A transformer-based blind area trajectory prediction and planning method for an autonomous vehicle

Abstract

The application is suitable for the technical field of automatic driving, and provides a kind of automatic driving vehicle blind area trajectory prediction and planning method based on Transformer, method first constructs key entry point and calculates the occlusion area, and then generates and maintains the belief state of ghost target in combination with risk score;Then the real and ghost target are uniformly coded as Token, input into Transformer for interactive modeling, and output multi-modal trajectory distribution;Finally, based on the predicted risk, a variable anisotropic repulsive potential field is constructed, and safe control instructions are obtained by optimization and hard safety filtering.The application realizes continuous probabilistic estimation and interactive reasoning of potential risks in blind area, and can generate flexible defensive trajectory according to risk confidence and vehicle dynamics state, which improves safety while ensuring traffic efficiency.

Description

Technical Field

[0001] This invention belongs to the field of autonomous driving technology, and in particular relates to a method for predicting and planning blind spot trajectories of autonomous vehicles based on Transformer. Background Technology

[0002] As autonomous driving technology evolves towards Level 3 and above, addressing the complexity of urban road scenarios has become the biggest challenge for system deployment. Among these challenges, blind spots or occlusion problems caused by limited sensor field of view are a major cause of traffic accidents. In scenarios such as turning at intersections, vehicles obstructing crosswalks, and merging into narrow streets, vulnerable road users (VRUs) often suddenly appear in the vehicle's path—a phenomenon commonly known as "ghosting." To improve driving safety, autonomous driving systems not only need to handle currently visible obstacles but also need the ability to anticipate potential dangers. This means that even in situations where perception is lacking, systems must infer risks within blind spots based on environmental semantics and historical information, and make proactive defensive plans in advance.

[0003] Existing top-tier Transformer-based algorithms all presuppose that the input must be an observable historical trajectory sequence. While TrajPT has mastered rich interaction patterns through large-scale pre-training, its input layer still relies on vehicle nodes and edges captured by sensors. When dealing with "ghost peek" scenarios, due to the lack of observation data during the target's occlusion, these models cannot construct corresponding input tokens, resulting in a complete absence of the interaction inference layer. Although Bufen uses kinematic extrapolation to fill in the data, this simple linear recursion cannot quantify the multimodal intentions that the target may have, such as changing direction or sudden stopping, within the blind spot.

[0004] Accessibility-based solutions often rely on worst-case assumptions, leading to overly conservative or even impassable behavior when vehicles pass through intersections (Freezing Robot). In contrast, patented solutions based on rules or fixed potential fields typically generate repulsive fields with fixed shapes, making it impossible to dynamically adjust the avoidance space according to the uncertainty (covariance) of the prediction. Summary of the Invention

[0005] The purpose of this invention is to provide a method for predicting and planning blind spot trajectories of autonomous vehicles based on Transformer, aiming to solve the technical problems existing in the prior art as identified in the background art.

[0006] This invention is implemented as follows: a Transformer-based method for predicting and planning blind spot trajectories for autonomous vehicles. The method is executed cyclically in each processing cycle and specifically includes the following steps:

[0007] 1. Construct a set of key map entry points: Construct a set of key map entry points to represent potential entry points where "ghost peeking" is frequently observed. The entry points are obtained through manual annotation, data-driven clustering, or online updates, which constrains potential risks to a limited number of high-probability locations, providing a priori information for the subsequent generation of controllable ghost targets.

[0008] 2. Calculate the occlusion area: based on the input visibility / occupancy information. With the status of the vehicle Calculate the spatial area that the vehicle's sensors cannot directly detect at the current moment, i.e., the obstructed area. This can be achieved through methods such as raycasting, occupancy grid analysis, or visibility propagation.

[0009] 3. Visibility determination: For each real target detected by the sensor, based on its detection location... Is it located in an obstructed area? Inside, calculate the visibility mask. This is to distinguish visible targets from those that are occluded or lost.

[0010] 4. Ghost target spawning trigger: in obscured areas Inside, from the key entry point collection Select those that meet the criteria (located within the occlusion area and within a threshold distance from the boundary). The entry points constitute the set of trigger points. For each trigger point Calculate the comprehensive risk score If the score exceeds the threshold Then a ghost target will be generated at that point. and add to the set .

[0011] 5. Ghost Target Belief State Definition: For each ghost target... Maintaining a state of belief This state includes at least the probability of existence. State mean State covariance matrix Category Identifier Generation time and the most recent match time .

