4. Collision Risk Quantification and TAN Inference:

[0360] Step S401. Vehicle prediction and occupancy construction:

[0361] Based on the current state of the vehicle, a sequence of future states of the vehicle within the prediction window is constructed, and this sequence is mapped to a set of geometric occupancy volumes, thus obtaining the future occupancy set of the vehicle. The state of the vehicle at time t is:

[0362]

[0363] in, The coordinates of the vehicle's planar position; For the vehicle's heading angle, The longitudinal and lateral speeds of the vehicle are... The yaw rate is obtained from the multi-source fusion and time alignment in step S1.

[0364] Define the MPC control vector as , For longitudinal acceleration, For the front wheel steering angle, the predicted step size is... The predicted time is denoted as .

[0365] Generate a reference state sequence within the prediction domain. It includes reference locations Reference heading Compared with the reference longitudinal velocity The reference is determined by the centerline geometry and the desired vehicle speed rule.

[0366] To ensure the interpretability and constrainability of predictions for commercial heavy-duty vehicles (commercial trucks) under curve conditions, this embodiment employs a single-track 3-DOF vehicle model and performs discretization recursion. Let the vehicle mass be m and the yaw moment of inertia be... , The lateral stiffness of the front and rear wheels is At the prediction step, the tire slip angle is defined as:

[0367]

[0368] in, To prevent division by zero constants, Prediction steps Tire slip angles of the front and rear wheels; For the prediction of the car's steps Yaw angular velocity at that point The predicted steps for each vehicle The lateral and longitudinal velocities at that point.

[0369] The lateral forces on the front and rear axles are modeled using a linear lateral slippage method:

[0370]

[0371] The longitudinal resultant force can be approximated by the longitudinal acceleration as follows:

[0372]

[0373] in, For the equivalent vehicle weight, For the prediction of the car's steps The longitudinal acceleration at that point.

[0374] Therefore, the discrete state of the vehicle in the prediction domain can be recursively written as:

[0375]

[0376] in, The predicted steps for each vehicle The global planar coordinates x and y components at the location, Let be the yaw moment of inertia of the vehicle about the vertical z-axis. For the prediction of the car's steps The heading angle at that location. The initial value (from step S1) is used to uniquely obtain the state sequence given the control sequence.

[0377] To obtain a future state sequence that conforms to road constraints and vehicle feasibility within the prediction domain, this embodiment calculates the control sequence using MPC, ensuring that the predicted state follows the reference as closely as possible and avoiding unnecessary drastic maneuvers. First, lateral error and heading error are defined: at the reference point, the lateral error is... The heading error is defined as Longitudinal velocity error is defined as Therefore, the objective function of MPC is taken as:

[0378]

[0379] in, This is a weighting coefficient used to balance "Lane Fit / Speed ​​Reference" and "Control Smoothness".

[0380] Define the constraint as a control amplitude constraint: Acceleration rate of change constraint , This is the maximum allowable value of the longitudinal acceleration change rate of the vehicle, i.e., the maximum jerk (maximum impact rate) constraint.

[0381] By solving the above problem, the predictive control sequence is obtained. With the corresponding predicted state sequence ,in This refers to the sequence of future states of the vehicle output in this step.

[0382] To obtain the optimal prediction state Then, the geometric occupancy of each prediction step is defined as a rotated rectangle (expandable in height to form a 3D occupancy), denoted as... The expression is:

[0383]

[0384] in, , , These are the vehicle's external dimensions. After optimization for MPC, the autonomous vehicle predicts steps The optimal planar position x and y coordinates, After optimization for MPC, the autonomous vehicle predicts steps The optimal heading angle. Let the two-dimensional rotation matrix be... , Represented by the optimal heading angle The constructed two-dimensional rotation matrix. From this, we obtain... .

[0385] Step S402. Construction of Multi-Input Metrics and TAN Learning:

[0386] For each predicted time of target j, define the relative position and relative velocity:

[0387]

[0388] in, The target vehicle j and the self vehicle are respectively in Location at any given moment The target vehicle j and the self vehicle are respectively in The speed of time.

[0389] Construct a set of multi-input indicators related to vehicle AEB triggering and collision point prediction.