[0012] 6. Belief transmission and detection correction:

[0013] When the ghost target is not visible, a uniform speed model is used to propagate and update the state and covariance.

[0014] At the same time, its existence probability is updated to achieve risk-driven probability maintenance.

[0015] If a real target is subsequently detected and matched with the ghost target, a state merging and correction process will be performed to convert the ghost target into a real tracking target.

[0016] If the probability of the existence of ghost targets is below the lower limit Survival time exceeds Or the status is outside the valid area If it is not found, then delete it to prevent the set from growing indefinitely.

[0017] 7. Construct a Token and input it into a Transformer for interactive modeling: This involves creating a set of real-world targets. With ghost target set Unified as a set of participants Construct a Token vector for each participant a. This vector integrates state, uncertainty, visibility, existence probability, semantic attributes, guiding features, and occlusion features. By inputting the historical token sequence into the Transformer network, a unified spatiotemporal interaction model for all participants is achieved.

[0018] 8. Transformer Output Multimodal Trajectory Distribution: The Transformer network outputs K candidate future trajectories and their weights for each participant a. This constitutes a multimodal trajectory distribution. For future location sequences, These correspond to the weights. The maximum weight mode can also be defined. For future planning purposes.

[0019] 9. Calculate collision risk and trigger defense mode based on predicted distribution: Construct a set of candidate trajectories for autonomous vehicles. For each candidate trajectory Calculate the upper bound of its collision risk with all participants. If the minimum risk of all candidate trajectories is... Exceeding the risk trigger threshold If this occurs, the defense mode is triggered, restricting subsequent actions to a conservative strategy and adjusting the planning parameters.

[0020] 10. Planning optimization and hard security filtering:

[0021] Planning and optimization: Construct an objective function and obtain the vehicle's reference trajectory by selecting the minimum cost from the candidate set or by directly solving it using an optimizer. .

[0022] Hard safety filtering: Through secondary programming, the reference control... By modifying the safety constraints, the final safety control quantity to be executed is obtained. This forms a closed loop.

[0023] The beneficial effects of this invention are:

[0024] 1. Improved continuity and foresight in blind spot risk management: Through continuous updates of the "belief state", the probability estimate of potential targets can be maintained even under long-term occlusion, overcoming the problem of traditional methods failing due to data interruption, and realizing early warning for scenarios such as "ghost peek".

[0025] 2. Enhanced the system's ability to utilize implicit interaction information: By inputting ghost targets into the Transformer along with real targets in the form of tokens carrying rich uncertainty information, the model can infer the game relationship between visible vehicles and potential vehicles in the blind spot, achieving a deeper level of scene understanding and intent prediction.

[0026] 3. Achieved flexible planning with a dynamic balance between safety and efficiency: The proposed variable anisotropic artificial potential field can map the uncertainty (covariance) and existence probability of the predicted output into dynamic changes in the potential field strength and shape. This allows vehicles to maintain traffic efficiency under low-confidence risk, while generating smooth avoidance trajectories that match their own motion capabilities in advance under high-confidence risk, fundamentally solving the problems of conservative and rigid planning behavior.

[0027] 4. A multi-layered security protection system was constructed: The method forms a complete technical chain from "probability prediction → risk quantification → potential field constraint → hard security filtering". The final hard security filter, as an independent security layer, ensures that even under extreme conditions, the system's output control follows basic security constraints, greatly improving the overall robustness and security of the system. Attached Figure Description

[0028] Figure 1 A flowchart of the overall predictive planning system;

[0029] Figure 2 Image showing a scene where pedestrians are being blocked from crossing a zebra crossing;

[0030] Figure 3 This is a scene depicting a sudden pedestrian crossing the road. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0032] like Figure 2 , 3As shown, this invention primarily addresses typical urban road boundary scenarios: when a vehicle approaches a crosswalk or is driving normally on a general urban road, its line of sight is easily obstructed by large vehicles or static facilities on the roadside, making it difficult for the perception system to confirm the existence of potential risks within the blind spot. To address this, this invention aims to generate and maintain the probability distribution of virtual ghost targets within the sensor blind spot by establishing a belief state that can be recursively updated online; utilize the Transformer architecture to perform unified spatiotemporal interaction modeling of the vehicle, obstructions, and ghost targets; and finally, transform the predicted potential collision risks into continuous trajectory constraints that conform to vehicle dynamics through a variable anisotropic potential field, thereby guiding the vehicle to execute a defensive driving strategy that balances safety and traffic efficiency in the high-risk area illustrated.