[0390] Furthermore, the Frenet coordinates are obtained by projecting the positions of the vehicle and the target at prediction step h onto the lane centerline. The longitudinal arc length coordinates of the vehicle and the target are denoted as follows: The horizontal offsets are respectively And define the longitudinal relative quantity and the transverse relative quantity as Let the longitudinal velocity difference be... ,in , These are the velocity components along the lane tangential direction for both the target vehicle j and the driver vehicle.

[0391] Longitudinal approximation of risk indicators: when and At that time, there is a longitudinal TTC. Otherwise, when there is a car front, the distance to THW is ,in To prevent longitudinal velocity stabilization small constants with denominators of zero or close to zero.

[0392] To avoid the dangers caused by longitudinal collisions, the deceleration required to avoid longitudinal collisions is defined. and normalized braking demand ratio ,in This represents the upper limit of the maximum available deceleration under the current attachment and load conditions. This indicates saturation cutoff.

[0393] In this embodiment, the longitudinal hazard index group constructs a multi-dimensional longitudinal risk assessment system; among them, TTC and THW characterize collision risk from the perspective of time urgency, reflecting 'how long until the collision' and 'follow-the-car safety margin' respectively; while the normalized braking demand ratio characterizes the physical feasibility of avoidance from the perspective of vehicle dynamics, by comparing the theoretical required deceleration with the maximum physical braking capacity ( By comparing the time indicators, it is possible to identify 'dynamic boundary risks' where the TTC is still large but the vehicle speed is too high to stop, thus overcoming the limitations of a single time indicator in the case of heavy-duty trucks.

[0394] Based on step S3, the N predicted future trajectories of target j are given. The frequency of lane intrusion is statistically analyzed, and the probability of lane intrusion is introduced. ,in Lane width, For lateral safety margin, This is an indicator function. That is, it takes the value 1 when the condition inside the parentheses is true, and takes the value 0 when the condition inside the parentheses is false.

[0395] Furthermore, the predicted vehicle occupancy volume output in step S401 is combined with... and the occupied body corresponding to the nth trajectory of the target. Define the probability of occupancy conflict .

[0396] The probability of lane intrusion Conflict probability with occupation The purpose is to quantify the uncertainty of predictions using the statistical characteristics of large samples. Unlike traditional geometric methods that rely solely on a single deterministic trajectory for a binary decision of presence or absence, this step transforms the fuzzy risk caused by the target vehicle's instability or the tail-wagging of a long trailer into continuous probability values ​​by statistically analyzing the distribution frequency of N predicted trajectories (trajectory bundles) within the lane boundary and the vehicle's own body. This allows the system to detect the presence of the target vehicle even before it has fully entered the lane. It can detect risks in advance when the risk is low but there is a high tendency to intrude, and effectively suppress false triggers caused by errors in the prediction of a single trajectory.

[0397] Let the set of stages output by step S3 be... The posterior of the stage is Then the stage danger weight is This forms a set of multiple input continuous variables. .

[0398] Based on the stage posterior and confidence influence space and the multi-trajectory prediction set from step S3, a discrete input is constructed. It is used to characterize whether there is a possibility of encroachment into this lane in the future and whether there will be a collision with the occupying body; the above quantity is not used as the sole criterion for the final collision point, but as one of the multiple input variables to participate in the subsequent TAN-BN relationship discovery and probability inference.

[0399] Based on the data, each continuous variable By threshold Discretized into five levels of states .

[0400]

[0401] Finally, we obtain the multi-input state vector:

[0402]

[0403] in, express Discretized state symbols express Discretized state, express Discretized state symbols These represent the discrete state symbols for the lane intrusion probability and the lane occupancy conflict probability, respectively. The state symbol is represented by the discrete instability risk weights obtained from the stage posterior weighting. To represent the discrete labels for the target vehicle's instability / crossing phase, For can be The maximum a posteriori state is obtained, i.e. .

[0404] Discretize continuous variables and construct state vectors The role of this step is to reduce noise and align formats. By mapping threshold intervals, this step effectively filters high-frequency noise from the original sensor data (such as small distance fluctuations or probability jumps) and extracts state features with significant distinguishability, such as safe, low-risk, and high-risk. At the same time, the discretized state vectors meet the input requirements of Bayesian networks for evidence nodes, laying a data foundation for subsequent complex logical reasoning based on probabilistic graphical models.

[0405] Furthermore, this invention no longer assumes that the input variables are independent of each other. Instead, it uses Tree-Enhanced Naive Bayes Network (TAN-BN) to characterize the dependence strength between input variables using conditional mutual information under class variable conditions, and automatically learns the structural relationships between each variable, thereby more realistically reflecting the dangerous coupling mechanism under multi-input conditions.

[0406] Let class variables This represents the cell number of the possible collision point at a future time t+h, which means dividing the area of ​​interest ahead of this lane into several grid cells in Frenet. If a collision occurs, the point of collision will fall into a certain unit.

[0407] For random variables Calculate conditional mutual information under class variable conditions:

[0408]

[0409] in, The discrete states of the random variable take values. Assign values ​​to the collision point element numbers. For random variables Values , Values Class variables Values The joint probability, In order to be in conditions, and The conditional joint probability, In order to be in under conditions The conditional marginal probability.

[0410] by As edge weights, a maximum weight spanning tree is constructed to obtain the parent node mapping, thereby ensuring that each input variable satisfies... After the structure is determined, a conditional probability table is obtained from historical data statistics.

[0411] The conditional mutual information formula is the core of constructing the TAN network topology, and its function is to mine and quantify the coupling dependencies between risk variables. This formula calculates the information overlap between any two input features given the class variables. Using this as weights, a maximum-weight spanning tree is constructed, enabling the final generated Bayesian network to break the strong assumption of feature independence in traditional Naive Bayes algorithms, automatically retaining the strongest feature dependency edges. This makes the inference model more consistent with the real physical causal chain of vehicle loss of control leading to intrusion and subsequent collision.

[0412] Step S403. Rolling Inference and High-Risk Time Window:

[0413] The TAN structure learned in step S402 is used for time-by-time rolling inference within the prediction domain. At each future time, the posterior distribution of the grid cell where the collision point is located is output, and the predicted collision point and high-risk duration window are obtained accordingly.

[0414] To improve robustness to noise and sudden changes, and to avoid unnecessary updates in the absence of cross-line events, this embodiment employs an event-triggered and exponentially decaying Markov prior update strategy: statistical updates to the state transition of discrete input variables are only performed when the target vehicle j enters the cross-line analysis window (triggered by step S3). Let a certain discrete input variable be... Its five-state set is For discrete sequences within a short time window Define a five-state transition matrix The elements are ,in The coefficient for exponential decay is denoted by N, with the weighting increasing as the transition approaches the current time; N is the short-time window length. For indicator functions; To prevent extremely small constants with a denominator of zero, let the prior distribution vector at the current time be... Then the prior information for future moments can be recursively updated as follows: .

[0415] At prediction step h, the conflict concern area for this lane is defined as... Divide it into K non-intersecting grid cells. : And record the first The geometric center coordinates of each grid cell are .

[0416] Define the class variable as the grid number to which the collision location belongs: That is, if a conflict occurs in the future prediction step h, the collision point falls into the k-th grid cell, denoted as . For ease of explanation, a mesh mapping operator is introduced. Map the plane points to their corresponding grid numbers: ;in Let h be the collision point of target j in prediction step h.

[0417] For each future prediction step, under the TAN structure learned in step S402, using multi-input discrete state observations as evidence, the collision grid cell class variables are... The class variable is inferred from the posterior probability. This indicates the grid cell number where the predicted collision point of target j is located at prediction step h.

[0418] Under the TAN structure, the class posterior probability is:

[0419]

[0420] in, The multi-input discrete state observation constructed for step S402; For nodes in the TAN structure The index of the parent node. The prior distribution of the class variable is initialized by offline sample statistics. The representative indicates the class variable. Values and nodes Attribute parent node Values Under the condition of input attribute variables Values The conditional probability of nodes in a TAN network. The conditional probability table entries. In rolling prediction, the class posterior pair from the previous prediction step can be further used. Perform recursive updates to ensure timing consistency.

[0421] The predicted collision point at time t+h for the geometric center of each grid cell can be taken as the posterior expectation:

[0422]

[0423] Furthermore, to obtain the set of high-confidence collision regions, a set of high-confidence collision units is given:

[0424]

[0425] in, This is the confidence threshold.