[0033] like Figure 1 This paper presents a Transformer-based method for predicting and planning blind spot trajectories of autonomous vehicles, characterized in that the method includes:

[0034] Step 1: Construct a set of key entry points K on the map to represent the potential entry points for pedestrians / vehicles appearing from behind obstacles;

[0035]

[0036] in:

[0037] : No. The two-dimensional coordinates of each key entry point in the global / local coordinate system;

[0038] Number of key entry points.

[0039] The critical entry point is obtained in any of the following ways:

[0040] 1) Manual annotation method: Extract pedestrian crossing endpoints, intersection entrances and exits, corners, building entrances and exits, etc. from HDMap;

[0041] 2) Data-driven approach: Clustering historical trajectory starting points / turning points (e.g., DBSCAN / KMeans), with cluster centers serving as... ;

[0042] 3) Online update method: During operation, the high-frequency locations of traffic participants are counted to form a heat map and extract the maximum value points.

[0043] Note: The purpose of this step is to constrain the "potentially appearing targets" to a limited set of high-probability entry points, thereby making the subsequent generation of "ghost targets" interpretable and controllable.

[0044] Step 2: Based on visibility / occupancy information With respect to the vehicle's own condition Calculate the occlusion area ;

[0045] Obstructed area Indicates at time Space areas that cannot be directly detected by the vehicle's sensors.

[0046] ;

[0047] in:

[0048] : A two-dimensional set representation of the occluded region (which can be a set of polygons, a set of grids, or a continuous implicit function);

[0049] : Function for constructing the occluded region.

[0050] The occlusion region can be constructed using the following equivalent method:

[0051] Ray projection is achieved by using the vehicle's sensor as the origin and adjusting the angle. Sampling ray; if the ray first intersects with the obstruction, then the maximum detection distance is reached from the ray behind it. The area is marked as obscured;

[0052] Occupied grid implementation: In the BEV grid, given the free / occupied / unknown components, the unknown connected components are selected and combined with sensor field-of-view clipping to obtain the desired result. ;

[0053] Visibility propagation is achieved by performing a visibility flood-fill from the vehicle's position within the grid, treating any unreachable drivable space as occlusion.

[0054] Step 3: Analyze the detection set For each real target in the image, calculate the visibility mask. ;

[0055] ;

[0056] in:

[0057] Index of real traffic participants;

[0058] :Target The detection location (from sensor detection, the hat indicates the detection measurement);

[0059] Visibility mask: 1 indicates that it is currently visible (directly detected by the sensor), and 0 indicates that it is in an obstructed area or is temporarily lost.

[0060] Note: The visibility mask is not only used to "determine whether the target exists", but also serves as an input feature of the Transformer, enabling the network to distinguish between "visible and definite targets" and "invisible and highly uncertain targets".

[0061] Step 4: Within the occluded area, filter out points located within the occluded area from the set of key entry points K. And the distance from the occlusion boundary does not exceed the boundary neighborhood threshold. The entry points constitute the set of trigger points. And calculate a risk score for each trigger point. Generate ghost targets And add to the ghost target set ;

[0062] ;

[0063] in:

[0064] : The boundary of the obstructed area;

[0065] The minimum Euclidean distance from a point to the boundary of a set;

[0066] : Boundary neighborhood threshold, indicating that "entry points near occlusion boundaries are more likely to suddenly appear as targets".

[0067] For each Calculate risk score If the threshold is exceeded, a ghost target will be generated. :

[0068] ;

[0069] in:

[0070] Historical pass rate / occurrence frequency of entry points;

[0071] Vehicle speed;

[0072] Area of ​​the obstructed area and number of grid cells;

[0073] Dynamic enhancements, such as the currently detected pedestrian / vehicle density near the entrance, traffic light status, and prior knowledge of school / shopping mall areas;

[0074] Weighting coefficient;

[0075] Risk trigger threshold;

[0076] :time A set of ghost targets;

[0077] Ghost Target This indicates the hypothesis of "possible but currently invisible traffic participants".

[0078] Note: This step is one of the key coupling points: ghost targets are not hypothetical out of thin air, but are determined by "occlusion geometry". +High-risk entry + Risk Function "Joint triggering ensures controllability and project implementation."

[0079] Step 5: For each ghost target Maintaining a state of belief ;

[0080] ;

[0081] in:

[0082] Ghost Target At any moment The probability of its existence (representing "how credible is this ghost hypothesis");

[0083] State mean vector; Recommendation use It can also be extended to Add acceleration;

[0084] State covariance matrix, describing the uncertainty of position and velocity and their correlation;

[0085] Category identifier (pedestrian / vehicle / cyclist, etc.) is used for subsequent potential field parameters and velocity priors;

[0086] Ghost target spawn time;

[0087] : The most recent time that a match was made with a real probe (can be left blank if there has never been a match).

[0088] Step 6: When the ghost target When invisible, a uniform propagation model and covariance update are used, while simultaneously updating the existence probability. If a real detection occurs in subsequent moments With ghost targets During matching, merge and correct the state; if a probability exists... Below the lower limit of the probability of existence Survival time exceeds the maximum number of frames a ghost can survive. If the status exceeds the valid area S, then delete the ghost target. ;

[0089] 1. Propagation when invisible (occlusion / loss)

[0090] When a ghost target is invisible (which is essentially its nature), it propagates using a uniform propagation model:

[0091] ;

[0092] in:

[0093] : State transition matrix; when hour:

[0094] ;

[0095] Process noise covariance is used to represent uncertainties such as "the target may accelerate / change direction"; different values ​​are set according to categories. (Pedestrians have more flexibility, while vehicles are more restricted).

[0096] 2. Probability update exists

[0097] ;

[0098] in:

[0099] : Memory coefficient; The larger the value, the more it "believes in history" (the probability of its existence decays slowly). The smaller the value, the more "convinced of current risk" (rapid changes).

[0100] Sigmoid: Maps any real number to a sigmoid function. This is used to smoothly map risk scores to "presence tendency".

[0101] : Bias term, which controls the baseline presence tendency of ghost targets when the risk is 0.

[0102] Risk gain: The intensity of the impact of risk on the probability of existence.

[0103] The entrance point corresponding to this ghost target. At any moment Risk assessment (see step 4).

[0104] : Truncation function to ensure the result strictly falls within interval; for example Then it returns 0. Then it returns 1.

[0105] This updated form realizes "occlusion geometry and risk-driven probabilistic ghost existence maintenance", which is an important coupling link in the continuous technology chain of this invention.

[0106] 3. Merging and correction when visible (ghost becomes real)

[0107] When real detection occurs in subsequent moments With ghost targets During matching, perform a merge:

[0108] ;

[0109] like Then, the ghost target will be converted into a real target trajectory ID and its state will be corrected.

[0110] in:

[0111] Mahalanobis distance (the distance considering the uncertainty of covariance, which is more suitable for matching uncertain targets than Euclidean distance);

[0112] : Probe vector (e.g.) );

[0113] : Detection matrix (maps the state to the detection space; if the detection only contains position, then the corresponding component is taken);

[0114] Merging threshold.

[0115] 4. Delete / prune (to prevent ghosts from multiplying indefinitely)

[0116] Delete a ghost target if any of the following conditions are met:

[0117] ;

[0118] in:

[0119] : The lower bound of the probability exists;

[0120] The maximum number of frames a ghost can survive (if it exceeds this number and still doesn't convert, it's considered no longer credible).

[0121] : Effective scene area (such as the union of drivable area / pedestrian area) to prevent ghosts from drifting to impossible areas.

[0122] Step 7: Set up a real set of detectable targets With ghost target set Unified as a set of unified participants For each participant Constructing Token Vectors Input the Transformer for interactive modeling;

[0123] Unify real targets and phantom targets into a single set of participants:

[0124] ;

[0125] in:

[0126] The set of real detectable targets (by...) get);

[0127] Ghost target set;

[0128] : A unified set of participants (the input basis for truly achieving "predictability even when unseen").

[0129] For each participant Construct a token vector :

[0130] State mean, uncertainty expansion, visibility, existence probability, semantic-physical properties, intersection guidance features, occlusion geometric features

[0131] ;

[0132] in:

[0133] :like If the target is real, take the average of its tracking state; if the target is ghost, take the average obtained from propagation in step 6.