[0426] To characterize the overall probability of a collision occurring within the conflict zone of this lane, let Let represent the set of grid cell numbers falling within the conflict zone of this lane. Then, the collision probability of prediction step h is defined as... The moment when the first collision hazard occurs is defined as the prediction step in which the collision probability first exceeds the trigger threshold: then ,in This is the collision probability trigger threshold. Therefore, the time of first collision danger is... ,in To predict the step size.

[0427] High risk is defined as a period of time in which the probability of collision in future sequences remains higher than a threshold. The threshold is:

[0428]

[0429] The duration of high risk and its corresponding time window are as follows:

[0430]

[0431] in, This is the first step in predicting the entry of the first consecutive high-risk time window. This is the first step in predicting the exit from the first consecutive high-risk time window. For high-risk duration, This corresponds to a high-risk time period in the future. This output is used in subsequent step S5 to adaptively adjust the trigger threshold and control strategy based on whether there is a continuous high-risk window during the duration of the preceding intrusion.

[0432] Step S5. AEB graded triggering and braking output:

[0433] Specifically, this step first determines the set of times and locations of the nearest possible collision based on the time-series output of step S4, which is used to form the minimum forward braking requirement; then, under the premise of intending to perform braking, the most dangerous rear vehicle is selected using the set of rear vehicles and a rear-end collision constraint is formed, thereby obtaining the braking feasible region; finally, within this feasible region, the impact rate (jerk) constraint and the damage cost to cargo and driver are introduced to complete the joint optimization of the AEB trigger threshold and the deceleration magnitude, and output the graded triggering results of warning / light braking / forced braking / cooperative braking.

[0434] Step S501. Earliest collision step and braking feasible region:

[0435] From the set of temporal collision points output in step S4, find the prediction step where the earliest spatial conflict occurs.

[0436] First, define the collision indication: , The term represents the high-confidence collision zone of target vehicle j in prediction step h.

[0437] If and only if the first When there are overlapping regions at each prediction time (i.e., the set of collision points is not empty), This indicates that a collision is possible at that moment. The earliest collision step is determined based on this.

[0438]

[0439]

[0440]

[0441] in, This is the earliest possible prediction step for a collision. It is the earliest collision hazard warning time, that is, how long after the current time t will the system first determine that a collision hazard has been detected. This represents the high-confidence collision region corresponding to the earliest collision step.

[0442] Furthermore, to facilitate the conversion of PCT into a distance quantity that can be used for braking, the nearest distance to the set of collision points is defined: ;in This refers to the current position of the vehicle, i.e., the time. The planar coordinates of the forward reference point of the vehicle are specified, wherein the forward reference point is preferably the center point of the vehicle's front end or the center point of the front axle. Representative indicates region The planar coordinates of any point within the area. Then, the high-risk duration segment, duration, and collision probability sequence output in step S4 are aggregated into a forward hazard demand, based on the nearest distance to the collision point set and the vehicle speed. Minimum braking requirements for construction:

[0443]

[0444] in, For safety margin.

[0445] Subsequently, assuming the vehicle may trigger AEB braking, a threat assessment is performed on the group of rear vehicles (R), selecting the vehicle most sensitive to the vehicle's braking and most likely to cause a rear-end collision. Define the speed relative to the vehicle. ;in Given the speed of the following vehicle, and the braking speed of this vehicle being 'a', the safety margin for a rear-end collision is defined as follows:

[0446]

[0447] in, For the reaction time of the following vehicle, Let r be the distance between the following vehicle and this vehicle. For the maximum braking capacity of the following vehicle, Assuming the longitudinal braking deceleration amplitude applied by the vehicle at the current moment, this is used to assess whether the following vehicle still has a safe reaction and braking space under this braking intensity. For rearward safety margin, the most dangerous following vehicle is selected based on the minimum necessary braking, resulting in... Based on the collected data, the maximum permissible rearward braking intensity is calculated:

[0448]

[0449] in, It is a supremum function. When the vehicle applies braking deceleration magnitude 'a', the most dangerous following vehicle is... The corresponding rear-end collision safety margin function; when The value indicates that the rear-end collision prevention constraint is still satisfied under this braking intensity. Here, it represents the maximum acceptable braking intensity to satisfy the constraint, thus yielding the AEB braking feasible region. .

[0450] Step S502. Graded triggering and smooth braking command:

[0451] Combination Within the feasible braking domain, select the appropriate AEB trigger level and braking intensity, and make the braking smoother by using jerk constraints and the costs of cargo and driver damage (cargo damage and comfort costs), and finally output the AEB graded trigger result and braking command.

[0452] Specifically, the AEB hierarchical triggering logic: when the condition is met... (Minimum duration) further Determine the urgency level and trigger level:

[0453]

[0454] in, For the graded time threshold, For high-risk durations, the above trigger levels are mapped to AEB mode:

[0455]

[0456] WARN indicates only a warning; These are light braking and forced braking, respectively. COOP represents "maximum feasible braking + rearward cooperative warning" under the braking-only framework, used to reduce the probability of the most dangerous rear-end collision, and introduces jerk (impact rate) constraints and cargo and driver damage costs on top of this. Discrete jerk is defined as follows:

[0457]

[0458] in, To determine the upper limit for jerk, construct a cost-benefit analysis of cargo damage and comfort. ,in , These are the weighting coefficients.

[0459] To ensure that the tiered triggering falls within the executable braking intensity, the braking intensity range corresponding to each tier is preset:

[0460]

[0461] in, These are the parameters for graded braking.

[0462] The braking sequence is determined under the above constraints, resulting in:

[0463]

[0464] in, This is the lower bound of the minimum forward braking requirement. The future braking deceleration sequence output by AEB. This is the upper limit of the maximum permissible braking in the rear direction.

[0465] The present invention also provides an electronic device, comprising: one or more processors and a memory; wherein the memory is used to store one or more programs, and when the one or more programs are executed by the one or more processors, the one or more processors implement the AEB triggering method described above for reducing false triggering in curve scenarios.

[0466] The present invention also provides a computer-readable medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the above-described AEB triggering method for reducing false triggering in curve scenarios.

[0467] Those skilled in the art will understand that all or part of the functions of the various methods / modules in the above embodiments can be implemented by hardware or by computer programs. When all or part of the functions in the above embodiments are implemented by computer programs, the program can be stored in a computer-readable storage medium, which may include: read-only memory, random access memory, disk, optical disk, hard disk, etc., and the above functions can be implemented by executing the program with a computer. For example, the program can be stored in the memory of a device, and when the program in the memory is executed by the processor, all or part of the above functions can be implemented.

[0468] In addition, when all or part of the functions in the above embodiments are implemented by computer programs, the programs can also be stored in storage media such as servers, other computers, disks, optical discs, flash drives, or portable hard drives. They can be downloaded or copied to the memory of the local device, or the system of the local device can be updated. When the program in the memory is executed by the processor, all or part of the functions in the above embodiments can be implemented.

[0469] The above-described specific examples are for illustrative purposes only and are not intended to limit the scope of the invention. Those skilled in the art can make various simple deductions, modifications, or substitutions based on the principles of this invention. Therefore, the scope of protection of this invention should be determined by the scope of the claims.