[0134] : Straighten the covariance matrix into a vector by rows / columns to express the magnitude and direction of uncertainty;

[0135] Visibility (real targets are obtained from step 3, ghost targets are fixed at 0 or defined according to the visibility of occlusion boundaries);

[0136] : Probability of existence (real target is 1, ghost target is 1) );

[0137] Semantic / physical attributes, such as category one-hot, length and width, velocity modulus, acceleration modulus, heading angle, whether it is within the zebra crossing / lane, etc.

[0138] Intersection / guiding features, such as "embedding of the nearest entry point ID", "distance to the nearest entry point", and "entry point direction vector", are used to make the prediction more "guided".

[0139] Occlusion geometric features, such as "distance to occlusion boundary", "occlusion area", "occlusion sector angle range", and "type of occlusion source object".

[0140] Recently Frame token sequence input Transformer:

[0141] ;

[0142] in:

[0143] History window length (1-3 seconds corresponds to 10-30 frames), used to capture long-term intent and short-term dynamics simultaneously.

[0144] Step 8: The Transformer processes each participant Output The candidate future trajectories and their weights constitute a multimodal trajectory distribution;

[0145] For each participant Transformer output Candidate future trajectories and their weights:

[0146] ;

[0147] in:

[0148] : Number of multimodalities (e.g., 3 to 10), representing "the most likely paths";

[0149] : No. The weights (probabilities) of the trajectories satisfy... ;

[0150] :Future sequence;

[0151] :Future No. Step-by-step two-dimensional position prediction;

[0152] Predict the temporal length (e.g., 3-8 seconds corresponds to 30-80 frames);

[0153] Transformer prediction network (including encoder, attention layer, feedforward layer, and output head).

[0154] Optional output of covariance per step As a form of uncertainty, it is used for more robust risk calculations.

[0155] Define the maximum weighted modal index:

[0156] ;

[0157] in:

[0158] : The candidate trajectory number with the highest weight;

[0159] : Retrieves the index corresponding to the maximum value.

[0160] Step 9: Construct a set of candidate trajectories for the vehicle Calculate the risk for each candidate trajectory of the autonomous vehicle. For each candidate trajectory of a vehicle, calculate the correlation with all participants. The upper bound of collision risk is determined when the minimum risk of all candidate trajectories for autonomous vehicles exceeds the risk trigger threshold. If so, the defense mode will be triggered;

[0161] The candidate trajectory is represented as:

[0162] ;

[0163] in:

[0164] Candidate number;

[0165] Candidate trajectories in the th The position of the step.

[0166] For each candidate trajectory, calculate the upper bound (probability-weighted) on the collision risk with all participants:

[0167] ;

[0168] in:

[0169] The safe distance threshold can be obtained by expanding the shapes of two objects (forming a safe circle / ellipse by adding margins to the length and width of the vehicle).

[0170] Euclidean distance;

[0171] Smoothing coefficient A smaller risk function has a more "hard threshold," while a larger one results in a smoother risk.

[0172] Risk is defined as "maximum time + maximum object", meaning "if the probability of collision with a certain object is high at a certain time in the future, the risk is considered high".

[0173] Conditions for triggering defense mode:

[0174] ;

[0175] in:

[0176] Risk trigger threshold;

[0177] Defense mode indicator;

[0178] When the defense mode is 1, the candidate trajectory set is restricted to conservative action templates (deceleration / stopping / exploratory forward movement), and the potential field coefficient is increased (see step 10).

[0179] Step 10: In the planning optimization, construct an objective function that includes smoothness term, efficiency term, risk term and repulsive potential field cost. Through the objective function, obtain the vehicle reference trajectory based on the vehicle candidate trajectory set, and introduce a hard safety filter to perform secondary planning correction on the reference control to form a closed loop execution.

[0180] 1. Plan the objective function (smoothness + schedule + risk + potential field)

[0181] Choose the candidate trajectory with the minimum cost, or solve it directly using an optimizer:

[0182] ;

[0183] in:

[0184] The trajectory of the vehicle to be determined;

[0185] The smoothness term is defined as the sum of the squares of acceleration and jerk (jerk).

[0186] Efficiency term, defined as the deviation from the desired speed / target point;

[0187] Risk item: Calculated using S9's risk calculation method for trajectory sampling points;

[0188] The cost of a repulsive potential field;

[0189] Weight; can be increased in defensive mode. To be more conservative.

[0190] 2. Variable anisotropic repulsive potential field (the greater the velocity, the thinner and longer it becomes, and the forward fan shape is enhanced).