Claims

1. A method for reducing false triggering of AEB (Automatic Emergency Braking) in curve scenarios, characterized in that, The method includes the following steps: Step S1. Multi-source sensing data acquisition, time alignment, and preprocessing; Step S2. Process the road image data obtained in Step S1 and input it into the constructed lane line recognition and lane geometry modeling module to form stable lane boundaries and center lines, establish a Frenet coordinate system, and complete the lane affiliation determination and curve arc length distance compensation for radar-detected target vehicles. If the target vehicle belongs to this lane, proceed to Step S4 to perform occupancy conflict assessment and collision risk inference; if the target vehicle is located outside the lane boundary but within the buffer zone, it is classified as a candidate dangerous target in the adjacent lane, and proceed to Step S3. Step S3. Calculate the lateral distance from the candidate hazardous target to the nearest lane boundary and determine whether it shows a tendency to cross the lane. If so, integrate appearance, motion and interaction features to identify the out-of-control stage of the cross-lane target, and generate a trajectory bundle and a three-dimensional dynamic occupancy envelope under the stage label constraint. Step S4. Based on the current state of the vehicle, predict the occupants corresponding to the future state and trajectory, and calculate the occupancy conflict probability by combining the occupancy envelope corresponding to the predicted trajectory of the candidate dangerous target; calculate the lane intrusion probability of the cross-line target based on the trajectory bundle of the cross-line target; construct multi-input indicators, use tree-enhanced Bayesian network to learn the dependencies between multiple dangerous variables, and infer the collision point and the future high-risk duration window in a rolling manner. Step S5. Combine forward collision avoidance requirements and rear-end collision avoidance constraints to construct a braking feasible domain. Determine the urgency level and AEB trigger level based on the predicted nearest collision time. At the same time, introduce impact rate constraints and damage cost to optimize braking commands. Step S3 specifically includes the following steps: Step S301. Cross-line trigger: Calculate the lateral distance from the candidate hazardous target to the nearest boundary. If the lateral distance in multiple frames within the detection window is greater than the boundary threshold, it is determined that there is a tendency to cross the line, and a set of cross-line targets is constructed. For each cross-line target, a set of neighborhood objects is constructed within the Frenet system, including the vehicle in front, the vehicle behind, and vulnerable traffic participants. Then, the neighborhood interaction priority score is calculated for each candidate traffic participant in the neighborhood object set of the cross-line target, and the top-k objects are selected to participate in the subsequent discrimination. Step S302. Multimodal feature comparison and out-of-control evidence extraction: Based on the original image of the cross-line target and the detection box, the target ROI image is cropped; the lane boundary output from step S2 is processed to generate a binary mask of the lane lines; the binary mask of the lane lines is cropped based on the detection box of the cross-line target to obtain the mask of the lane lines within the ROI; a context vector is constructed based on the positional relationship between the cross-line target and neighboring objects; the acquired data is processed using a multi-branch encoder, and then the multi-branch features are concatenated and fused using MLP to obtain the final frame-level fused features; for the extracted multi-frame data, the abrupt change intensity between adjacent frames is calculated, and the observation vector is constructed by comprehensively considering the sideslip intensity, roll amplitude, roll growth rate, longitudinal front and rear vehicle spacing and the closest distance of vulnerable traffic participants of the cross-line target; Step S303. Phased inference of the coupled state machine: Based on the observation vector constructed in step S302, a coupled state machine is established. ,in, Let m be the state of the target m at time t. For state geometry, This is the normal state. This indicates a normal cross-line status. Indicates the start of a sideslip. Indicates a continuous sideslip state. This indicates the critical state of sideslip. This indicates that a sideslip has occurred; the loss of control process is subdivided into multiple stages, and the target state is matched to the corresponding stage at the moment of crossing the line; Step S304. Defining the 3D envelope space based on stage labels: Based on the target vehicle's position, geometry, heading angle, and roll angle, a proxy calculates the vehicle's three-dimensional directed envelope at time t. A three-dimensional semi-ellipsoidal reachable region is adaptively constructed based on the longitudinal, lateral, and vertical semi-axis lengths of the identified stage. Both are then processed using Minkowski summation to obtain the stage. The three-dimensional influence space of the next future moment, i.e. the three-dimensional dynamic occupancy envelope; at the same time, the future H-step trajectory is predicted.

2. The AEB triggering method for reducing false triggering in curve scenarios according to claim 1, characterized in that, The lane line recognition and lane geometry modeling module includes a lane line recognition network that outputs a pixel-level probability map based on the original road image, and a lane geometry modeling module that outputs lane boundaries, lane centerlines, and centerline curvature functions based on the pixel-level probability map. Step S2 specifically includes the following steps: Step S201. Image Processing and Input Tensor Construction: Grayscale images are extracted from the acquired original road images to obtain grayscale values. These grayscale values ​​are then processed to obtain horizontal and vertical gradient components, which are combined to form a gradient magnitude. This gradient magnitude is then combined with the original three-channel road image in the channel dimension to form a four-channel input tensor. Step S202. The input tensor is processed based on the lane line recognition network. The lane line recognition network first maps the four-channel input to shallow features to obtain shallow feature maps; the encoder downsamples step by step; after downsampling, the intermediate features obtained by downsampling are added element-wise to the result after mapping by the residual branch, and the output of the l-th layer is obtained by nonlinear activation; the multi-scale fusion and attention enhancement module selects feature maps from multiple encoding layer outputs to participate in the fusion, aligns them and splices them to obtain multi-hop features; at the l-th fusion node, the features from different receptive field branches are jointly described to form a three-dimensional description map, and then the three-dimensional description map is aggregated along the branch dimension to generate the two-dimensional description quantity required for spatial attention gating, and then spatial attention weights are generated. The spatial attention weights are used to perform weighted fusion of multi-hop features; finally, decoding and recovery are performed, and after processing and fusion through two parallel coarse structure branches and detail repair branches, the feature of each pixel is mapped to an unnormalized score, and then mapped to a pixel-level probability map. Step S203. Based on the lane geometry modeling module, perform lane line fitting and establish boundary functions: After obtaining the pixel-level probability map, a binary lane line map is generated by thresholding. Then, connected component filtering and thinning are performed to extract the pixel sets of the left and right lane lines. The pixel sets are then mapped to the road surface plane coordinate system through inverse perspective mapping. After that, a fifth-order polynomial fitting is performed on the left and right lane line point sets to establish a boundary function with the vertical coordinate y as the independent variable, thereby outputting the lane center line and the curvature function of the center line. Step S204. Lane attribution identification of surrounding vehicles and curve distance compensation: First, the distance and azimuth measured by the millimeter-wave radar are converted into planar coordinates in the vehicle coordinate system. Based on the lane boundary function, it is determined whether the target vehicle belongs to the current lane. If it does, it is classified as a major hazardous target in the current lane. If it does not belong to the current lane, it is determined whether the target vehicle is located outside the lane boundary and within the buffer zone. If so, it is classified as a candidate hazardous target in the adjacent lane. Based on the identified lane centerline, the curvature correction is applied to the radar straight-line distance to obtain the actual distance between the current vehicle and the target vehicle. Step S205. Coordinate system transformation: Based on the lane centerline obtained in step S203, a Frenet coordinate system is established, and the target vehicle position obtained in step S204 is mapped into the Frenet coordinate system to form the longitudinal arc length coordinate and lateral offset coordinate of the target vehicle.

3. The AEB triggering method for reducing false triggering in curve scenarios according to claim 1, characterized in that, Step S4 specifically includes the following steps: Step S401. Vehicle prediction and occupancy construction: Based on the current state of the vehicle, construct the sequence of future states of the vehicle within the prediction window, and map it to a set of geometric occupants, thereby obtaining the future occupancy set of the vehicle; Step S402. Construction of Multi-Input Metrics and TAN Learning: Constructing a set of continuous variables with multiple inputs Among them, the stage hazard weight is: The probability of conflict is The lane intrusion probability is determined based on the vehicle's predicted occupant and the occupant corresponding to the nth trajectory of the cross-lane target; Based on the determination of N predicted trajectories for the cross-line target in the future; the normalized braking demand ratio is... The longitudinal TTC is determined based on the deceleration required to avoid longitudinal conflict and the upper limit of the maximum available deceleration under current attachment and load conditions; the longitudinal TTC is... The headway of the train is THW. ; Each continuous variable is discretized into five states based on a threshold, resulting in a multi-input state vector. A tree-reinforced Naive Bayes network is used to characterize the dependency strength between input variables using conditional mutual information under class variable conditions, automatically learning the structural relationships between variables. Class variables represent the unit numbers of possible collision points at future time t+h. For random variables... Calculate conditional mutual information under class variable conditions, and construct TAN network topology based on conditional mutual information; Step S403. Rolling Inference and High-Risk Time Window: For each future prediction step, under the TAN network topology, the posterior probability of the collision grid cell class variable is inferred using the multi-input state vector as evidence, the grid cell number where the collision point is located is predicted, the collision probability of prediction step h is predicted, and the moment when the first collision danger event occurs is defined as the prediction step when the collision probability first exceeds the trigger threshold, thus determining the first collision danger time and the future high-risk time period.

4. The AEB triggering method for reducing false triggering in curve scenarios according to claim 1, characterized in that, Step S5 specifically includes the following steps: Step S501. Earliest collision step and braking feasible region: Based on the time-series collision point set output in step S4, the prediction step where the earliest spatial conflict occurs is determined; based on the nearest distance to the collision point set and the vehicle speed, the minimum braking is constructed, the most dangerous following vehicle is selected based on the minimum required braking, the maximum allowable braking intensity in the rear is calculated, and the AEB braking feasible region is obtained. Step S502. Graded triggering and smooth braking command: When the high-risk duration window exceeds the set threshold, the appropriate AEB trigger level and braking intensity are selected within the braking feasible domain, taking into account the earliest collision risk advance time. The braking is made smoother by considering the impact rate constraint and the damage cost to the cargo and driver. Finally, the AEB graded trigger result and braking command are output.