[0191] For each participant Prediction (taking the maximum weight mode) (or probability fusion mean), calculate relative pose:

[0192] ;

[0193] in:

[0194] A rotation matrix constructed from the vehicle's heading angle transforms the global coordinates to the vehicle's coordinates;

[0195] Forward relative distance (>0 in front of the vehicle);

[0196] : Horizontal relative distance (>0 on the left or as agreed).

[0197] Define the major and minor axis scales of the potential field related to velocity:

[0198] ;

[0199] in:

[0200] Vehicle speed;

[0201] The minimum turning radius function can be derived from the upper limit of the vehicle's lateral acceleration. Give

[0202] ;

[0203] Adjustable parameters;

[0204] This design makes The larger, The larger, therefore The larger the potential field, the more "slender" its forward influence becomes.

[0205] Define anisotropic distance:

[0206] ;

[0207] Introduce directional weights for "forward sector enhancement":

[0208] ;

[0209] in:

[0210] Forward enhancement coefficient;

[0211] Indicator functions;

[0212] : Lateral attenuation scale, controlling the "angle" of the sector.

[0213] Repulsive potential energy (repulsive force only):

[0214] ;

[0215] in:

[0216] Potential energy coefficient, set by category (small for pedestrians, large for vehicles), can also be amplified overall in defense mode:

[0217] ;

[0218] The cost of the power field:

[0219] ;

[0220] 3. Hard safety filter (maximally reduces the risk of collision)

[0221] Because prediction is probabilistic, the plan may still output dangerous trajectories in extreme cases. Therefore, a hard safety filter is introduced before execution to reference control. Revised to security control .

[0222] One possible implementation method is quadratic programming (QP):

[0223] ;

[0224] ;

[0225] in:

[0226] : The control quantity to be determined;

[0227] : The reference control output by the planner;

[0228] : No. Safety constraint functions, such as distance constraints to the nearest obstacle, road boundary constraints, and maximum deceleration constraints;

[0229] This filter ensures that "no obvious hazard controls are implemented even if the prediction is incorrect".

[0230] Final output execution:

[0231] ;

[0232] Then it enters the next frame loop, forming a closed loop.

[0233] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0234] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

[0235] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting and planning blind spot trajectories of autonomous vehicles based on Transformer, characterized in that, The method includes: Step 1: Construct a set of key entry points K on the map to represent the potential entry points for pedestrians / vehicles appearing from behind obstacles; Step 2: Based on visibility / occupancy information With respect to the vehicle's own condition Calculate the occlusion area ; Step 3: Analyze the detection set For each real target in the image, calculate the visibility mask. ; Step 4: Within the occluded area, filter out points located within the occluded area from the set of key entry points K. And the distance from the occlusion boundary does not exceed the boundary neighborhood threshold. The entry points constitute the set of trigger points. And calculate a risk score for each trigger point. Generate ghost targets And add to the ghost target set ; Step 5: For each ghost target Maintaining a state of belief ; Step 6: When the ghost target When invisible, a uniform propagation model and covariance update are used, while simultaneously updating the existence probability. If a real detection occurs in subsequent moments With ghost targets During matching, merge and correct the state; if a probability exists... Below the lower limit of the probability of existence Survival time exceeds the maximum number of frames a ghost can survive. If the status exceeds the valid area S, then delete the ghost target. ; Step 7: Set up a real set of detectable targets With ghost target set Unified as a unified set of participants For each participant Constructing Token Vectors Input the Transformer for interactive modeling; Step 8: The Transformer processes each participant Output The candidate future trajectories and their weights constitute a multimodal trajectory distribution; Step 9: Construct a set of candidate trajectories for the vehicle Calculate the risk for each candidate trajectory of the autonomous vehicle. For each candidate trajectory of a vehicle, calculate the correlation with all participants. The upper bound of collision risk is determined when the minimum risk of all candidate trajectories for autonomous vehicles exceeds the risk trigger threshold. If so, the defense mode will be triggered; Step 10: In the planning optimization, construct an objective function that includes smoothness term, efficiency term, risk term and repulsive potential field cost. Through the objective function, obtain the vehicle reference trajectory based on the vehicle candidate trajectory set, and introduce a hard safety filter to perform secondary planning correction on the reference control to form a closed loop execution.