5. The AEB triggering method for reducing false triggering in curve scenarios according to claim 2, characterized in that, In step S202, the intermediate features obtained by downsampling are mapped by residual branches and the result is composed of a two-level convolution-normalization-activation mapping concatenation. When constructing the three-dimensional description map, channel average pooling is first performed on the features output by different receptive field branches to obtain the two-dimensional response maps of each branch. Then, the two-dimensional responses of each branch are stacked on the branch index dimension to form the three-dimensional description map. When generating spatial attention weights, spatial attention weights are obtained by weighted aggregation of neighborhood responses and Sigmoid transformation. Output coarse-structured branches for each channel c ;in, , For interpolation weights, , , , For feature map In the On each channel, the feature values ​​at four integer grid points in the continuous coordinate neighborhood are respectively the top left, top right, bottom left, and bottom right of the four adjacent sampling points in bilinear interpolation; Detailed fixes for branch output ;in, For the fine-branch convolution kernel parameters, For the corresponding bias term, where is the convolution kernel radius, and k is the kernel size. To refine the details, the total number of channels in the input feature map of the branch is adjusted. For detail repair, the input feature map is in the first branch. Input channels, spatial location Eigenvalues ​​at; Then the two are added together element by element at the same output resolution.

6. The AEB triggering method for reducing false triggering in curve scenarios according to claim 1, characterized in that, The multi-branch encoder in step S302 includes an appearance branch, a motion branch, and a geometric interaction branch. The appearance branch is used to extract the tire, body, posture, and relative relationship with the lane lines. The input is the ROI image and the mask of the lane lines within the ROI. First, the channels are stitched together to form the appearance input tensor, and then the appearance features are output after processing by ResNet-18. The motion branch is used to extract motion consistency / abruptness in adjacent ROI images. First, dense optical flow is estimated on adjacent ROI images, and the dense optical flow is input into a lightweight convolutional neural network to obtain motion features. The geometric interaction branch is used to perform two-layer MLP encoding on the context vector to obtain context geometric interaction features.

7. The AEB triggering method for reducing false triggering in curve scenarios according to claim 1, characterized in that, In step S302, the target vehicle velocity vector is estimated based on the tracking displacement, and the motion direction angle is obtained from the velocity vector; based on the vehicle heading angle... , direction angle of motion Constructing side angle proxy Rolling corner agent ;in, For the roll-off estimation module, The target region, i.e., the target ROI image; the rollover rate. , These represent the inter-frame abrupt change intensity, sideslip intensity, roll amplitude, and roll rate of the target vehicle at time t, respectively; inter-frame abrupt change intensity , , For inter-frame feature difference, This represents the fusion feature of adjacent frames.

8. The AEB triggering method for reducing false triggering in curve scenarios according to claim 3, characterized in that, In step S402, the random variable The expression for calculating conditional mutual information under class variable conditions is: ； in,, For the discrete states of a random variable, Assign values ​​to the collision point element numbers. For random variables Values , Values Class variables Values The joint probability, In order to be in conditions, and The conditional joint probability, In order to be in under conditions Conditional marginal probabilities, In order to be in under conditions The conditional marginal probability.

9. The AEB triggering method for reducing false triggering in curve scenarios according to claim 4, characterized in that, In step S502, the earliest collision risk advance time is first determined. Determine the level of urgency Then map to AEB mode : ； in, The time threshold for the classification; ； WARN indicates only a warning; These are light braking and forced braking, respectively; COOP indicates "maximum feasible braking + rearward coordination warning" under the braking-only framework. The discrete impact rate is: ； in, The upper limit of the impact rate, The braking intensity at adjacent time points. represent Impact rate at any moment; Construct cargo and driver damage cost functions ;in, , These are the weighting coefficients; Preset braking intensity ranges for each level: ； in, These are the parameters for graded braking; Determining the braking sequence under constraints yields: ； in, This is the lower bound of the minimum forward braking requirement. The future braking deceleration sequence output by AEB. This is the upper limit of the maximum permissible braking in the rear direction.

Images