2. The method according to claim 1, characterized in that, In step 1, the set of key entry points ; in: : No. The two-dimensional coordinates of each key entry point in the global / local coordinate system; Number of key entry points; The set of key entry points K for the map can be constructed in any of the following ways: Manual annotation method: Extract the two ends of pedestrian crossings, intersection entrances and exits, corners or building entrances and exits from HDMap; Data-driven approach: Clustering historical trajectory starting points and turning points, with cluster centers serving as... ; Online update method: During operation, the frequency of traffic participants appearing at different locations is counted, locations with a frequency higher than a preset threshold are filtered out, a heat map is generated, and maximum points are extracted.

3. The method according to claim 1, characterized in that, In step 2, according to Calculate the occlusion area Specifically: ； in: : A two-dimensional set representation of the occluded region; : Function for constructing the occluded region.

4. The method according to claim 1, characterized in that, The set of trigger points ; in: : The boundary of the obstructed area; The minimum Euclidean distance from a point to the boundary of a set; Boundary neighborhood threshold; For the risk score ； in: Historical pass rate / occurrence frequency of entry points; Vehicle speed; Area of ​​the obstructed area and number of grid cells; Dynamic enhancements; Weighting coefficients; Risk trigger threshold; :time A collection of ghost targets.

5. The method according to claim 1, characterized in that, In step 6, the update has a probability. Specifically: ； in: : Memory coefficient; Sigmoid: Maps any real number to a sigmoid function. This is used to smoothly map risk scores to a tendency to exist. Bias term, which controls the baseline presence tendency of ghost targets when the risk is 0; Risk gain: controlling the intensity of the impact of risk on the probability of existence; : Truncation function to ensure the result strictly falls within interval; If a real detection occurs in subsequent moments With ghost targets During matching, a merge and state correction are performed, specifically as follows: ； like Then the ghost target will be converted into a real target trajectory ID and its state will be corrected; in: Mahalanobis distance; : Probe vector; : State mean vector; : The probe matrix is ​​used to map the state to the probe space; Merging threshold; Ghost Target At any moment The state covariance matrix.

6. The method according to claim 1, characterized in that, Each participant Constructing Token Vectors Specifically, it includes: ； in: State mean, if If the target is a real target, the average value of its tracking state is used; if the target is a ghost target, the average value obtained from the uniform speed model propagation is used. Uncertainty expansion flattens the covariance matrix into a vector by rows / columns, expressing the magnitude and direction of uncertainty; Visibility; : Probability of existence, real target is set to 1, ghost target is set to 1. ; Semantic / physical attributes; Intersection / Guiding Features; Occlusion geometric features; The input Transformer performs interactive modeling, specifically including: Recently Frame token sequence input Transformer: ； in: : Historical window length, used to capture both long-term intent and short-term dynamics simultaneously.

7. The method according to claim 1, characterized in that, In step 8, the Transformer performs a process for each participant. Output The candidate future trajectories and their weights include: ； in: : Number of multimodalities, representing the most likely paths; : No. The weights of the trajectories satisfy... ; : Future order; :Future No. Step-by-step two-dimensional position prediction; Predict the length of the time domain; Transformer prediction network, including encoder, attention layer, feedforward layer, and output head.

8. The method according to claim 1, characterized in that, In step 9, the risk is calculated for each candidate trajectory of the autonomous vehicle. Specifically: ； in: The safe distance threshold can be obtained by the expansion of the shapes of the two objects. Euclidean distance; , Candidate trajectories in the th The position of the step, Candidate number; Smoothing coefficient A smaller risk function has a harder threshold, while a larger risk function has a smoother threshold. The conditions for triggering the defense mode are as follows: ； in: Risk trigger threshold; Defense mode indicator.

9. The method according to claim 1, characterized in that, The method of obtaining the vehicle reference trajectory based on the candidate trajectory set of the vehicle through the objective function is one of the following methods: Select the trajectory with the lowest cost from the candidate trajectories; Solve directly using the optimizer; The objective function is expressed as: ； in: The trajectory of the vehicle to be determined; The smoothness term is defined as the sum of the squares of acceleration and jerk. Efficiency term, defined as the deviation from the desired speed / target point; The trajectory of the vehicle to be determined Risk items; The cost of a repulsive potential field; Weight.

10. The method according to claim 1, characterized in that, In step 10, the introduction of a hard safety filter performs secondary programming correction on the reference control, achieving the following through secondary programming: ； ； in: : The control quantity to be determined; : The reference control output by the planner; : No. One safety constraint function; Final output execution: ; Then it enters the next frame loop, forming a closed loop.

Images